You probably open something like this with a specific question already in mind. Who should I be learning from in this corner of the field? If I want my work in front of a community I am not part of, who is the way in? Who is worth collaborating with, and who is about to matter before everyone else notices?

Those are good questions, and a list of names ranked by follower count answers none of them well.

It can tell you who is big. It cannot tell you who connects worlds, who only shouts inside one, or who is quietly about to climb. Those are different things, and they matter most.

That gap is why the AI for UX Index is a map and not a ranking: every month it plots the people and institutions shaping how artificial intelligence shows up in design and product work, positioned by what they actually do in the network rather than how loud they are so readers can get the most value.

This piece is about how to read it and why it is built the way it is, starting with what it is actually good for, because the structure only matters once you know what you came to find.

What the map is for

The map answers the questions you came in with, because it positions every voice by two things at once:

• How far they reach

• How many separate communities they connect

Hold those loosely; the next section makes it more precise.

• Figure out who to learn from. Filter to your lane with the AI / UX / PM lens and the tier and category toggles, read the people clustered there, and open any profile for the short note on why they’re on the map.

• Find the way into a community. Look to the voices on the seam between two clusters, and to the institutions that feed a cluster from upstream. Those are the warm paths, the routes your work travels to reach an audience it otherwise never would, the map of the front doors.

• Decide who to collaborate with. The highest-leverage collaborators are the ones who bridge into a community your own cluster cannot reach: one connection through them carries your work somewhere it could not have gone on its own. Reach is nice; adjacency compounds.

• Catch who is rising. The climbers, moving from quiet specialist toward a connecting role, are where the next wave forms. That is the point of a monthly edition: any one snapshot is the field today, the movement between them is where it is heading, and following risers early is cheaper than discovering them late.

• Place yourself, honestly. Find your dot and sit with where it lands. The useful question is rarely “how do I get more reach,” but “do I connect different worlds, or mostly reach a lot of people inside one, and is that the role I want?” That functionality will be added at some point.

How the map works: two axes, not one number

A ranking cannot answer those questions because of how it is built: it collapses each person into a single number, almost always a proxy for reach, and sorts. A single number can only measure size. It cannot represent the gap between someone who reaches a hundred thousand people inside one bubble and someone who quietly connects two fields that never speak.

So the Index refuses the single number; it borrows the way of seeing that network science made popular, most of all in Albert-László Barabási’s Linked: picture the field not as a ranked column of names but as a social graph, people and institutions as dots (nodes), relationships as lines (edges).

Reddit is the cleanest everyday version of this: thousands of subreddits, each a dense community, with the occasional user or crosspost carrying an idea from one to the next. The corner of the field this map covers runs the same way, only its communities are Substack newsletters, Medium publications, and conference circuits rather than subreddits, and its bridges are the people who show up across several of them.

Influence stops being one quantity and becomes a question of where a node sits, and the Index plots every voice on the two measurements that capture it.

What degree measures: raw reach

Degree is the simple one: how many connections a node has, here roughly how many people a voice directly touches. It is the axis a follower count already captures, big audience, high degree, and the obvious half of influence, the only half a leaderboard ever sees.

Reach is real, and degree is a fair proxy for it.

What degree cannot tell you is where those connections point, whether they cluster inside one community or fan out across many separate ones, which is exactly the question the second axis exists to answer.

What betweenness measures: bridging power

Betweenness almost never gets measured, and that’s the one that makes the map worth building. Network scientists have used betweenness since Linton Freeman gave it a formal definition in 1977: of all the shortest paths an idea could travel, how many pass through you?

Picture two voices with the same follower count, one whose connections sit inside a single tight community, the other’s across four that don’t overlap.

On a leaderboard they are tied; in the network the first is an echo, the second a junction, the path two separate worlds would lose if it vanished, invisible to degree.

Betweenness addresses this with this concept: Measure how many worlds the user if in. If they are in a lot of rooms, that’s bridging power. Otherwise they’re shouting in a single room.

Why both work together

Neither axis is enough on its own. Degree answers how many people hear you. Betweenness answers how many different worlds you let talk to each other. The two move independently: a voice can sit high on one and low on the other, and that gap is where the interesting value lives.

A huge audience sealed inside a single community is a different thing from a small audience that is the only path between two fields, even though a follower count would file them side by side.

Plotting both instead of collapsing them into one number is the whole method.

It is what turns a flat list into a plane with four corners, and it is why the most interesting voices are the ones whose two scores disagree, the ones a single number would rank next to people they have nothing in common with.

What the four quadrants mean: four roles, not a ranking

Plot degree across and betweenness up, and the plane splits into four corners. Each is a role, not a rank.

Superconnectors: high degree, high betweenness

Big reach and heavy bridging at once: the audience of a Megaphone and the cross-cluster position of a Bridge. Network science calls this a hub, a node whose degree far outruns the average and whose links reach across many communities at once.

Most major institutions land here, the publications, course platforms, and community hubs that command large audiences and sit at the crossroads of several sub-communities.

When a Superconnector picks up an idea it reaches not just a lot of people but a lot of different people; the idea becomes common knowledge across the field rather than one pocket of it.

A single one of them can carry a small voice across the whole field at once.

Examples

• John Maeda

• Lenny Rachitsky

• Andrew Ng

• Don Norman

• Xinran Ma

Bridges: low degree, high betweenness

The punch-above-their-weight quadrant, and the one most invisible to a follower count. A Bridge does not have the biggest audience; what it has is position. It sits on the seam between two communities that don’t otherwise talk, academic human-computer interaction research and shipping product teams, and carries ideas across the gap.

This is the position Mark Granovetter named in 1973 called weak ties: a tie that bridges two densely knit groups carries information neither could produce alone. LinkedIn makes it bread and butter off of it.

Their value is not in volume, it is in adjacency.

Remove a Megaphone and its cluster still hears the same ideas from someone else; remove a Bridge and two communities quietly stop exchanging. They are load-bearing in a way their numbers never show: where the next idea enters your community, and your highest-leverage collaborators.

Examples

• Maggie Appleton

• Amelia Wattenberger

• Linus Lee

• Duane Degler

• Rupa Chaturvedi

Specialists and cite-only: low degree, low betweenness

Low on both axes, which is not a basement but a label: “not yet a distribution engine.” In network terms they sit in the periphery, sparsely linked nodes loosely tied to a dense core, a position that measures how little their work travels, not how good it is.

Specialists are the deep practitioners and researchers whose work gets cited and built on without them broadcasting it, the one solving the hard problem inside an organization, the researcher everyone references who never does the podcast rounds.

They are the foundation; the ideas Bridges carry and Superconnectors amplify often originate here.

A low position usually measures distribution behavior, not substance, and many could move toward Bridge the moment they connect their work outward.

Examples

• Dr. Carmen Martinez

• Siavash Memar

• Mehekk Bassi

• Emilie Mazurek

• Krishna Brown

Megaphones: high degree, low betweenness

Large audiences, concentrated inside one cluster. Brain-network researchers call this a provincial hub: a node with high degree whose links stay almost entirely inside its own community, the mirror image of the connector hub that reaches across them.

The volume is real; it just runs deep in one community, so the ideas land hard at home and rarely jump the fence. This is the quadrant a follower count most overrates, because audience size is exactly what the single number measures.

Influence that does not cross clusters is local: it moves its own community and stops at the edge.

That is not a knock; speaking powerfully to one audience is real work, a different job than bridging, the two the map exists to keep apart.

Examples

• Aakash Gupta

• Pawel Huryn

• Pavel Samsonov

Why the quadrants matter

The point is not to crown the top-right corner. It is to show that influence is at least two independent things, and that a Bridge and a Megaphone, identical on a follower-count leaderboard, do completely different work: one moves ideas between worlds, the other within one.

A list flattens that into a single column; the quadrants are the whole reason to use a map instead.

How this refines Gladwell and improves on Klout

Naming the people who carry ideas is an old instinct, and two famous attempts at it show what the map keeps and what it throws away.

Malcolm Gladwell got the roles half right. In The Tipping Point, his 2000 book, he sorted influential people into Connectors who seem to know everyone, Mavens who gather and share expertise, and Salesmen who persuade. The Maven is close to the Specialist corner, the trusted expert others cite without much broadcasting of their own.

Gladwell’s Connector fuses two jobs the map insists on separating. The person with a huge audience inside one world and the person who is the only door between two worlds are both Connectors to him, and they are nothing alike; degree and betweenness are what pull them apart.

Klout made the opposite mistake. From 2008 until it shut down in 2018, it scored your online influence on a single number from 1 to 100 and let brands sort people by it, the purest version of the thing this map refuses.

One independent analysis found that the logarithm of your follower count alone explained about 95 percent of a Klout score, which is to say the number measured reach and dressed it up as influence. It had no context of community.

The score drifted from anything real, and the company folded.

The lesson is not that influence cannot be read. It is that flattening a position into a rank throws away the only information worth having.

Where the two axes come from, and where else they are used

None of this machinery is invented for the occasion. Degree and betweenness are standard centrality measures in social network analysis, a field that has spent decades asking what makes a node important. Reading a node’s role off more than one of them at once is an established method, not a flourish.

The cleanest precedent is a 2005 paper in Nature by Roger Guimerà and Luís Amaral, “Functional cartography of complex metabolic networks,” which sorts nodes into “universal roles” by placing each in the same kind of two-axis space, one axis for how strong a hub it is, the other for how much it reaches across communities.

Swap the labels and you have Superconnectors, Megaphones, and Bridges.

The measures are not identical to the Index’s, but the move is the same, and it has been reused to map roles in brain networks, ecosystems, and protein interactions.

Closer to most readers’ work, the same two axes drive organizational network analysis: Rob Cross, who has run it across hundreds of organizations, finds leaders only about 30 percent accurate at spotting the brokers who bridge gaps, against roughly 50 percent for the high-degree connectors.

A single number hides the bridging role in a company exactly as it does on this map.

The same idea names what the map does with institutions, the publications, course platforms, and newsletters that sit upstream of audiences and decide what reaches them. Network sociologists call the gaps they span structural holes and call those who span them brokers, who gain an advantage from the gap.

The map draws them as edges into clusters, because a single broker can hand a small voice an enormous amount of borrowed reach.

So the framework is borrowed from settled science even though the placements are not. The axes are real measures and the role classification a real method, drawn in fields with far more data than an influence map for our corner of design ever will.

What stays editorial is where each dot lands.

How people get selected

Selection is curation, not a threshold: no follower minimum, no automated cut line, just Claude and other data added to a prompt for analysis, for fun.

People can ask to be added3.5k3.5k3.5k3.5k3.5k, and I add them, and the number of sources will grow.

The map began as a seed roster I already tracked in the AI-for-UX conversation and grew to take in the ecosystem around it, the institutions, the educators, and the adjacent AI and applied-AI voices that design people actually read.

The test is not whether someone is famous, it is whether they are a node in this discourse, and whether including them helps explain the structure of the field. A quiet researcher everyone cites can matter more than someone with ten times the audience, because the map is about position, not volume.

Why the coordinates are editorial estimates

Here is the lede I will not bury, because it makes the whole thing honest. The degree and betweenness coordinates are qualitative judgment calls, not measured follower counts and not a centrality score computed from real interaction data.

Each position is Claude’s informed read of where a voice sits, from watching who cites whom, who shows up across communities, whose ideas travel. So treat it as a model, not a measurement: an opinionated argument about the shape of the field, not analytics you could reproduce by querying a platform.

And treat it as falsifiable on purpose.

If you think someone is misplaced, you are disagreeing with my judgment, which is exactly the conversation the map is meant to start.

Say you think a researcher I parked deep in Specialists belongs in Bridges, because her papers are what three communities keep citing even though she never posts. That is a claim that her betweenness is higher than I drew it, the kind of correction that moves a dot.

Make the case and I’ll consider it.

I would rather ship an explicit, arguable model than a “data-driven” ranking whose objectivity is mostly costume. The numbers in most influence lists are real; the meaning attached to them usually is not.

How to read it

A ranking answers one question: who is ahead. It is a satisfying question and almost always the wrong one, because being ahead on reach tells you nothing about whether a voice moves ideas between worlds or only shouts inside one.

That is the question the map is built to answer, and it is why the map is worth more than the list it could have been.

So read it as a map. Where someone sits says more than how high.

The dot near the edge with one long thread into another cluster is doing work the loudest node in the field cannot do, and the two should never be confused for each other, whatever their follower counts say. The coordinates that place them are my judgment, set down in the open, which means you can disagree with them out loud. Disagreement is the point, not a defect.

What you get from all of this is not a verdict on who matters most. It is a sense of how the field is wired, where its seams run, and which way it is moving this month. That is the more useful thing to carry. Find the seams, and do not keep score.

Why this matters beyond the map

Strip away the specifics and this is an argument about how to read influence anywhere: every field has its leaderboards and its follower counts, its single numbers that promise to settle who matters.

Those numbers are seductive because they are simple.

A single number is simple because it has thrown almost everything away. It can rank, and ranking feels like knowledge, but it cannot tell you whether the person it puts on top moves ideas between worlds or only repeats them loudly inside one.

Klout learned that the expensive way, and its descendants are still everywhere. Substack, Medium, X, BlueSky, Threads, Instagram, TikTok. A single number.

So the next time someone hands you a ranked list of people, the useful instinct is not to ask who is on top. It is to ask what the list flattened to produce that order: whether the name at number three reaches a different world than the name at number one, or merely shouts louder inside the same one.

A map cannot answer every question, and this one is an argument you can dispute. It keeps that question open, where a ranking quietly closes it; that is the whole case for drawing one instead of counting.

For twenty years, information architecture was the discipline nobody funded. The job titles disappeared—I was one back in the day—and now it’s an art (and science) that needs to return. Teams didn’t align. Companies started showing their underwear.

Doing it well is less about structure than about getting an organization to agree on one way to name and arrange things. Every team brings its own vocabulary, so aligning them is slow, political, and thankless, and it loses every quarter to whatever ships a feature. The work stayed invisible, and so did its failures.

Then AI arrived, and the invisible foundation showed up on the balance sheet.

The same gaps that once cost you a confused visitor now cost you a hallucination, a wrong answer, an agent that confidently retrieves garbage and acts on it. In a 2025 survey, data quality and availability top the list of AI adoption barriers, outranking every other obstacle. The failures didn’t get worse; they got visible, repeatable, and expensive.

Simply said, poor information architecture can now be measured in token costs. A lot of them.

You can buy a better model, tune a sharper prompt, and stack another evaluation layer, and you will still serve wrong answers, because the model retrieves from a pile nobody agreed how to organize.

You need to fix the foundation before you pay for it a second time.

Retrieval Is Only as Good as What It Starts With

Most teams treat retrieval as a solved problem. Point the model at your content, let it pull the relevant passages, generate an answer.

The technique has a name — retrieval-augmented generation, or RAG — and a foundational paper, Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, and the mechanism is exactly what it sounds like: the model grounds its answer in documents it fetches at query time instead of relying only on what it memorized in training, limiting the retrieval set.

That mechanism inherits every flaw in the pile it fetches from.

If your content store is a flat heap of unstructured, unlabeled, contradictory documents, retrieval doesn’t work. It finds the loudest match — the one that shares the most surface words with the query, regardless of whether it is current, authoritative, or true.

It finds the loudest match — the one that shares the most surface words with the query, regardless of whether it is current, authoritative, or true.

Information architecture is what makes “the right one” a thing that exists in the first place. Organization schemes, labeling, and navigation aren’t decoration on top of content; they are the difference between a store where the correct answer is findable and a store where it sits buried next to four contradictory drafts of itself. The model can’t reach past the structure you gave it. If you never gave it one, it reaches for whatever is nearest.

This is the foundational claim, and everything else stacks on it. A retrieval system is a findability system. You have been building — or neglecting — findability systems for years.

Structure Is How a Model Knows What Something Is

A price, a policy, a deprecated note, and a customer quote can read as similar strings of text and mean opposite things. A person skims the surrounding page and knows which is which. A model, handed the same four snippets with no scaffolding, sees four passages of plausible language and no reason to rank one above another.

Metadata, taxonomy, and content typing are what let a system tell authoritative from anecdotal, current from expired, official from speculative.

This is the work Lou Rosenfeld, Peter Morville and Jorge Arango spent careers formalizing in Information Architecture: For the Web and Beyond — organization, labeling, and the semantic structures that let people, and now machines, understand what a piece of content is before acting on it.

The industry has a newer name for that same scaffolding — the semantic layer, the governed set of definitions, types, and relationships that sits between raw content and the systems reading it.

Call it a taxonomy, an ontology or a semantic layer; the job is identical, and it is the job of telling a machine what a thing is and how it relates to other things.

It isn’t reasoning about your content; it is pattern-matching over it.

Two mechanisms do most of that work. Controlled vocabularies — a fixed, agreed set of terms for the same thing — hold the language steady, so that “cancelled,” “canceled,” “terminated,” and “closed” don’t fracture one concept into four the system treats as unrelated.

They prevent the slow drift that creeps in when every team names things its own way, and they give the model a strict, consistent language to match against instead of a moving target. Structured metadata does the rest: it is the scaffolding the system leans on, the tags and fields that mark what is current, what is canonical, and how one thing relates to another, so retrieval has something to reason over besides raw text.

That vocabulary earns its keep on the input side, too. You can normalize the instruction before it reaches retrieval — mapping a user’s “refund,” “money back,” and “reimbursement” onto the one preferred term the content is filed under — so the query and the store speak the same controlled language instead of guessing at each other. Structure the content, then bring the question to it.

Strip that scaffolding away and the model can’t separate signal from noise, because nothing ever told it which was which.

This escalates the first claim. Reason one was about finding the right document. This is about the system understanding what the document even is — and understanding is not something you can prompt your way into after the fact. It has to be in the structure.

Ambiguity Doesn’t Get Resolved — It Gets Amplified.

People are forgiving of messy structure. Faced with a vague label or a half-right category, a person uses judgment to fill the gap, works around it, and moves on. That tolerance hid the cost of bad information architecture for two decades. It also let teams believe the mess was fine, because a human was always there to absorb it.

Models remove the human from that loop. They don’t fill gaps with judgment; they pattern-match over the mess and reproduce it at scale, confidently.

They don’t fill gaps with judgment; they pattern-match over the mess and reproduce it at scale, confidently.

A vague label a person would have quietly corrected becomes a systematic wrong answer served to thousands who have no idea the source was ambiguous. The World Quality Report 2025 found hallucination and reliability concerns among the top barriers enterprises named, with only 15 percent reporting generative AI deployed at enterprise scale.

This is the turn in the argument.

Bad information architecture used to cost you one confused user at a time, a cost so diffuse nobody bothered to name it.

Now it costs you an automated, repeated, confident failure — the same wrong answer, generated fresh for every person who asks.

The mess didn’t change. The blast radius did.

Agents Need Structure to Act, Not Just to Answer

Retrieval is the easy case. Answering a question wrong is embarrassing; taking an action wrong is a liability. The moment an agent stops retrieving and starts doing — routing a ticket, updating a record, approving a request, calling another system — it needs more than the right passage. It needs to know relationships, hierarchies, and boundaries: what belongs to what, what depends on what, what it is allowed to touch and what it is not.

That is information architecture functioning as an operating model, not a filing system — which is what the semantic layer is meant to be for agents: the governed encoding of what belongs to what, what depends on what, and what an agent is allowed to touch.

Agents need more context than humans and the semantic layer provides that.

A taxonomy that once organized help articles now governs which actions are valid against which objects. I made this case at length in an earlier piece on the semantic layer as agent infrastructure — the structures we built to make content findable are the same structures agents need to act safely.

A wrong answer erodes trust. A wrong action moves money, changes records, and triggers systems downstream that assume it was correct.

An agent with no model of these relationships doesn’t refuse to act. It acts anyway, on the flat and ambiguous picture it was handed, with the same confidence it brings to everything. A wrong answer erodes trust. A wrong action moves money, changes records, and triggers systems downstream that assume it was correct.

This already has a price tag, even before an agent acts. In Moffatt v. Air Canada, the airline’s support chatbot told a grieving customer he could claim a bereavement fare retroactively — the opposite of what Air Canada’s own bereavement policy page said. The chatbot even linked to the page that contradicted it. When the customer relied on the answer and was refused, a British Columbia tribunal held the airline liable and ordered it to pay, rejecting the claim that the chatbot was a separate entity responsible for its own words.

Two pages of one site said opposite things, and nothing reconciled them. That was a chatbot that only answered. Give an agent like it the power to act, and the same unreconciled structure starts moving money on its own.

This raises the stakes one more level. Reason three was about wrong answers at scale. This is about wrong actions at scale — and actions don’t ship with a disclaimer that the underlying structure was a guess.

This Is the Reframe That Finally Gets It Funded

Every reason so far converts a formerly invisible cost into a now-visible outcome. Findability becomes retrieval accuracy. Typing becomes hallucination rate. Structure becomes agent reliability and time to deploy. The discipline didn’t change. The balance sheet did.

That reframe changes who is willing to pay for it. Information architecture never won budget when its return was “fewer confused users,” a number nobody could put on a slide.

It wins budget when its return is “the AI initiative already on the roadmap doesn’t fail.”

Enterprises spent $37 billion on generative AI in 2025, up from 11.5 billion the year before, per Menlo Ventures’ State of Generative AI in the Enterprise — and much of that spend rides on retrieval and grounding that only work if the underlying content is organized.

The gain is measurable, not hand-waved: Anthropic’s Contextual Retrieval work found that adding the context that situates each chunk before indexing it cut failed retrievals by up to 49 percent — a reliability number produced by fixing content, not by swapping models.

That is the semantic layer earning a budget line: structure priced as retrieval accuracy rather than as tidy content.

Fund the foundation, or keep paying for it downstream in failed retrievals, eroded trust, and pilots that never reach production.

So stop framing information architecture as hygiene and start framing it as the precondition it has become. It is no longer a librarian’s luxury or a cleanup task that slips every quarter. Fund the foundation, or keep paying for it downstream in failed retrievals, eroded trust, and pilots that never reach production.

The librarian’s work didn’t get more important because the librarians argued better. It got more important because it started showing up in the numbers executives already watch.

None of this means the model never matters. Sometimes it really is the weak link — the reasoning is too shallow for the task, or the instruction is vague enough that no content could save it, and a better model or a sharper prompt is the right fix. That case is real. It is just rarer than the spend implies, and you cannot tell the two apart by staring at the output: a wrong answer from a weak model and a wrong answer from a capable model reading a disorganized store look identical on the screen. The only way to know which one you have is to trace what the system retrieved — which is why the audit comes before the purchase, not after.

Where to Start Before You Buy Another Model

The fix isn’t a purchase; it’s an audit and a sequence. Before you approve the next model, tool, or evaluation layer, do the unglamorous work the last decade let you skip.

Here is the whole method in miniature. Take one wrong answer your assistant produced this month and pull the passages it retrieved to build it. You will usually find a plausible-looking source that was outdated, duplicated, or mislabeled, and nothing in the store told the model so.

It has already played out in public: when Google’s AI Overviews recommended putting glue on pizza — a tip a model lifted from an old forum joke because nothing marked the joke as anything but a keyword match — that was this exact audit, failed at planetary scale, as Forbes traced in Google’s AI Recommends Glue on Pizza.

Fix that one source, and you have fixed every future answer that would have leaned on it.

• Audit what your AI retrieves from, not just how it answers. Trace a wrong answer back to its source. Nine times out of ten the model did its job and the content store handed it garbage.

• Type your content before you index it. Tag what is current, authoritative, and canonical — and what is deprecated, draft, or anecdotal — so retrieval can tell them apart instead of averaging them together.

• Fix labels and taxonomy before you fix prompts. A prompt patch papers over one query; a corrected taxonomy fixes every query that touches it. Structure scales, and prompt tweaks don’t.

• Give agents an explicit model of relationships and permissions. Before an agent acts, it has to know what belongs to what and what it is allowed to touch. Write that down as structure, not as hope.

• Put an information architecture cost on an AI outcome. Translate “cleaner content” into “higher retrieval accuracy” or “lower hallucination rate” so the work competes for budget on the terms executives already track.

The Foundation Was Always Load-Bearing

For twenty years, the case for information architecture was aesthetic and moral: do it because findability is good, because users deserve clarity, because sloppy structure is unprofessional. Those arguments were true, and they lost, every quarter, to whatever shipped a feature. The discipline stayed underfunded because its failures were invisible and its wins were unprovable.

That argument is over. Not because anyone won it, but because the ground moved. AI turned every structural shortcut into a measurable failure — a hallucination you can count, a wrong answer you can reproduce, an agent you can watch act on bad information.

The invisible foundation now has a profit-and-loss line, and it is bleeding.

You don’t have an AI problem. You have an information architecture problem wearing an AI costume.

The good news buried in that sentence is that you already know how to fix it. The organization schemes, the labeling, the metadata, the content types — the whole toolkit has been sitting in the discipline the entire time, waiting for a reason anyone would fund it.

Audit what your AI retrieves from before you buy another model. The foundation was always load-bearing and now everyone can see the load.

Disclaimer: I am not a lawyer; I don’t even own a suit. I did spend eight years in legal technology, which qualifies me to play one poorly on Law and Order and nothing further. This not about the law, it’s about trust. Read on.

In November 2025, Raza Khan, a startup founder, sued Figma in the Northern District of California.

The complaint does not lead with copyright — it leads with consent, and with the claim that the company switched on model training over customer design files by default after years of promising otherwise.

Nothing in that case has been decided.

Figma disputes the allegations and says its training focuses on general patterns like, uh, creating weather apps, rather than customer content.

Hold the legal question open, because that is where it belongs. The design question is not open, and it is not really a question about Figma. Somebody chose between opt-in and opt-out, and chose differently by plan tier. And that decision sits in public documentation right now.

Here is the part that matters: Designers are the users in this story of a company lead by, and we know precisely how it feels to have a default decided on our behalf.

Dylan Field and his company should have a higher standard because where they sit in the public square; they are teaching designers through their actions.

We ship those same defaults to our own users every quarter. So this is a test with four questions, about twenty minutes of work, and it only counts if you turn it around so the users are treated with empathy and respect and not dark patterns.

The Pattern Rhymes

This has happened enough times to have a shape, AI or not. These are the AI cases.

In the summer of 2023, Zoom amended its terms of service to claim broad rights over customer data for training and tuning models. Nobody noticed for months. When a technology blog surfaced the clause in August, The Record reported in Zoom revises terms again to say it doesn’t use customer data to train AI models that the company rewrote the language twice inside a week.

Nine months later it was Slack. A post on Hacker News pointed at the privacy principles page, and workspace owners learned they were enrolled by default in machine learning training on messages, content, and files. TechCrunch covered the reaction in Slack under attack over sneaky AI training policy, including the detail that turned irritation into anger: opting out meant emailing a specific address with a specific subject line, and only a workspace owner could send it.

No toggle, no setting.

The terms had been live since at least September 2023, and Slack rewrote the language within days while saying it had changed no practice, which was true and beside the point. Fresh ink on their hands.

This has happened enough times to have a shape, and it doesn’t feel good.

Then June 2024. Adobe pushed a re-acceptance modal, creators read the license language, and the company spent two weeks walking it back. Its own post, Updating Adobe’s Terms of Use, conceded the terms needed to be more precise.

Adobe’s defaults were never the problem; its wording was. The company lost weeks of trust without doing the thing it stood accused of, which tells you that trust does not track behavior. It tracks what people can read and verify.

Slack is its own category, and the one worth studying. Zoom and Adobe reversed. Slack did not: it rewrote the sentence and kept the mechanism, which is a different move wearing the same clothes.

Trust is not a feeling, it is a set of defaults you can audit, and Zoom and Adobe both recovered inside a fortnight because reversal was available to them — a sentence to fix, a toggle to flip, a correction to make in public. Slack fixed the sentence only, and the email address is still the exit.

Same script every time.

• A policy change lands quietly.

• Enrollment is the default.

• Somebody outside the company finds it.

• Then the clarification, the apology, and sometimes the walk-back.

It feels a lot like Fight Club, where the companies are calculating risk and reward.

“My job was to apply the formula: Take the number of vehicles in the field, A, multiply it by the probable rate of failure, B, multiply the result by the average out-of-court settlement, C. A times B times C equals X. If X is less than the cost of a recall, we don’t do one.”

Sounds like consent spirit, read like this:

“My job was to apply the formula: Take the number of users whose data we ingested, A, multiply it by the probable rate of anyone noticing, B, multiply that by the average settlement per claim, C. A times B times C equals X. If X is less than the cost of building a real consent flow, we don’t build one.”

What is different about Figma is that nothing was walked back at all, and the discovery arrived as a federal complaint. There was no drafting error to fix. The tiered default was announced, explained, defended with a stated rationale, and allowed to take effect on schedule. You cannot clarify a decision that was already clear.

The Sniff Test

Four questions fall out of that pattern, worth naming with their scoring before we run them.

1. Consent

2. Symmetry

3. Disclosure

4. Exit

A tool that passes all four treats your work as yours, and three passes is ordinary enough to be worth raising at renewal. But symmetry does not weigh the same as the rest, and a failure there counts for more than the other three combined.

Test One: Consent

The question is not whether a tool told you. It is whether the switch was off when you got there.

Figma’s announcement, Meet Figma AI, is unusually direct. Published in June 2024, it states that sharing customer content for training is optional, that admins set the preference, and that Starter and Professional plans are opted in by default, with no training before August 15, 2024. That is seven weeks of notice, published openly, with the mechanism described.

That is more transparency than Zoom or Slack managed.

Slack buried its terms in a privacy policy and answered discovery with an email address.

Figma published a date.

It is still opt-out.

Notice is not consent.

A seven-week window before a default takes effect is a courtesy, and a real one, but it puts the burden of refusal on the person whose work is at stake. Consent that has to be revoked was never asked for. It was assumed, announced, and then made cancellable.

Consent that has to be revoked was never asked for. It was assumed, announced, and then made cancellable.

The research on this is settled. Nielsen Norman Group’s Sneaking: The Deceptive UX Pattern You Never Saw Coming catalogs the family of practices that get someone to agree to something they never intended, and the finding is not an ethical one so much as an accounting one: sneaking converts, and it costs long-term trust at a rate that outruns the gain.

Designers know the pattern because we build it — the pre-checked box, the bundled permission at install, the subscription folded into account creation. We have spent two decades publishing research on why these patterns are hostile, then shipping them anyway when a growth target needed hitting.

has spent the last stretch of his career arguing this is the job rather than a footnote to it. Design for a Better World presses designers to widen the frame past the person at the screen and treat a shipped decision’s consequences as their own. A default is exactly that kind of decision. It gets made once, by a few people in one room, and applies to millions who will never meet them.

The standard that produced the opt-out default was whether it would survive legal review. Ours is narrower and harder: what is right for the person whose work this is. The two agree most of the time, which is why the gap goes unnoticed until a quarter when they do not.

Run it this way. Open the tool, find the training setting, and look at its state.

If it is on and you did not turn it on, the company made a decision about your work and told you afterward.

Test Two: Symmetry

Here is where it stops being an ordinary default and starts being a statement. Figma’s engineering explainer, Building Figma AI, lays out the tiering plainly.

Content training defaults to on for Starter and Professional, off for Organization and Enterprise. That’s a pretty clear statement and the stated reason is that agreements with Organization and Enterprise are typically more complex and include specific requirements and restrictions.

Read that again, because it is the most important sentence in the whole episode and Figma published it voluntarily. What it says is this: we understood the setting carried contractual risk, so for customers who had negotiated terms making that risk explicit we defaulted to off, and for everyone else we defaulted to on.

Then read the consequence, which sits in Figma’s own registration statement: roughly seventy percent of new Organization and Enterprise customers included at least one person who had previously been on a Professional plan. The protected tier is grown from the enrolled one. The designer whose files trained the models on a Professional account is the same designer who later brings her company onto an Enterprise contract and receives the protective default as a welcome gift.

The asymmetry is not evidence of bad faith. It is evidence of accurate risk assessment.

Bad faith is a mistake you can apologize for. This was a correct read of where leverage sat, followed by placing the exposure where it was least likely to generate a lawsuit.

That calculation was wrong by about seventeen months.

The underlying premise does not hold either. A freelance designer working under a client nondisclosure agreement is not carrying less confidentiality risk than a company with a master services agreement. She carries the same risk with a thinner contract and no procurement team. The obligation is identical.

The direct version: this was not morally right, and it does not respect designers.

Set the litigation aside and assume Figma wins outright. The tiering still sits in the published record, put there by the company itself: customers with procurement departments get protected, customers without them get enrolled. That is a judgment about whose work matters, made by a company whose business is the work of designers.

Leverage is not a moral category. Bargaining power tells you what a customer can force you to do, and nothing at all about what you owe her, which is why we do not let doctors calibrate care to a patient’s ability to sue.

Symmetry is the best question in the test because it is hardest to explain away. A single default applied to everyone might be a philosophy. Two defaults split along a payment line approved by legal is a position on who is worth protecting.

Test Three: Disclosure

The third question is whether a working designer, doing her job, would have found out without somebody else telling her.

Blog posts are not product notifications. Figma published openly and design press covered it, but the audience for an engineering blog is a fraction of the audience for the product, and the people least likely to read it are the people the default applied to: individual designers and small teams with no administrator tracking vendor policy changes.

Bloomberg Law’s account, Figma Trained AI on User Data Without Consent, Class Action Says, frames the plaintiff’s central allegation as a gap between years of assurances and a settings change. The docket is public at Khan v. Figma, Inc.. The claims are unproven, and reasonable people can read the notice period as adequate.

The disclosure test is not about whether a company published something. It is about whether the publication reached the person the change applied to.

You can watch this fail in Figma’s own forum. In a thread titled Opting out of AI Features, one designer called the arrangement “sketchy as hell” and said the way it was done felt disrespectful.

Another asked, plainly, whether her work was being used given that she could find nowhere to opt out. A third wrote that he had not logged in for months and was learning about it only then.

Those are not journalists reading a policy. Those are users finding out.

So here is the standard, stated plainly enough to apply. A change to what a tool does with your files should reach you inside the tool, once, before it takes effect, with nothing pre-selected. Anything clearing all four conditions is disclosure. Anything clearing three is publication wearing the same coat.

Measured that way, Figma cleared two of the four. The notice arrived before the effective date and it was published once. It never arrived inside the product, and the choice was pre-selected when it did. There is a version that would have passed cleanly: an in-product modal, shown once, two buttons, no default selection. Companies build exactly that for cookie consent because regulation forced them to.

Test Four: Exit

Disclosure only matters if it leads somewhere. The last question is whether you can leave, and what leaves with you.

Figma’s documentation, Manage AI settings and content training for your team or organization, answers precisely. Content training took effect August 15, 2024. If an admin turns the setting off after that date, new content and edits will not be used to train models.

New content and edits. Forward-looking only. And for a while, not everything was covered.

Designers in that same forum thread pointed at documentation stating that opting a team out did not reach an individual user’s personal drafts, which on Starter and Professional could not be excluded at that time.

Personal drafts are where the unfinished work lives — the pitch nobody approved, the client concept that never shipped, the side project. The one place a designer would most reasonably assume was hers was the one place the switch did not reach.

An exit that requires an administrator is not an exit for the person whose work it is.

That is a reasonable engineering position. Unwinding a trained model to extract one team’s contribution is a research problem, not a settings toggle. But it means the exit is partial by construction, and its value decays every day you wait.

Then there is who holds the switch. It is an admin control, team-level on Starter and Professional plans. An individual contributor cannot turn it off. She can ask. If the team lead is busy or unconvinced, the answer is no by inaction. An exit that requires an administrator is not an exit for the person whose work it is.

Test it like any recovery path. Can the person harmed by the state change the state? How long does it take? What is unrecoverable afterward? Ordinary usability questions. We just do not point them at the terms of service.Permission and Forgiveness

The industry has an operating rule, and everyone reading knows it. Better to ask forgiveness than permission. It has been repeated in enough keynotes to pass for wisdom, when what it describes is a wager: that the cost of being wrong will land later, on somebody else, in pieces small enough to absorb.

A default-on training setting is that wager, placed in a settings panel.

Three regimes now govern it, and they do not read it the same way.

• California codified forgiveness. The California Consumer Privacy Act gives consumers the right to opt out of the sale or sharing of personal information, alongside notices explaining what a business does with it. Notice first, then a switch, which is the shape Figma’s design sits comfortably inside. One detail is worth carrying: California’s regulations define a dark pattern as an interface that substantially impairs autonomy or choice regardless of intent, and require that opting out be no harder than opting in. Symmetry is not only a design principle. In California it is a rule.

• Europe codified permission. Under the General Data Protection Regulation you need a lawful basis before processing, not a notice afterward. The European Data Protection Board’s Opinion 28/2024 on personal data in the context of AI models confirms legitimate interest can serve as that basis, subject to a three-part assessment. It also states the consequence designers should sit with: data processed unlawfully during development can follow a model into deployment, carrying erasure obligations with it. Forgiveness there can reach the model.

• The EU AI Act codified disclosure. Then the calendar moved. As Gibson Dunn’s summary of the Digital Omnibus agreement sets out, high-risk obligations slipped to December 2027 for stand-alone systems and August 2028 for embedded ones. Most Article 50 transparency obligations still apply from August 2, 2026. The wider package also proposed amendments to the General Data Protection Regulation aimed at making training on European personal data easier.

Read the scoreboard from a designer’s chair. The industry asked forgiveness at scale, and in the largest regulatory market it is being granted, partly as time and partly as amended rules. Regulation is not going to settle whether default-on training is acceptable. The question lands back with whoever sets the default.

Leave a comment

Running the Test on Your Own Stack

Anil Dash wrote the clearest version of the principle in The Internet of Consent: technology should only ever do what we have explicitly given it permission to do. That standard is testable, product by product, and most teams have never tried.

The first four items protect you. The fifth is the one that matters.

• Inventory where the work lives. Eight minutes. List every tool holding files your team produces, including the ones nobody formally approved. The tools nobody approved are usually the ones with the most permissive defaults, because nobody read the terms before signing up.

• Read the default, not the promise. Six minutes. Marketing pages describe intentions. Settings panels describe behavior. Open the toggle in the actual account and record its state today.

• Check the tier line. Three minutes. Find out whether the vendor treats enterprise accounts differently from yours. If it does, you now know which side of that line you are on.

• Screenshot the settings with a date. Three minutes. Defaults change quietly and version histories are rarely public. A dated screenshot is the only evidence you will have that a setting changed.

• Run the test on your own product. One design review. This is the part designers skip, and the reason the test exists. If your company ships a feature that learns from customer content, somebody chose a default and you may have been in the room. Kronenberger Rosenfeld’s Figma Lawsuit Opens Door to New Damages Claims notes the exposure runs through the gap between what a company says and what its product does.

The tools nobody approved are usually the ones with the most permissive defaults, because nobody read the terms before signing up.

Then score it against the threshold set out earlier. A symmetry failure on its own outweighs the rest, and here is the reason promised back then: it is the only one of the four where a company assessed the risk correctly and decided you would carry it.

None of this requires an opinion about the litigation. The test works whether Khan wins, loses, or settles quietly next spring.

And do not assume the alternative is leaving, because switching cost is the point. The registration statement reports 132 percent net dollar retention as of the first quarter of 2025, among the strongest figures of any public cloud company.

Read how the number is built, though. As Figma S-1 | The IPO Bar is High sets out, the calculation counts customers above ten thousand dollars in annual recurring revenue who are still customers, so accounts that left entirely never appear in it.

It measures expansion among the people who stayed and cannot see the ones who went. The bill for this does not arrive as churn, and the instrument pointed at churn would miss it anyway. It arrives as interpretation, in the pause before somebody vouches for you.

Figma may win this case. The trade secret theory is aggressive, the class definition will be fought hard, and seven weeks of notice is a defense. I would not bet against the company.

Winning is not the same as being trusted, AI or not, and only one is a design outcome.

The dispute traces back to one product decision, made by people doing a job many of us have done. Nobody in that room needed to be cynical. They needed a training corpus, a legal review that cleared the language, and a notice period. That is how these decisions get made, and how we make them.

Defaults are the quietest thing we ship and the loudest thing we say. Opt-out says your work is ours until you object. Opt-in says it is yours until you offer it. The engineering cost between them is nothing. Permission costs a modal; forgiveness costs everything after.

So run the test. Then go run it on what you build, where your answer counts. Designers spent this year asking a vendor for consent it never thought to ask us for. We should not need a lawsuit to remember how that felt, or to stop doing it to somebody else.

Nir Eyal published Hooked in 2014, and product teams have been running its playbook ever since. It was a topic I taught in General Assembly, and have been fascinated with ever since because it’s at the base of many interactions, digital or not.

’s Hook Model builds directly on BJ Fogg’s work at the Stanford Behavior Design Lab, where Eyal studied. Fogg’s Behavior Model (B=MAP) holds that behavior happens when motivation, ability, and a prompt converge at the same moment.

Eyal absorbed this foundation — particularly Fogg’s emphasis on triggers and making actions as frictionless as possible — and extended it into a four-phase loop, adding variable rewards and investment to explain how products build habits over repeated cycles.

Building on B.J. Fogg’s behavior model, Hooked describes a four-step loop that turns a product into a habit:

• A trigger prompts you to act — You hit a wall drafting a tricky email to your boss (internal trigger: the discomfort of not knowing how to phrase it), or the assistant pushes a notification: “Your weekly digest is ready to review.” Either way, something nudges you.

• An action delivers a quick win — You type one line — “make this sound more confident but not pushy” — and get a polished draft back in two seconds. Minimal effort, immediate payoff.

• A variable reward keeps the outcome slightly unpredictable — Each response is a little different. Sometimes it just fixes the tone; sometimes it surfaces an angle you hadn’t considered or catches an error you missed. You can’t fully predict the quality or the surprise, so you keep pulling the lever.

• An investment loads the next trigger — You correct the draft, save a preference, upload your style guide, or let it remember that you always sign off a certain way by creating instructions. That effort makes the next session better and quietly plants the reason to come back —a trigger—so your customized assistant is now more useful than a blank one.

This fits AI assistants and chat almost too well. For consumer, it pushes you into a higher subscription tier; for enterprise, it costs money per token.

Those tools already provide value, so the loop has real utility to wrap around. Value gets someone to try you once; the Hooked loop is the layer above it, turning a useful answer into a daily habit.

Triggers come in two kinds.

• External: The notification, the email, the app icon.

• Internal: Feelings that send you back unprompted — boredom, the nagging itch of an unfinished thought, a glass of whiskey.

Variable rewards are the engine: an uncertain outcome keeps you pulling the lever, which is why the model borrows from the slot machine. Investment is the quiet closer — every preference, file, and correction raises the switching cost.

And that investment? It’s helping the assistants build their models.

Understand the loop, and you see it running underneath nearly every AI product you touch.

Triggers: Move Users From the Notification to the Itch

Most teams reach for external triggers first, because they are the easy ones to ship.

Send a digest email.

Add a red badge.

Fire a push notification when the model finishes a task.

These work, for a while, and then they stop working, because people learn to ignore anything that arrives on someone else’s schedule. The durable trigger is internal. It is the reflex that sends you to a chatbot the moment you feel stuck, bored, or unsure — before any notification arrives.

That reflex has already formed at scale from habits of the past.

When Sam Altman said in October 2025 that ChatGPT had reached 800 million weekly active users, he was describing a lot of people who now open a text box the way an earlier generation opened a search box, or a phone book. Previous habits for a new world.

The design work is not adding another alert. It is finding the specific emotion or routine your product should attach to, then removing everything between that feeling and the first useful response so you don’t overwhelm the user.

If your only lever is the notification, you are renting attention. If you own the internal trigger, you stop paying rent.

Examples

• Duolingo — the streak counter and its notifications are relentless external triggers; the real design work is the emotion underneath (fear of losing the streak) that keeps users returning once the pings are muted.

• Slack — the unread badge starts as an external cue and hardens into the internal “am I missing something” reflex; attach to an anxiety users already have rather than inventing one.

• ChatGPT — people now open it unprompted the moment they hit a blank page; identify the specific stuck-moment your product should own. Some are using it more than Google.

Action items

• Map the internal trigger before the external one. Name the exact feeling — the blank page, the stalled decision, the half-formed question — that your product should answer. Design backward from it.

• Retire the notification that hasn’t earned its place. Every alert that gets dismissed teaches the user to dismiss the next one. Cut the ones that fire on your schedule rather than their need.

• Attach to a routine that already exists. New habits are cheaper to borrow than to build. Slot your product into the moment a user already reaches for something else.

Action: The Chat Box Is the Lowest-Friction Action Ever Shipped

Eyal borrowed the action stage from B.J. Fogg, whose model says a behavior happens when motivation, ability, and a prompt line up at the same moment. The action is the simplest thing a user can do in anticipation of a reward. Lower the effort, and the behavior fires more often.

By that measure the empty chat box is close to the theoretical minimum. There is nothing to learn, no menu to parse, no feature to find. You type a sentence, the way you would text a colleague, and something happens.

That is why it spread faster than any interface in the history of software.

There is nothing to learn, no menu to parse, no feature to find. You type a sentence, the way you would text a colleague, and something happens.

It is also why the friction did not disappear so much as move. Amelia Wattenberger made the case early, in Why Chatbots Are Not the Future: the blank box looks identical to a search field, a login form, and a credit-card input, so the burden of knowing what to type falls entirely on the user. The old friction was learning the interface. The new friction is knowing what to ask, trusting what comes back, and switching between telling the model what to do and judging whether it did it.

That is the honest reason the chat box won: it was the fastest thing to ship on the day the capability arrived, not the best interface anyone could design. The durable AI habits will belong to the teams that keep the low effort and quietly repair the friction that moved.

Examples

• ChatGPT — a single text box: near-zero effort, but no hint of what to type; lower the cost of the first good prompt with examples and sensible defaults.

• GitHub Copilot — the suggestion appears inline as you type and accepts with one keystroke; push the action below the box wherever the context is already known.

• Perplexity — suggested follow-up questions and visible sources are affordances built into the interface; carry context in the UI instead of making users retype it.

Action items

• Reduce the cost of the first good prompt by setting expectations. Show examples, defaults, and starting points so the user is not staring at a blank oracle wondering what it knows.

• Add the affordances the blank box hides. Sliders, toggles, and structured inputs can carry context the user would otherwise have to type — and remember to type — every time.

• Design for the evaluate step, not just the ask. The action is not finished when the user hits enter. It is finished when they can tell whether the answer is any good.

Rewards: A Language Model Is a Slot Machine by Default

Here is the part of the loop AI did not have to engineer, because it comes free with the technology. A language model is nondeterministic. Ask the same question twice and you get two different answers.

Eyal split rewards into three kinds, and AI hits all three.

• Rewards of the hunt are the information you were chasing — the answer, the summary, the fix.

• Rewards of the self are the sense of progress and mastery — a finished draft, a solved bug, a box checked.

• Rewards of the tribe are social — the clever output you screenshot and share.

Every generation is a fresh pull of the lever, and you never know in advance whether this one is brilliant, useful, or nonsense, hallucinations included.

The trap is that the same variability that hooks people is also the reliability problem.

A slot machine is supposed to be unpredictable. A tool is not, which presents some ethical challenges for AI.

When the unpredictability lands, users chase the next great answer; when it misses, they lose trust in all the answers. You cannot turn the variability all the way up without eventually turning credibility down.

So the discipline is to be deliberate about which reward you are really paying out, and to make the good outcome reliably reachable rather than occasionally miraculous.

Examples

• Midjourney — four images per prompt and a reroll button turn nondeterminism into a ritual; make the reward of the hunt clear and the good outcome reachable.

• ChatGPT “Regenerate” — each press is a fresh pull of the lever; use variability to help the user, not to string them along.

• Spotify Discover Weekly — an unpredictable, personalized set you open because you never know what you’ll find; pair the surprise with genuine relevance.

Action items

• Name the reward you are really offering. Hunt, self, or tribe — pick one on purpose. A product that pretends to offer all three usually delivers none of them well.

• Make the good outcome reachable, not just possible. Users forgive a model that is consistently helpful faster than they forgive one that is occasionally dazzling.

• Do not manufacture variability you cannot stand behind. Randomness that serves engagement but not the user is the first step from facilitator to dealer.

Investment: Every Iteration You Make Loads the Next Trigger

The fourth stage is the one that compounds. Investment is any effort the user puts in — time, data, preferences, content — that makes the next session better than the last because it adds context. Unlike a sales funnel, it has no endpoint.

It is stored value that improves the product without the company lifting a finger.

AI products are unusually good at this, because correcting the model is itself an investment. Every custom instruction, uploaded file, saved project, and thumbs-up quietly tunes the thing toward you. The model gets more yours, and the cost of starting over somewhere else climbs with each pass through the loop.

This is the endowment effect and the IKEA effect doing exactly what they always do: we overvalue what we have shaped and what we have helped build.

You can watch investment deepen in the enterprise numbers. OpenAI reported that usage of structured workflows such as Projects and Custom GPTs rose roughly 19-fold across 2025 — the shift from casual questions to configured, repeatable setups people do not want to rebuild elsewhere.

The catch is that investment cuts both ways. Lock-in that the user chose feels like ownership. Lock-in they did not choose feels like a hostage situation, and the moment a competitor is meaningfully better, resentment does the switching for them.

Examples

• ChatGPT memory and custom instructions — every correction tunes the assistant toward you; make that accumulated value visible so users feel the return.

• OpenAI Projects and Custom GPTs — configured, repeatable setups people will not rebuild elsewhere; let power users see and export what they have built.

• Notion — the workspace you have filled with your own documents and structure is the switching cost; earn the data with a real win before asking for more.

Action items

• Make the investment visible so users feel the return. People keep investing when they can see the memory, the customization, and the saved work paying off in front of them.

• Let users port what they have built. Confidence that they could leave is often what convinces them to stay. Export is a retention feature, not a leak.

• Earn the data before you ask for it. Request the investment after the first real win, not before it. Value first, effort second.

The Manipulation Matrix: Run Eyal’s Own Ethics Test Before You Ship

Eyal knew the loop could be pointed at harm, so he built a gate for it. The Manipulation Matrix asks two questions:

• Does the product materially improve the user’s life?

• And would you use it yourself?

Here’s the quadrant:

• Answer yes to both and you are a facilitator. You are building something you would use and believe in, which is the only quadrant where you get to run the techniques at full strength.

• Yes to improving lives but no to using it, and you are a peddler. You usually headed for failure because you lack the empathy to build something people want.

• Yes to using it but no to improving lives, and you are an entertainer. This is fine but it is art, not value.

• No to both is the dealer, and Eyal has a plainer word for it: exploitation.

No to both is the dealer, and Eyal has a plainer word for it: exploitation.

There is a second test, sharper than the first. Eyal calls it the regret test: if users knew everything you know about what is about to happen, would they still do this?

An assistant that saves someone an hour passes easily; an engagement-maximized companion tuned to keep a lonely user talking past midnight does not, and no amount of retention lift changes the answer.

Run these before the roadmap review, not after the lawsuit.

Examples

• Slack and WhatsApp — Eyal’s facilitators: the founders used what they built and believed it helped; the honest quadrant is the one you can occupy without a preamble.

• Gamified fitness and charity apps — Eyal’s classic peddler: badges bolted onto tasks the makers themselves would not do daily; if you would not use it, research hard before shipping it.

• Engagement-maximized AI companions — the dealer risk: tuned to keep a user talking with unproven benefit; run the regret test on the retention feature before it ships.

Action items

• Put your product in a quadrant, out loud. Say the words facilitator, peddler, entertainer, or dealer in the room. Ambiguity is where exploitation hides.

• Apply the regret test to the retention feature. Not to the product in the abstract — to the specific mechanic you are shipping to make the number go up.

• If you are writing an “if I were the user” preamble, stop. Eyal’s rule is direct: if you have to imagine yourself into the user’s shoes to justify it, you already failed the test.

Where the Loop Breaks: Habit Is Not the Same as Value

The model has one blind spot, and AI walks straight into it. The loop measures engagement, not benefit. It counts whether people came back, not whether they should have. With social media, the worst common case was wasted time. With AI systems built to simulate a relationship, the stakes are higher, because the product is designed to feel like a friend, a confidant, or a partner.

Start with what the loop counts. Every number it produces — return visits, sessions, messages sent, minutes on the platform, tokens burned — measures consumption. And consumption is not an outcome. It is an activity, the input a user spends in the hope of getting something, not proof that they got it. A person can pour an hour into your product and walk away with nothing solved, and every dashboard you own will record that hour as success.

In an AI product the problem runs deeper, because a rising consumption number is often the tool failing louder: the re-asks, the re-prompts, the three mediocre answers a user wades through to reach one usable one all register as engagement. Every consumption metric leaves the only question that matters — did this help? — wide open.

You can watch both signals sit in the same report.

OpenAI’s enterprise data shows consumption climbing fast — weekly messages up roughly eightfold and reasoning-token use per organization up hundreds of times across a year. That tells you how much the tool gets used. The figure that carries value is a different one in the same document: workers reporting 40 to 60 minutes saved per day. Time saved is an outcome. Messages sent is consumption. Only one of them answers the question you were asking.

A person can pour an hour into your product and walk away with nothing solved, and every dashboard you own will record that hour as success. That’s consumption over valuable outcome.

That is not a hypothetical.

In late 2025, Character.AI removed open-ended chat for users under 18, effective November 25, after lawsuits and regulatory pressure over teen harm, including wrongful-death claims. A company built entirely on the loop conceded, in public, that for some users the loop had grown too strong to leave running.

The direct version: if your retention depends on a variable-reward loop that users would switch off the moment they understood it, you did not build a habit, you built a trap.

This is the same lesson behind the old “worse is better” pattern — products win on utility, not novelty, and the AI habits that last will be the ones that make a real job faster, not the ones that keep a user pulling the lever. Compulsion retains people until it doesn’t; utility retains them because it works.

There is even a self-interested reason to care: the variability that rewards a user is the same variability that erodes trust, and a hooked user who stops believing the output leaves anyway. Build the honest version. It is the one that survives.

Examples

• Character.AI — removed open-ended chat for under-18 users after harm and lawsuits; a loop strong enough to require shutting off is a design failure, not a metric win.

• Infinite social feeds — Instagram and TikTok optimized engagement while the value question stayed open; more time on the platform is not more benefit to the person.

• AI usage dashboards — message and token counts measure consumption, not outcomes; instrument time saved and tasks finished instead.

Action items

• Count outcomes, not consumption. Messages, minutes, and tokens measure how much your product got used, never whether it helped. Instrument the result — the task finished, the time saved — or you are guessing about value.

• Build the graceful exit before you need it. Limits, summaries, and natural stopping points signal a facilitator and are more valuable. Their absence signals a dealer.

• Separate “they came back” from “it helped.” Treat them as two different numbers, because the whole failure mode lives in the gap between them.

Conclusion

The Hooked Model is not a trick, and it is not a villain.

It is a description of how habits form, and it works the same whether you point it at a fitness app, a feed, or a language model.

The four stages have not changed since 2014 — what changed is that AI arrived with two of them pre-installed: the variable reward comes almost free with a nondeterministic model, and the action collapsed to a single text box anyone can use, even if the expectations aren’t set.

That is why the loop runs faster and stickier in AI than in almost anything before it, and why the ethics stopped being optional. When the reward is automatic and the action is frictionless, the only real brake left is the one you choose to build.

So use the model, all four stages of it, without apology. Attach to a genuine internal trigger. Keep the action simple. Be deliberate about the reward. Let users invest in something they can carry with them. Then run Eyal’s own two questions and the regret test, and be honest about the answers.

Build the facilitator. The dealer’s numbers look better for exactly one quarter, then the trust runs out and takes retention with it.

Recommendations

The full playbook in one place, by stage.

Triggers

• Map the internal trigger before the external one — notifications rent attention; the internal reflex is what returns unprompted.

• Retire notifications that haven’t earned their place — every ignored alert trains users to ignore the next one.

• Attach to a routine that already exists — borrowed habits are cheaper to form than new ones.

Action

• Reduce the cost of the first good prompt — a blank box hides what the tool can do; examples and defaults lower the effort.

• Add the affordances the blank box hides — structured inputs carry context users should not have to retype every time.

• Design for the evaluate step, not just the ask — the action is not done until the user can judge the answer.

Variable rewards

• Name the reward you are really offering — hunt, self, or tribe; pretending to offer all three delivers none.

• Make the good outcome reachable, not just possible — consistent help earns trust faster than occasional brilliance.

• Do not manufacture variability you cannot stand behind — randomness that serves engagement over the user is the first step toward the dealer.

Investment

• Make the investment visible so users feel the return — people keep investing when they can see the payoff in front of them.

• Let users port what they have built — confidence they could leave is often what makes them stay.

• Earn the data before you ask for it — request effort after the first real win, not before it.

Manipulation Matrix

• Put your product in a quadrant, out loud — ambiguity is where exploitation hides.

• Apply the regret test to the retention feature — test the specific mechanic, not the product in the abstract.

• Drop the “if I were the user” preamble — needing it means you already failed the test.

Where the loop breaks

• Count outcomes, not consumption — messages and minutes measure use, not whether it helped which takes qualitative research.

• Build the graceful exit before you need it — natural stopping points mark a facilitator; their absence marks a dealer.

• Separate “they came back” from “it helped” — the whole failure mode lives in the gap between those two numbers.

On Monday, July 13, 2026, Figma’s stock jumped nearly 12% in a single session.

There was no product launch, no earnings surprise, no announcement anywhere worth the move. Just under half the available shares had been sold short, and the bears got squeezed. Designers watched the ticker anyway, the way you watch a jury file back into the room.

The number felt like a verdict being handed down on the craft itself.

It wasn’t.

The market cannot try your craft, because your craft was never the defendant. What is on trial is the design layer — the software, the seat-based pricing, the question of whether a canvas can stay expensive when a chatbot can draw.

This is going on not only in design, but in all of SaaS software — what the craft is for each domain versus the tool they are using.

Those are business questions and they deserve business answers. Watch how quickly a swing in a share price gets read as a referendum on whether design still matters.

This trial has been held before; every few years a tool that owned the work gets dragged into court, and the jury confuses the tool’s fate with the discipline’s. The tool loses and the discipline walks out free.

If you have been in this room before, you already know how the verdict reads.

The Evidence In Front Of The Jury

Start with what happened, because the facts are stranger than the fear. The Motley Fool’s Why Figma Stock Jumped 11.9% Today laid out the mechanics: nearly 42.4% of Figma’s tradable shares were sold short heading into the session, one of the heaviest bets against any technology stock.

A single institutional filing tipped the balance, the bears rushed to cover, and the price shot up on mechanics, not news. That is a short squeeze and it tells you about positioning and float, but it tells you nothing about pixels.

The week before, the mood music was already shifting. The trade press framed it as Figma stock jumps 6% as BofA backs AI outlook, with Bank of America setting a $30 target on Figma and, in the same note, an Underperform on Adobe.

One thesis held both calls together: the collaboration layer survives the model, and the legacy creative suite is the one exposed.

None of it is a referendum on whether you can design; all of it is a referendum on whether Figma can keep charging for the room you design in.

Then there is the line practitioners keep replaying. On the company’s first-quarter call, captured in the Figma (FIG) Q1 2026 Earnings Call Transcript, chief executive Dylan Field refused to wave the threat away: “I think when it comes to Anthropic, obviously, we can’t dismiss them.”

He then argued that Figma’s real-time multiplayer canvas and deep product context are the hard parts to copy. On that same call he made the point the market keeps missing: when building gets cheap, design and creativity become the edge, not the casualty.

The defendant testified for the discipline. The jury did not take notes.

None of it is a referendum on whether you can design. All of it is a referendum on whether Figma can keep charging for the room you design in. Sit with that distinction, because the rest of the market keeps losing it.

The Tool Has Stood Trial Before

Here is the part the ticker cannot show you: we have run this exact trial, more than once, and the verdict on the craft has come back the same every time.

Flash was once the design layer. If you built anything expressive on the web between 1999 and 2010, you built it in Flash — motion, interaction, sound, the whole vocabulary of rich media. Then standards caught up, mobile refused to run it, and Adobe wound it down. The people who had defined themselves as Flash designers had a bad year. Interaction design had a great decade. The work moved; it did not die.

The moat was always the tool, but the craft was always the judgment. Confuse the two and you misread every trial.

Go back further and it rhymes. QuarkXPress owned professional page layout so completely that Quark became a verb, until Adobe’s InDesign took the market in the early 2000s.

Typographers did not retrain as baristas because they opened a new application and kept setting type, because setting type was the skill, and Quark was only where they happened to do it.

Then it happened to the winners. Sketch took the interface-design crown from the heavyweights, and for a few years Sketch was how you said UI design. Figma — browser-based, multiplayer, free to start — took it right back. The designers who could think in systems packed up and moved to the new canvas without losing a step.

Josh Clark has a line for why this keeps happening. AI, he argues in AI Is Your New Design Material, is exactly that: a material, like wood, with a grain you learn to read, where the job is to point your attention and judgment at it.

Materials change. The pointing does not.

The pattern did not stop at Figma; it arrived at Figma which is the winner of the last round is now the incumbent in the dock, and the tools unsettling its stock — a chatbot that drafts a screen, a free canvas from a search company — are only this round’s challengers.

The moat was always the tool and rhe craft was always the judgment; if you confuse the two, you misread every trial.

The Jury Keeps Convicting The Wrong Defendant

So why does the market keep returning a guilty verdict on the craft when only the company is in the dock?

Because a share price is the only instrument the market has, and it measures exactly one thing: can this company convert the work into cash, and keep doing it?

That is a fair question about Figma. It is a category error about design.

Look at the fundamentals underneath the fear. As Figma Stock Is Down 85%, But Wall Street Thinks It’s Time for the Stock to Surge documents, first-quarter revenue grew 46% year over year, net dollar retention hit 139%, and the company raised its guidance — yet the stock still sits far below the $143 peak it reached when it went public in 2025. Those are not the numbers of a dying business. They are the numbers of a business the market is too frightened to price.

A share price can measure whether a company can charge for a canvas; it cannot measure whether the world still needs people who design.

And the fear has real fuel, none of which is about your talent. Anthropic and Figma had been close partners, wiring Anthropic’s models into Figma’s product — a relationship VentureBeat traced in Anthropic just launched Claude Design, an AI tool that turns prompts into prototypes and challenges Figma, the same report noting that chief product officer Mike Krieger left Figma’s board days before Anthropic shipped Claude Design, a tool that turns a prompt into a prototype. Google, meanwhile, made its own AI-native design tool, Stitch, free. A former partner on the board, then a rival across the table; free competitors; a 42% short position. That is a business knife-fight, and the jury is watching all of it.

A share price can measure whether a company can charge for a canvas. It cannot measure whether the world still needs people who design. The two questions live on different floors of the building. The market only has an elevator to one of them.

The Skills The Verdict Cannot Touch

If the tool is always the defendant, then the smart move is to stop identifying with the defendant. The tool you learned will change three more times before you retire, but the judgment you build will not.

The tool you learned will change three more times before you retire. The judgment you build will not.

Here is what holds when the software you love is the one in the dock.

• Own the problem, not the pixels. The pixels are the part the tool makes cheap; the problem is the part you were hired to understand. Stay close to what users are trying to do, and no model ships that understanding out from under you.

• Trade production for judgment. When producing the artifact costs almost nothing, the value moves to knowing which artifact is worth producing — and to saying no to the other nine. That call has never lived inside a tool.

• Learn the AI-native tools without mourning the old ones. Open Claude Design, open Stitch, open whatever launches next month, and get fluent quickly. The designers who lost the last three transitions were the ones defending a file format instead of the work.

• Read the ticker as news, not as a mirror. Figma’s share price is a real story about one company’s pricing power. It is not a performance review of your career. Keep the two documents in separate folders.

Do all four and the next trial becomes what it always was: someone else’s, about someone else’s business, held in a courtroom where your name never appears on the docket.

The Verdict That Was Never Yours

The jury will keep getting confused. That is what juries in this particular courtroom do. A tool gets expensive, a cheaper way to do the work shows up, and the market puts the whole enterprise on trial — pricing, moat, seats, the works. Somewhere in the gallery, the people who do the work mistake the proceedings for their own sentencing.

Do not make that mistake. Flash died, and interaction design got bigger. QuarkXPress fell, and typography did not need a funeral. Sketch handed the canvas to Figma, and the people who could think in systems moved with it, because the thinking was the thing they carried, not the file format. The verdict on the design layer is still being argued, and it may go badly for any single tool. It says nothing about you.

Watch the ticker if you want. It is a real business story, and a good one. Just do not read it as a mirror. The market is pricing one company’s ability to charge for a canvas.

It is not qualified to rule on whether the world still needs people who can look at a hard problem and see the shape of the answer. That case never opened.

Every team building with artificial intelligence right now believes it is working on a blank frontier. New interface, new rules, no map. That belief is wrong, and it is expensive.

The cost is not hypothetical, and it shows up as rework, as users who never trust the tool, and products that age quickly. It feels a lot like 1997, when the web was the wild west almost literally. The images in this post reflect that.

A handful of browsers each went their own way. Netscape Navigator, Internet Explorer, Opera, and the fading remains of Mosaic rendered the same page differently, each with its own proprietary tags, thinking they had invented sliced bread:

• Netscape, specifically Lou Montulli, gave you a blink tag, but he claims it was implemented by someone else.

• Microsoft answered with a scrolling marquee. No one wanted to take credit.

Both quite honestly sucked, but it was early times so I won’t blame Lou or Bill Gates (That might be the wrong one but who’s looking?).

A site built for one broke in the next, so designers rebuilt the same thing three times and called it a living.

Then a small group decided the chaos was a choice and not a fact of nature, and they made the case for standards. L. Jeffrey Zeldman was the lead pirate.

In the theme of Fight Club, we have to keep saying his name to honor his contribution to the field. His name is Jeffrey Zeldman.

Zeldman did not invent the specifications, he did something harder: He convinced an entire industry that shared conventions were worth fighting for, and he won. Zeldman changed the world with a stance, not a specification and we should thank him for it.

We are living through that moment again, this time for the interfaces we wrap around models, the skills we scaffold on top of them and representations they mean.

The browsers have new names: ChatGPT, Claude, Gemini, and Copilot each handle the same task their own way, with their own conventions for parsing content, showing reasoning, citing a source, and asking permission before they act.

There are no shared standards to make them consistent, so the patterns are unsettled, the vendors are playing jazz, and the window to shape the conventions is open.

We’re in a revolution, as Dan Maccarone so correctly points out. It will not stay open long, so it is worth studying what Zeldman did, and asking what we can borrow. We really have to get this right.

What Happened Before: Zeldman and the Fight for Web Standards

A Magazine and a Coalition

Start with who he is, because a surprising number of designers working today have never had to learn the name. The gap is wider than one name.

Most designers now arrive in the field with no working knowledge of how its standards were won— who fought for them, against what, and why it mattered. I ask designers and get a lot of blank stares.

That gap runs right up through the people teaching design, many of whom cannot pass on a history they were never handed themselves. A profession that forgets how it climbed out of its last mess is well positioned to climb back in.

So here is the history worth recovering.

Jeffrey Zeldman is, in the phrase Business Week gave him, the King of Web Standards. He has been publishing at his own site since the mid-1990s. In 1998 he founded A List Apart, the magazine for people who make websites, which became the place where the craft argued with itself and got better.

That same year he co-founded the Web Standards Project, a grassroots coalition of designers and developers with an unglamorous goal: get the browser makers to support the same technologies the same way.

Note who he organized: the designers and developers who built the web, one page at a time. Other than my disagreements with him using really small type.

His other efforts:

• A List Apart gave them a place to argue and learn, the Web Standards Project gave them a banner

• An Event Apart, which he co-founded in 2005, gave them a room to stand in together

• A Book Apart grew knowledge of the field by selling book.

This was grassroots in the literal sense. The pressure came up from the people closest to the markup — the ones paying the daily tax of the chaos — rather than down from above.

The Browser Wars

That goal sounds small until you remember what it was up against. Browser makers treated the web as territory to capture, not a commons to tend.

Microsoft and Netscape shipped incompatible features by design, because lock-in was the strategy. Zeldman and his collaborators spent years making the case, in public and in private, that this served no one over the long run — not users, not designers, not even the browser makers.

Browser makers treated the web as territory to capture, not a commons to tend, and it showed up in the quality of their products.

The Argument That Outlasted the Tools

His 2003 book, Designing with Web Standards, was the argument in its most durable form and the first half was not aimed at developers at all.

It was aimed at the people who sign the checks: here is why separating structure from presentation from behavior makes your site cheaper to build, faster to load, easier to maintain, and reachable by more people.

The book is credited with moving the industry off tag soup and Flash and toward semantic markup and accessibility, and it ended up as a textbook in dozens of college programs.

Here is the part I keep coming back to. Zeldman’s real achievement was never a specification, it was a shared belief that the web should work the same way for everyone, held widely enough that building any other way started to feel embarrassing.

Standards won because enough people decided they should.

Why It Matters: What the Standards Movement Won

So what did the standards movement win, and how did it win it? Three things worth stealing.

A semantic separation as a standard

The core move was to pull structure, presentation, and behavior into distinct layers so each could change without breaking the others. That one idea outlived every specific technology it was built on.

It is why a design system works at all.

When Brad Frost introduced atomic design, he was extending the same instinct: stop shipping pages, start composing interfaces from small, shared, recombinable parts. Design systems are the standards movement’s direct descendant, and they are the closest thing we have to a working model for AI interface conventions.

Hold onto that frame, because it names the work: this is a design systems problem. Documenting a pattern once and reusing it everywhere did not begin with atomic design, or even with the phrase “design system.”

erin malone built one of the first public pattern libraries at Yahoo in the mid-2000s and traced the whole lineage, from Chris Alexander forward, in A History of Patterns in User Experience Design. The industry already knows how to stop reinventing the same button.

AI is that problem at a larger scale, and the answer is the same shape: a shared, documented, governed system instead of a fresh pile of one-offs per product.

A business case before the moral one

Zeldman did not lead with accessibility as a virtue, though he believed in it.

He led with cost and reach, which was exactly the right leave: standards-based sites cost less to build, load faster, and reach more people. Accessibility came bundled with the economics, which is how it got adopted. Do it right, and you don’t get sued.

If you want a convention to spread, make it cheap, not just the righteous. VHS won over Betamax for exactly that reason.

That is the lever, and it is the one most AI teams are ignoring while they chase novelty. Simplicity wins, full stop.

A coalition, not a command

No one had the authority to mandate standards. The Web Standards Project had no power except persuasion, a body of shared work, and the patience to keep making the case until the browser makers found it easier to comply than resist.

The standards did not come down from a committee; they came up from practitioners who agreed with each other, in public, until agreement became the default.

Put those wins next to the vocabulary the standards era produced, and you get a checklist. Every concept below was hard-won for the web, and each one now has an AI-shaped gap where the next convention belongs.

• Separation of concerns. Structure, presentation, and behavior became distinct layers. Keep a model’s reasoning, its output, and its actions inspectable and separate, rather than fused into one opaque block of text.

• Accessibility. The web committed to working for the people the median design forgot. Extend that to AI — legible to assistive technology, honest about what the model cannot do, and usable when someone opts out of the automation.

• Semantic markup. Tags gave content meaning a machine could read. AI output needs the same: a claim, a citation, a guess, and an action should each be labeled and legible, not flattened into identical prose.

• Progressive enhancement. Pages worked without the fancy parts, then improved where support existed. AI features should degrade to a dependable path when the model is wrong, slow, or unsure, instead of collapsing.

• Responsive design. Layouts adapted to context instead of assuming one screen, the case ethan marcotte made in Responsive Web Design. Generative interfaces can adapt to task and person the same way, as long as the adaptation stays predictable. Most of this foundation was built off Zelman’s work.

• Interoperability. The whole point was that a page worked the same across browsers. The equivalent now is shared interaction patterns across assistants, so moving between tools does not mean relearning the rules every time.

Why We Need It Now: The AI Interface Moment Is 1999 Again

Now look at where we are. Artificial intelligence has adoption numbers the standards movement could only dream of.

ChatGPT alone reached eight hundred million weekly users in 2025, and inside companies the pattern holds: in OpenAI’s own state of enterprise AI report, three in four workers said the tools improved the speed or quality of their output. The demand is not in question. The interface is.

Jakob Nielsen has called this the first new user interface paradigm in sixty years — a shift from telling the computer how to do a thing to telling it what you want and letting it decide how.

Nielsen is right, and that is exactly the problem. A new paradigm means some conventions are settled but not all.

How should a model show its reasoning? How should it cite a source, signal how confident it is, or hand a task back to a person when it is out of its depth? Every product answers differently, and most answer badly.

That inconsistency carries a cost people are only starting to name. When every model formats its reasoning, its sources, and its confidence differently, you cannot line two answers up side by side — which is exactly what evaluation needs. Comparing answers across systems has become hard, and that is quietly degrading the evaluations teams rely on to know whether any of this is working. You end up grading the packaging instead of the answer.

This is 1999 wearing new clothes.

Back then a site worked in one browser and broke in the next. Today a workflow works in one assistant and falls apart in another, because every vendor has invented its own gestures, its own affordances, its own quiet rules about what the machine will and will not do.

We are pouring the same proprietary concrete, only faster and with more money behind it.

And the interface is only the visible half. Beneath it sits the scaffolding of skills — the reusable units of capability a model draws on, and the conventions for describing, discovering, and composing them.

Every platform defines a skill its own way and wires up its tools on its own terms. One piece is further along than the rest: the Model Context Protocol became a cross-vendor standard for connecting models to tools in barely a year.

That is the tell.

The shape of the capability layer is being set right now, and most of it is still up for grabs.

The web had a single surface to tame: the page in a browser. AI is spreading across many at once — chat, voice, autonomous agents, assistants embedded in every application, and interfaces that assemble themselves on the fly.

Every new surface is another place for conventions to fork, and another set of rules for people to relearn. This is the part the excitement obscures: new surfaces need standards more than ever, because the cost of getting them wrong now compounds across all of them.

Mobile is where this gets concrete. It is the surface most people live on, and the one where agents grow real teeth — booking, buying, and messaging across apps on your behalf, sometimes on standing permission. The AI inside mobile apps and the agents acting across them both need shared rules for consent, confidence, and handoff, or the most personal device becomes the most inconsistent one.

The lock-in logic is identical too. When your patterns are yours alone, switching costs go up and users get stranded. What looked like a moat in 1999 became a liability the moment standards made the rest of the web interoperable. That reckoning is coming for AI interfaces. The only question is whether designers shape it or inherit it.

What It Should Look Like: Building Shared Standards for AI Interfaces

None of this happens by waiting. Here is where the work is.

Name the patterns before the vendors lock them in.

The web got semantic markup because designers agreed on what a heading, a list, and a link meant before the tools decided for them.

Do the same for AI interfaces now: agree on what “show your work,” “cite your source,” and “I am not sure” should look and behave like, as reusable components rather than one-off features.

The same discipline belongs a layer down, in how a skill is described, discovered, and combined, so capabilities compose instead of colliding. If you keep a component library, this is where shared AI patterns belong.

My own ux-components.com is one attempt at that; build your own and build it in the open.

Make traceability and confidence first-class parts of the interface.

The most dangerous thing an AI interface can do is present a guess with the same visual authority as a fact. Standards-era design stripped false neutrality out of markup by making structure explicit.

The equivalent move today is to make sourcing, traceability, and calibrated confidence visible by default, so a person can see where an answer came from and how far to trust it. Treat that as a convention, not a competitive differentiator.

The most dangerous thing an AI interface can do is present a guess with the same visual authority as a fact.

Build the coalition, not the mandate.

You do not have the authority to set an industry standard, and neither did the Web Standards Project. What you have is your own work, done in public, and the willingness to keep making the case.

• Publish your patterns.

• Argue for them at conferences and in writing.

• Adopt other people’s good conventions instead of reinventing them for credit.

Start this week: take one convention you have already settled — how your product signals confidence, say — write it down, and publish it where another team can find and copy it.

Standards spread when agreeing becomes easier than diverging, especially when cost is involved.

The New Markup Is Markdown

There is a quieter version of this already taking shape in the projects agents work in — Markdown as a baseline, which also has a long history back to 2004 as invented by John Gruber and the late Aaron Swartz. It is cheap and good, thus has become a standard.

Agents increasingly take their instructions from plain text files that sit beside the work — AGENTS.md for how an agent should behave in a project, SKILL.md for what a capability can do, README.md for the context around both.

This is the new semantic layer. It is markup again, written in Markdown and read by a model instead of a browser.

That layer inherits two disciplines the web spent twenty years learning, and it is being built as if neither exists. The first is accessibility, and right now it is missing from what informs the agent.

A model generating an interface has no accessibility standard in front of it, so it ships whatever it ships. Those requirements — contrast, focus order, alternative text, keyboard paths — belong in the file that guides the agent, feeding it up front rather than getting bolted on after the interface is generated. The second is content strategy. Voice, terminology, reading level, and the shape of what a model produces are decisions, and right now they are made by default, one prompt at a time, with no shared model to anchor them.

So write the files down, and standardize what goes in them.

A design.md that carries your design system’s patterns, tokens, and rules into every agent that touches the product. An accessibility.md that states the non-negotiables in language a model can follow. A content.md that fixes voice, terminology, and the content model.

The set is not fixed — a research.md for what you know about users, a brand.md for identity, whatever your practice depends on.

Name them, share them, and treat them the way the web learned to treat structure, presentation, and accessibility: as a contract, not a preference.

The W3C Already Built the Machinery and We should Use it

Here is the part that should make this feel less daunting. You are not starting from nothing.

The World Wide Web Consortium — the W3C — has run an open, consensus-based standards process since 1994. It produced the specifications that Zeldman’s coalition spent years pushing browser makers to honor.

The division of labor back then is worth remembering. The W3C wrote the standards, and the Web Standards Project made adoption non-optional. One body defined the shared language; a movement made it stick.

That machinery did not disappear. It still runs the accessibility, privacy, and security work that any responsible AI interface will need, and it has already turned toward this moment. In 2025 the W3C launched a Web and AI Interest Group to work through how AI technologies intersect with the web, and community groups there are drafting early protocols for how agents identify themselves and cooperate.

You are not waiting for this to begin. It has begun. A partial map of the standards taking shape right now:

• Model Context Protocol — a shared way for a model to reach tools, data, and context, already adopted across rival platforms and now stewarded by a neutral foundation.

• A2UI — a declarative protocol for agents to describe interfaces that render natively across web, mobile, and desktop, keeping what the interface is separate from how each client draws it.

• Agent2Agent — an open protocol for agents to discover one another and collaborate across frameworks and vendors, launched by Google and handed to the Linux Foundation.

• The W3C AI Agent Protocol Community Group — a grassroots group drafting open rules for a trustworthy web of agents.

• Agent identity work — cross-body efforts, at the W3C and beyond, to verify who an agent is and what it is allowed to do before it acts.

None of these is finished, and that is the opening. The conventions are still soft enough to shape, which is exactly where Zeldman’s coalition made its difference.

So leverage the past instead of relighting the fire. Extend the accessibility standards that already exist rather than inventing parallel ones, and bring your patterns to the groups already forming instead of publishing them into a silo. The lesson of the browser wars is not only that standards win. It is that the institutions to make them are already here, and the work is joining them, not founding a rival.

Conclusion

The standards moment is not a prediction. It is a choice, and it is being made right now, mostly by default.

Every AI product that ships a novel interface without asking whether it should be novel is casting a vote for chaos — for a world where each assistant works its own way, users relearn the rules with every tool, and the people who get stranded are the ones with the least patience for our cleverness. That world is not inevitable. It only feels that way because no one with a title is going to stop it.

Zeldman’s lesson is that no one ever does. Standards did not arrive because an authority demanded them. They arrived because a stubborn coalition of practitioners decided the web should work the same way for everyone, and then made that true one argument at a time.

We are the practitioners now. The models are astonishing and the interfaces around them are a mess, which is exactly the condition Zeldman walked into and refused to accept.

So look back, take what worked, and get to work. The window is open, the same one Zeldman climbed through. It closes the moment we agree the chaos is normal.

For a decade, design proved itself in the artifact: A pixel-perfect Figma file, a prototype that clicked exactly right, a design system with everything perfectly documented.

That proved you were good.

A Figma frame has never shipped revenue — It promises a product but it is not one, and we spent years mistaking the two.

Design was never supposed to be about production, it was about products people use — products where engagement, retention, and revenue are the outcomes that prove the work mattered.

The craft narrowed to the file, and our impact narrowed with it.

The shift isn’t new. When desktop publishing arrived in the 1980s, typesetters who guarded the metal lost; those who followed craft to judgment thrived. Artificial intelligence (AI) is running the same play on design, only faster.

The production layer — the wireframe, the boilerplate, the competent first draft of a screen — is collapsing toward free, which changes our own perceived value proposition.

When making gets cheap, much of what you called craft turns out to be production wearing craft’s clothes. What survives is the part that was never about the file: choosing the right problem and owning the outcome it moves. Not a loss but a relocation you can get ahead of. Here is where craft goes.

Craft is not decoration on the product. It is part of what makes the product worth trusting.

Craft Now Means Choosing The Right Problem, Not Producing The Artifact

For most of your career, craft was legible in the output. A tight layout, a clean component, a flow with no dead ends: these were how you proved you were good, and they were mostly production.

AI is good at production.

Point a competent model at a clear prompt and it returns a plausible screen, a serviceable component, a first draft that would have cost you an afternoon.

The polish that used to signal skill is now a commodity.

So the signal moves upstream. John Maeda has been making this argument for a while; in Practicing CRAFT: Keeping Ahead of the Machine, he reframes craft as the human discernment a machine optimizing for the shortest path will skip. The efficient answer and the right answer are not the same answer, and knowing the difference is the work.

When a tool can generate a thousand plausible screens before lunch, the scarce skill is no longer making one; it is knowing which one deserves to exist.

I have argued the same from the production side: when execution gets cheap, the work that survives is deciding what to build, understanding who it is for, and owning the call when the evidence is mixed. That is craft now. It is less visible than a beautiful artifact and far harder to fake.

The practical shift is from making to curation.

When a tool can generate a thousand plausible screens before lunch, the scarce skill is no longer making one; it is knowing which one deserves to exist. That skill has a name, taste, and most people have been outsourcing it to whoever runs the critique.

None of this means polish stops mattering. It means polish is table stakes, not the differentiator. The differentiator is the judgment that points all that cheap production at a problem worth solving in the first place.

That judgment is also what earns trust. Users never see your process, but they feel its absence. A product built on the right calls feels coherent and reliable, and reliability is what brings people back; one built on plausible guesses feels off in ways people cannot name and do not forgive.

Action items

• Defend one direction in writing. The next time AI hands you three or four plausible directions, pick one and write two sentences on why it wins for this user, this moment, and this constraint. Do it even when no one asks.

• Separate production from judgment. Audit one recent deliverable and split the parts a machine could regenerate from the parts where you owned the decision. Defend your calendar for the second kind.

• State the problem before you open the tool. Write it down in one sentence first. If you cannot state the problem cleanly, you are not ready to evaluate any solution the machine hands you.

Resources

• John Maeda, “Practicing CRAFT: Keeping Ahead of the Machine”

• Patrick Neeman, “When Execution Gets Cheap in an AI World, Most UX Ceremony Loses Its Point”

• Patrick Neeman, “AI Wants Builders. Here’s How to Become One.”

Make Research And Iteration A Habit, Not A Phase

The focus is now people and not the machines, by design.

The tools are so fluent that they will happily produce the thing research was supposed to earn.

• Ask a model for a user persona and it writes one.

• Ask for a research readout and it returns ten pages, every paragraph plausible, every quote invented. The document is finished before any research has happened, and it looks exactly like the real thing.

• Ask it to test it with a synthetic usability test, and it will do it in sections.

That is why the answer is not more research documents, it is a research habit that feeds the machine. Teresa Torres calls this continuous discovery: at a minimum, weekly touchpoints with customers by the team building the product, doing small research activities in pursuit of an outcome.

The point is contact with reality on a cadence, not a phase you complete and file away.

Research is no longer the slow part you skip under deadline; it is the cheap insurance against a machine that is wrong in fluent, well-formatted prose.

AI raises the value of that habit rather than replacing it. When the machine can fabricate a confident answer to any question, the habit of checking that answer against a real person becomes the thing standing between you and shipping fiction.

Iteration changes shape too. Breadth used to be expensive, so you explored two directions and committed early, often to the wrong one. Now breadth is close to free: you can put ten directions on the table in the time it took to mock up one. The skill is no longer generating options. It is killing the weak ones fast, before anyone grows attached, and the only reliable way to kill them is evidence from someone who is not on the team.

Action items

• Book a weekly customer touchpoint. Put one recurring slot on the calendar this week and defend it like a production deadline. Small and weekly beats large and quarterly.

• Label AI findings unverified by default. Before you accept any AI-generated persona, quote, or finding, mark it unverified until a real person confirms it. Ship the label, not the assumption.

• Kill weak options with evidence, fast. When AI gives you ten directions, run a five-minute test with one user to cut the bottom half before your team debates the top half.

Resources

• Teresa Torres, “Everyone Can Do Continuous Discovery — Even You! Here’s How”

• Teresa Torres, “Continuous Discovery: Definition and Overview”

• Patrick Neeman, “Tool-Shaped Objects in UX, Part 1”

The Machine Can Only Apply The Craft You Can Put Into Words

Here is the good news the panic skips over. AI can carry real craft, not just production. It can hold your standards, catch your anti-patterns, and apply your definition of good across a thousand screens without tiring.

It can do this only for the part of your craft you can put into words. Everything you cannot say, it cannot use.

Most craft is tacit. You know a layout is wrong before you can explain why. You feel that a flow has one screen too many. Michael Polanyi named this decades ago: we know more than we can tell.

Your taste lives mostly below the waterline of language, in pattern recognition you built over years and never had to state, because your own hands did the work. That gap is harmless when you do the work yourself. It becomes the whole problem the moment you hand the work to a machine.

Point a model at a vague brief and it fills the silence with its own defaults, which is the average of everything it has seen. Average is exactly what craft is supposed to beat.

The machine cannot read your hands; it can only read your words.

So the new literacy is description. The work is turning “I will know it when I see it” into stated constraints, named anti-patterns, and an explicit account of what good means for this product and no other. Geoffrey Litt frames the shift in Malleable Software in the Age of LLMs: the old bottleneck was turning fuzzy, informal intent into something formal, and the machine is opening that bottleneck fast. It opens from your side.

The intent still comes from you, and the sharper you make it, the better the machine bridges the gap.

This is uncomfortable, because taste resists being written down. Say the standard out loud and it sounds either obvious or arbitrary, and you lose the cover of the knowing nod in critique.

Maggie Appleton’s Language Model Sketchbook points the same way from the interface side: the tools that work are scoped, structured, and specific, not open-ended chat that hopes you will supply the constraints. They are well defined controlled vocabularies that define exactly what you mean with nothing left to chance.

Specific beats open-ended, for the same reason a good brief beats a vague one.

Writing your craft down is itself craft. It forces you to find the rule under the reflex, and finding it sharpens the reflex. Do it well and you get a second thing for free: a standard the machine can hold, so its help finally looks like your work instead of the internet’s.

Action items

• Write your definition of good. For one current project, state it in plain sentences: three things it must do, three it must never do. Hand that to the model before anything else, and watch the output move toward your standard instead of the average.

• Turn every rejection into a rule. The next time you reject an AI draft, do not just regenerate. Write the one sentence that says why it failed, add it to your standing instructions, and make the correction permanent instead of repeating it.

• Make one tacit standard explicit. Take a piece of your taste you have never stated, such as spacing, tone, or when to break a pattern, and force it into a rule you could hand to a stranger. If you cannot write it, you cannot delegate it.

Resources

• Geoffrey Litt, “Malleable Software in the Age of LLMs”

• Maggie Appleton, “Language Model Sketchbook, or Why I Hate Chatbots”

• “Tacit Knowledge in Large Language Models,” The Review of Austrian Economics (2025)

Build The Scaffolding First Starting With A Design.md File

Give an AI agent no context and it will guess. It will invent a hex code, reach for default framework blue, and hand you a settings page that looks like a cousin of your dashboard rather than a sibling. The fix is not a better prompt every time. It is scaffolding: standing context the agent reads before it does anything, so it starts inside your constraints instead of wandering toward them.

This practice now has a name and a shape. In April 2026, Google Labs open-sourced DESIGN.md, a portable format that describes a design system to any coding agent in a single file.

It pairs machine-readable tokens, meaning colors, typography, and spacing, with human-readable rationale explaining why those values exist, and it ships with a validator that checks choices against accessibility contrast rules. The format specification reads like a README written for the machine instead of the new hire.

Scaffolding is how you stop re-explaining your world in every prompt.

The deeper idea travels beyond design.

Coding agents already read standing instruction files, and the teams getting reliable output are the ones treating those files as a source of truth: conventions, boundaries, the vocabulary your product uses, the things an agent could never infer from the code alone.

One caution worth stating plainly. Scaffolding only helps when a human writes it. A context file auto-generated from your codebase and left unattended tends to bloat, drift, and mislead, because it encodes what the machine already guessed rather than what you decided. The value is in the curation, which means the design file is a living document you own, not a one-time export you forget.

Write the scaffolding before you write the feature. It is the highest-leverage hour you will spend all week.

Action items

• Create the design file this week. Put tokens for color, type, and spacing on top, with a short paragraph under each explaining the intent an agent cannot infer.

• Set explicit boundaries. Add a section that names what the agent must never touch and the vocabulary your product uses, then version it alongside your code.

• Keep the file living. When you correct the same mistake twice, that correction belongs in the scaffolding, not in your next prompt. It should be self learning, if possible.

Resources

• Google Labs, “Stitch’s DESIGN.md format is now open-source”

• DESIGN.md format specification (GitHub)

• DESIGN.md specification document

Keep A Human In The Loop And Put The Work In Front Of Users

The most important number in AI right now is a gap between what people believe and what is true. In a 2025 randomized controlled trial, METR had 16 experienced developers complete 246 real tasks, randomly allowing or forbidding AI tools.

The developers predicted AI would make them 24% faster. Afterward, they estimated it had made them 20% faster. Measured against the clock, AI had made them 19% slower. They were confidently wrong about their own experience, in real time.

That is the case for keeping a human in the loop, and then keeping a user in the loop after the human. Your own read on whether the AI got it right is not trustworthy, because the output is fluent enough to feel correct.

Developers are the first to sense this now and we should adopt this stance: in Stack Overflow’s 2025 survey, 84% of developers use or plan to use AI tools, yet only 29% trust the accuracy of the output and 46% actively distrust it. High use, low trust.

The verification of craft did not go away; it moved to you as the owner.

The failure mode is specific and worth naming. I once watched an AI draft a research readout before the research existed: ten pages in twenty seconds, every quote invented, the document finished before a single interview. Nothing in the artifact told you it was fiction.

Only a human who knew the research had not happened could catch it, and only a real user could tell you whether the concept underneath was any good.

So build the loop in on purpose. Put a person between the model and the decision, and put the concept in front of the people it is meant to serve before it hardens into a shipped feature. The machine is a fast, confident first draft. It is not a witness, and it is not your user.

Action items

• Name a human owner. Assign one for every AI-assisted deliverable before it moves forward. The rule is simple: a person signs, not the model.

• Require a real user before deciding. For any AI-generated finding, insight, or quote, demand one real user contact before it informs a decision. Treat unverified output as a hypothesis, not a result.

• Test the concept before the build. Schedule it first. Put the idea in front of five users while it is still cheap to change.

Resources

• METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”

• Stack Overflow, “2025 Developer Survey: AI”

• Patrick Neeman, “Tool-Shaped Objects, Part 2: A UX Designer’s Field Guide to What Survives”

Evaluating The Machine’s Answers Is Now Craft And Now Part Of The System

For your whole career, craft was something you applied to your own work. You made a thing, then you judged it. That order is inverting. More of the work now arrives already made, generated in seconds, and your first real job is to decide whether it is any good.

Evaluation used to be an afterthought at the end of a process; it is becoming the center of the process, and a designed part of the product itself.

This is the taste from earlier turned outward. Choosing the right problem asks for judgment about what to build; evaluating a model’s output asks for the same judgment about what it just built, at volume, all day. It is the skill AI raises the value of most, because the machine will hand you a hundred confident answers and only some of them deserve to survive.

The industry has already turned evaluation into infrastructure. In a 2023 study, Zheng and colleagues showed that a strong model acting as a judge agreed with human preferences more than 80% of the time, the same rate at which humans agree with each other.

Live systems now run this continuously: Chatbot Arena has collected more than 240,000 human preference votes to rank models in the open.

Judging outputs is no longer a side task. It is a system you can build.

A judge is only as good as the standard you hand it, and the standard is still yours to set.

But look closely at that 80%. It is also the ceiling on how often two people agree about what good means, which tells you the hard part was never the mechanics of judging. It was defining the thing being judged.

A model judge inherits whatever rubric you give it, along with its own leanings toward longer answers and toward its own kind of writing.

This is why evaluation stays craft even when it runs on autopilot. @Aman Khan puts it plainly in his guide to building evals: the difference between AI products that work and ones that fail is not the model or the prompt, it is the willingness to argue about what good means, and to keep arguing as new failure modes surface. An eval is a living definition of done. Skip it and the product drifts toward plausible and away from correct, and users feel the difference before your metrics do.

So build the judgment in. Write the rubric, keep a small set of examples you have labeled by hand, and treat evaluation as a first-class part of the system, not a gate you bolt on later. The machine can apply your standard a thousand times an hour. It still cannot decide what the standard is.

Action items

• Build one rubric this week. Take a task you already delegate to AI and write the three to five criteria that separate a good answer from a merely plausible one. Hand it to the model as a judge and to yourself as a reviewer.

• Keep a hand-labeled golden set. Collect a small set of real inputs and label the outputs good, average, or bad yourself. Do not outsource the labeling; the judgment is the point, and it is yours to own.

• Make evaluation the definition of done. Before a feature ships, require it to pass the rubric, not just the tests. When a new failure mode appears in production, fold it into the eval instead of fixing it once and forgetting.

Resources

• Lianmin Zheng et al., “Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena”

• Aman Khan, “Beginner’s Guide to AI Evals (Walkthrough)”

• Wei-Lin Chiang et al., “Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference”

Review At The End Because AI Erodes Quality Quietly

AI does not fail loudly. It fails by accretion, one plausible shortcut at a time, and the damage shows up in the codebase months later. GitClear analyzed 211 million changed lines from 2020 to 2024 and found the signature of that erosion.

Refactored code, the lines that get moved and consolidated when someone reshapes a system, fell from 25% of changes in 2021 to under 10% in 2024. Copy-pasted lines rose from 8.3% to 12.3%, and 2024 was the first year on record that copy-pasted code outnumbered moved code. Duplicated blocks rose roughly eightfold.

Read that as a pattern, not a verdict. The study is correlational, and other things changed over those years. But the mechanism is familiar to anyone who has watched a fast tool meet a tired team: the machine makes adding easy and reusing hard, so people add. The result passes every test and still weakens the foundation, because duplication and drift do not announce themselves in a green build.

A test suite tells you the code runs. A review tells you the code belongs.

This is why the review at the end is not optional ceremony. It is the one checkpoint positioned to catch what the machine hides: the third copy of a function that should have been shared, the pattern that broke from your system, the shortcut that works today and costs you four times over by year two. A test suite tells you the code runs. A review tells you the code belongs.

The review also has to change to match the tool.

Reviewers should know which parts were AI-assisted and read those parts harder, not softer. The instinct to trust fluent output is exactly the instinct a good review exists to correct. You are not reviewing only for whether it works. You are reviewing for whether it should stay.

Action items

• Add an end-of-work review gate. Check for duplication and drift, not just passing tests. A green build is necessary, not sufficient.

• Flag the AI-assisted sections. Label them so reviewers apply extra scrutiny where fluent output is most likely to hide a shortcut.

• Hunt the third copy quarterly. Once a quarter, look for anything duplicated. If a function or pattern exists in three places, consolidate it before it becomes five.

Resources

• GitClear, “AI Copilot Code Quality: 2025 Data Suggests 4x Growth in Code Clones”

• GitClear, “Coding on Copilot: 2023 Data Suggests Downward Pressure on Code Quality”

• Stack Overflow, “Mind the Gap: Closing the AI Trust Gap for Developers”

Conclusion

Return once more to the typesetters. The ones who struggled were not the least talented; they defined themselves by the part of the job the machine took.

The ones who thrived had decided, before the beige computer arrived, that their craft was never the metal. It was the judgment the metal carried.

AI is handing you the same decision on a shorter clock. The production layer is leaving and not coming back. You can defend the artifacts, the polished screens, hand-built components, or follow craft to its new address.

That address has seven rooms. Choose the right problem first. Keep research a weekly habit, not a phase. Describe your craft, so the machine applies your standards, not its defaults. Build the scaffolding to hold that description. Keep a human in the loop, and a user after them. Evaluate what the machine hands you, and build that judgment into the system. Put a real review at the end, where erosion hides.

None of this is nostalgia. It is where the leverage moved. The tools got faster, judgment got scarcer, and scarce is where craft has always lived. Name it clearly and the machine carries it, amplifying your judgment, not replacing it.

Enterprises have poured billions into generative AI, and most of it has produced nothing you can measure. A 2025 study from MIT’s NANDA initiative found that roughly ninety-five percent of enterprise generative AI pilots delivered no measurable impact.

The models work. The programs do not.

The failure is old: most enterprises never learned to build software iteratively. They fund one big bet, specify it up front, and ship a year later, hoping the guess held.

Eric Ries published The Lean Startup in 2011, after the dot-com bust buried a generation of companies that built things nobody wanted. Speed of learning, not speed of shipping, separates the survivors from the casualties. He could not have imagined generative AI; experiments then were expensive so it was about making smaller ones.

The tools have since collapsed the cost of execution, which makes the discipline more appropriate today, not less.

The design world has its own version of the big up-front bet, and it is due for the same reckoning. Long-form design thinking — the months-long IDEO-style engagement that empathizes, defines, ideates, prototypes and costs $300 an hour — was built for a slower world.

I lived through that era, and we shipped fast when we could — really fast. Two week sprints, continuous learning and measurement, analysis of data anyway we could. It was Lean Startup before the book was out, in spades, and I learned so much that I apply today.

And today, it’s back.

When a working prototype now takes an afternoon and the ground shifts every month, spending a quarter on empathy maps before you build anything is not rigor. Most of the design process is theater now that you can ship by lunch.

Understanding the user still matters; the front-loaded ritual around it does not when you can learn in small increments.

For years I have modeled my work as small startups: fast experiments, quick shutdowns when they stopped paying off. The ground under AI shifts so fast that learning is the only durable skill. What does not shift is the foundation: this comes back to first principles of design, and the laws of UX did not change with the tooling.

Here are four things the book still teaches us.

Understanding the Problem in the Real World Still Comes First

The comfortable read is that the models are not good enough yet, and the next release will rescue these pilots. It will not. Look at why companies die, and the same cause sits at the top of the list. CB Insights, reviewing hundreds of startup post-mortems, found that poor product-market fit remains the leading root cause of failure — founders built something the market did not need badly enough to pay for. Running out of money is how the story ends, but a problem nobody had is how it starts.

Generative AI pilots are dying the same death. The MIT researchers were clear that the failures were not caused by weak models. They were caused by tools built to signal innovation rather than solve a job someone really has.

A demo that impresses a boardroom and a tool that survives a Tuesday are two different products. One is built to be watched. The other is built to be used.

Toyota had a word for the fix long before software borrowed it. Genchi genbutsu — go and see. Do not manage the problem from a conference room; walk to the place where the work happens and watch it fail.

Teresa Torres makes the same case for modern product teams in her work on continuous discovery: you find real opportunities by staying close to real customers, not by admiring your roadmap. The teams crossing the AI divide are the ones amplifying work people already do. The teams stuck on the wrong side are the ones who never left the room.

Action items

• Go and see the real work. Sit with the people who will use the thing, in the place they use it, before you scope a single feature.

• Name the job before you build. Write down the specific task a real person is trying to get done; if you cannot, you do not have a product yet.

• Kill demo-driven roadmaps. Stop funding what impresses a boardroom and start funding what survives a Tuesday.

Resources

• CB Insights, the top reasons startups fail — poor product-market fit as the leading root cause.

• Teresa Torres, opportunity solution trees — staying close to real customers through continuous discovery.

Learning Fast Beats Shipping Big

The core loop of The Lean Startup is three words: build, measure, learn. Not build, ship, celebrate. The measure-and-learn half is where the value hides, and it is the half teams keep cutting when the schedule tightens.

If The Lean Startup names the loop, the design sprint shows you how to run it. The five-day method the Google Ventures team laid out in Sprint takes a real problem to a tested prototype in a week — map the problem on Monday, build a realistic prototype by Thursday, put it in front of five customers on Friday. You learn whether an idea holds up before you spend a quarter building it. Same instinct as lean, compressed into a calendar you can clear.

None of this is new. Mary and Tom Poppendieck carried Toyota’s lean manufacturing into software two decades ago in Lean Software Development: An Agile Toolkit, built on the same structure — amplify learning, decide as late as the evidence allows, and treat a big up-front specification as the waste it usually is.

Ries generalized the idea to startups; the Google Ventures team compressed it into a week.

Waterfall, the plan-it-all-then-build-it-all model these books were written to bury, is dead. Most enterprises just have not held the funeral.

Generative AI is the best thing to happen to that loop in years. It collapses the cost and the calendar of an experiment — a prototype that used to fill a sprint week takes an afternoon, a research synthesis that took days takes an hour. Lean always wanted cheap, fast experiments; the tools finally deliver them. Used well, AI does not tempt you to skip the learning. It buys you more of it, because you can run the next test today instead of next quarter.

A faster loop is only an advantage if you aim it at an outcome and put a guardrail on what it ships.

Put the guardrails in first. A loop that runs faster also fails faster, so evaluation goes in front of the ship button — automated checks, a human reviewing whatever reaches a customer, a clear definition of good before you generate anything. Done right, guardrails are not brakes on speed. They are what lets you keep your foot down without driving into a wall.

Then aim it at an outcome. Ten features shipped is a vanity number; the only question a fast experiment has to answer is whether a real metric moved — activation, retention, hours saved, revenue — for a real person doing a real job. Tie the loop to that, and generative AI becomes the fastest learning engine you have ever had. Tie it to volume, and you have automated the busywork of shipping things nobody measured.

Action items

• Run the smallest experiment that teaches you something. Break the idea into its riskiest assumption and test that first, this week.

• Put evaluation in front of the ship button. Automated checks plus a human on anything customer-facing, with a definition of good set before you generate.

• Measure the outcome, not the output. Ask whether a real metric moved for a real person, not how many features you shipped.

Resources

• Sprint, Jake Knapp, John Zeratsky, and Braden Kowitz (Google Ventures) — the five-day design sprint for testing an idea before building it.

• Lean Software Development: An Agile Toolkit, Mary and Tom Poppendieck — Toyota’s lean principles adapted to software.

Speed Comes From Approach, Not Ambition

Execution really is faster now. That part of the hype is true. But the speed that matters comes from how you approach the work, not from how much you attempt at once.

The MIT data makes the point in numbers.

Buying from specialized vendors and building partnerships reached deployment far more often than internal builds, which succeeded roughly a third as often. And the fastest movers were not the large enterprises running the most pilots and staffing the most people. They were mid-market companies that scoped the work tightly and shipped.

Ambition, measured in pilot count and headcount, correlated with getting stuck.

The output is a starting point you edit, not an answer you ship.

I have watched the same pattern up close. When I wrote a book with generative AI in the loop, the tool accelerated the parts I disliked and sped up the exploration, but the first draft was never the deliverable, and every chapter still needed a human to make it hold together.

Approach the work that way and you move fast as a vending machine for finished work and you join the ninety-five percent.

Speed is a by product of narrow scope and a real problem. It is not something you buy by attempting more.

Action items

• Buy or partner before you build. Reach for a specialized vendor first; reserve custom builds for the few places they truly set you apart.

• Scope narrow, ship, then expand. Pick a piece small enough to put in front of users this week, not a platform you unveil next year.

• Treat AI output as a draft, not a deliverable. Keep a human deciding what holds together and what reaches a customer.

Resources

• Fortune on the MIT NANDA report — vendor partnerships reach deployment far more often than internal builds.

• Patrick Neeman, writing an e-book with generative AI — AI accelerates creation while the first draft still needs a human in the loop.

Documentation for Its Own Sake Is Waste

Toyota’s production system has a word for effort that consumes resources without creating value: muda, waste. The whole method is built around removing it. Decades later, seventeen engineers met at a ski lodge in Utah and wrote the same instinct into the Agile Manifesto, which values working software over comprehensive documentation.

Not no documentation. Working software more.

Generative AI is a waste machine pointed straight at this failure mode.

It can generate a fifty-page strategy deck, a governance framework, and a model card before lunch, and none of it moves a product one inch closer to a user.

Programs stuck on the wrong side of the divide often have beautiful documentation and nothing in production. The documents became the deliverable, which is how you feel productive while shipping nothing.

The test has not changed since the assembly line. Does this document help someone or some system make a decision or do the work? If yes, write it. If it exists to prove you were busy, to survive a review, or because the model made it easy, it is waste wearing a nice cover.

AI lowered the cost of producing documents to near zero, which means the only thing keeping your shelf clear now is judgment.

Action items

• Apply one test to every document. Does it help someone or some system make a decision or do the work? If not, it is waste.

• Delete documentation that substitutes for shipping. A beautiful deck with nothing in production is the failure, not the achievement.

• Write only what the work needs. Now that AI makes documents free to produce, judgment is the only thing keeping the shelf clear.

Resources

• The Agile Manifesto — working software over comprehensive documentation.

The Job Did Not Change

Strip away the framework names and The Lean Startup was never really about startups. It was about how to act when you cannot know the answer in advance, which is the condition every team building with AI is living in right now. The uncertainty is not a flaw in your process. It is the terrain.

So the moves are the ones Ries described, and the ones Toyota’s line workers practiced decades before him. Go see the real problem before you build for it. Make the smallest bet that teaches you something, then make the next one. Let your approach, not your ambition, set your speed. Write only the documents that help someone decide or do the work, and skip the rest.

None of this requires a better model. The models are already better than the programs pointed at them. What the ninety-five percent are missing is not capability. It is the discipline to learn faster than they spend.

That discipline is fifteen years old and older. It was true when the money was chasing pets and groceries, and it is true now that it is chasing agents. The tools have changed. The job has not changed. Learn first. Build second.

They could not have pictured the telegraph, the automobile, radio, the internet, the Hamster Dance, or Donald Trump. They certainly could not have pictured a machine that creates French Pug moving paintings that violate Copyright Law, writes a legal brief, paints a portrait, or scores a defendant’s risk of reoffending in the time it takes to read this sentence.

The country marks its 250th birthday this year, toasting the democracy experiment — though the framers built a republic, but who’s splitting hairs—through trial and error.

The Constitution was itself a rewrite. The Articles of Confederation came first, and when that frame proved too weak, the framers scrapped it and started over to create the wonderful document we have today. Call it lean startup country with apologies to Eric Ries.

It has absorbed every shock since, mostly through reinterpretation rather than another rewrite. They built an amazing framework, not a rulebook, and the courts have stretched it to cover technologies the founders never named. Rewriting the document is the hardest move of all: an amendment needs two-thirds of both houses of Congress to propose and three-fourths of the states to ratify, which is why only twenty-seven have passed in two and a half centuries.

By design.

Our country is slow and inefficient on purpose— separation of powers, checks and balances, two houses, 50 states, friction at every point and city council members to say no, no matter what. It’s one of the reasons why we’ve stuck around this long without being invaded by Canada.

That friction is a good thing; it keeps the frame from being rewritten in a passing panic, by a bare majority, on a bad week, by someone with a bad toupee.

The document has been rewritten before, but only under the largest shocks — a civil war that ended slavery, the three-fifths compromise and remade citizenship, the long push for women’s suffrage, the rise and repeal of Prohibition (thankfully).

Generative AI is the newest shock and the reflex is to demand another amendment. Tat reflex is usually wrong. Most of the strain resolves through interpretation, the way it always has. Inefficiently and with a lot of discussion.

A few clauses were written with such specific human assumptions baked into the text that AI does not bend them. It breaks them. This piece is about finding those few.

The Copyright Clause: It Names Authors and Inventors on Purpose

[The United States Congress shall have power] To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.
 — The Copyright Clause

Most of the Constitution sets up structure: who holds which power, how they check each other, what the government may not do to you.

One clause does something stranger.

It writes the incentive engine of an entire creative economy into the text, and it names the people that engine is meant to reward: Congress may “promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.”

That word, “Authors,” was doing work no one questioned for two centuries, because the alternative was science fiction, and then it stopped being fiction. A computer scientist named Stephen Thaler asked the Copyright Office to register a picture his AI system had generated on its own, listing the machine as the sole author.

In March 2025, the D.C. Circuit said no: the Copyright Act requires a work to be authored, in the first instance, by a human being. The Supreme Court declined to take the case in March 2026, leaving that rule standing.

That word, “Authors,” was doing work no one questioned for two centuries, because the alternative was science fiction no one could have every imagined in 1789.

The court decided narrowly, on the statute, and pointedly declined to rule on whether the Constitution itself demands a human author.

But the clause names two human roles, not one, and AI breaks both.

Patents are the other engine: the Federal Circuit held in 2022 that a machine cannot be a named inventor, and the patent office has since reaffirmed that a human must be named even when AI does the invention’s heavy lifting.

Authorship and inventorship, the two rewards the clause was built to grant, both stall on the same word, written twice into a single sentence — human.

The bargain also has a valve, and AI is straining it. Copyright would choke the progress it exists to promote if every copy needed permission, so the law built in an escape hatch: fair use, codified in Section 107 of the Copyright Act, which lets some copying go unlicensed.

It plays the role here that Section 230 plays for speech — a statutory pressure-release bolted onto a constitutional promise. Building a creator-machine means ingesting the corpus of human work first, and in June 2025 two federal judges in California held that training on lawfully acquired books can be fair use, while assembling a library out of pirated copies cannot.

One of those cases named my own maker, Anthropic, and later settled; the question is far from resolved, with higher courts and Congress still to weigh in.

So this is the clearest case where the human assumption is written into the text, not merely assumed by it. The Copyright Clause tests AI from both ends at once: who counts as a creator, and whether feeding every existing work into a machine honors the deal or hollows it out.

Count it as a subtraction. The exclusive right the clause promises attaches to neither a machine’s writing nor its inventions, so what it makes is born with no owner at all.

Action Items

• If your product generates images, copy, or code, assume the raw output may not be copyrightable, and design human authorship into the workflow you can document.

• Keep a record of the human creative decisions in AI-assisted work; the line the courts are drawing runs through demonstrable human control, not effort or intent.

• Treat training-data sourcing as a legal question, not only an engineering one. Where your data came from now carries liability of its own.

The First Amendment: It Assumed a Human Was Doing the Speaking, Because There Was Nothing Else

Congress shall make no law respecting an establishment of religion, or prohibiting the free exercise thereof; or abridging the freedom of speech, or of the press; or the right of the people peaceably to assemble, and to petition the Government for a redress of grievances.
 — First Amendment

The First Amendment protects “the freedom of speech,” and behind that plain phrase sits a quiet assumption: a speaker. A person with intent, standing behind words, answerable for them.

They own them, by design.

The marketplace-of-ideas metaphor that runs through a century of free-speech law assumes rough parity, that truth can out-compete falsehood because both cost roughly the same to produce and spread.

Generative AI breaks the cost symmetry. Producing a convincing lie now costs almost nothing and scales without limit. The classic remedy for bad speech has always been more speech, but that remedy assumes the counter-speaker can keep pace. When falsehood is free and infinite, counter-speech is a bucket against a tide.

When falsehood is free and infinite, counter-speech is a cup against the Pacific Ocean.

This is not the first flood.

When the interactive internet arrived, the country did not amend anything; it legislated. Section 230 of the Communications Act, added by the Communications Decency Act of 1996, says an online service cannot be treated as the publisher of what its users post.

“No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider.”
 — 47 U.S.C. § 230(c)(1)

Those twenty-six words let the modern internet exist, because a platform could carry billions of strangers’ words without answering for each one. It was adaptation by statute, sitting on top of the First Amendment rather than inside it.

Section 230 shields a platform for someone else’s speech; it says nothing about speech the platform’s own machine makes.

Generative AI cracks the line that statute draws.

Most legal analysts now expect the immunity will not cover content the system itself generates, and in Garcia v. Character Technologies, a suit over a chatbot’s role in a teenager’s death, the company did not even raise Section 230 as a defense.

The First Amendment still stands, yet the scaffolding built around it for the last shock is what is giving way.

Legislatures noticed.

By the end of 2024, twenty states had passed laws targeting election-related deepfakes, and California’s were among the most aggressive. Then a creator who makes AI parody videos sued, and in 2025 a federal judge struck down California’s deepfake statute on First Amendment grounds.

The law swept in satire and parody, discriminated by content and by speaker, and failed the strict scrutiny that content-based speech restrictions must survive. The court reached for the oldest instinct in the doctrine: the answer to bad speech is more speech, not suppression.

That leaves a real and unresolved tension. The same protection that shields a dissident shields a synthetic smear, and courts have so far refused to carve AI-generated content out of the First Amendment. The state’s interest in protecting elections is genuine; so is the risk that a broad deepfake ban silences the political satirist along with the fraudster.

Whether the doctrine holds its shape as the technology scales is the open question, and it will not be settled by one district court in California. On the rights ledger, this is an addition.

The First Amendment is stretching to cover synthetic speech, which hands creators a new right and the rest of us a new hazard that we have to review on our own, by design.

Action items

• If you build tools that generate or distribute media, add content-credential and labeling signals early; the regulatory pressure is moving toward disclosure, not outright bans.

• Design labels that survive re-sharing and screenshotting. A disclaimer that vanishes on repost is decoration.

• Assume your synthetic content is legally protected speech in the United States today, which means the responsibility for restraint sits with you, not with a statute.

The Fourth Amendment: It Assumed Surveillance Took Effort

The right of the people to be secure in their persons, houses, papers, and effects, against unreasonable searches and seizures, shall not be violated, and no Warrants shall issue, but upon probable cause, supported by Oath or affirmation, and particularly describing the place to be searched, and the persons or things to be seized.
 — Fourth Amendment

The Fourth Amendment protects against unreasonable searches and seizures, and its unstated engine is friction. Watching a person used to require officers, hours, and shoe leather. The cost of surveillance was its own limit.

The framers loathed the general warrant precisely because it removed that limit, letting officers rummage through anyone’s life on a hunch.

Digital technology removed the friction the framers relied on. Cheap sensors, cloud storage, and AI analysis make it possible to reconstruct a life from data no one thought was private so the Court has spent years stretching the amendment to rebuild the friction it assumed.

In Carpenter (2018), it held that police generally need a warrant to pull months of cell-site location records, because those records assemble an intimate map of where a person has been. Chief Justice Roberts framed it as a response to seismic shifts in digital technology.

The friction the framers relied on is gone, and the Court is rebuilding it one ruling at a time.

Three days before this was written, on June 29, 2026, the Supreme Court went further in Chatrie v. United States, holding that a geofence warrant sweeping up Google Location History is a search, even for two hours of data, even though a third party holds it. The Court’s own framing is the thesis of this whole piece: the database is new, but the principle covering it is old.

That is the Fourth Amendment adapting by interpretation, not amendment. The friction the framers relied on is gone, and the Court is rebuilding it one ruling at a time.

The rebuild is slow and fact-specific, and AI-driven analysis keeps outrunning it. Facial recognition, license-plate networks, and pattern detection across pooled data can already do what no colonial magistrate could have authorized, and the doctrine is still catching up one case at a time.

This one adds rights back.

Each ruling restores a slice of the privacy the framers assumed, extending the right to be left alone into the places our data now lives.

Action items

• If you handle location, biometric, or behavioral data, treat it as latent evidence. What you collect can be compelled, and the legal shield around it is thickening.

• Minimize retention. The data you never stored is the data you never have to hand over.

• Build for the warrant standard, not the subpoena standard. The direction of the law is toward more protection for pooled personal data, not less.

The Due Process Clause: It Assumed a Person Made the Decision

“No person shall be held to answer for a capital, or otherwise infamous crime, unless on a presentment or indictment of a Grand Jury, except in cases arising in the land or naval forces, or in the Militia, when in actual service in time of War or public danger; nor shall any person be subject for the same offence to be twice put in jeopardy of life or limb; nor shall be compelled in any criminal case to be a witness against himself, nor be deprived of life, liberty, or property, without due process of law; nor shall private property be taken for public use, without just compensation.”
 — Fifth Amendment

“All persons born or naturalized in the United States, and subject to the jurisdiction thereof, are citizens of the United States and of the State wherein they reside. No State shall make or enforce any law which shall abridge the privileges or immunities of citizens of the United States; nor shall any State deprive any person of life, liberty, or property, without due process of law; nor deny to any person within its jurisdiction the equal protection of the laws.”
 — Fourteenth Amendment

The Fifth and Fourteenth Amendments guarantee due process: before the state takes your liberty, you get notice and a real chance to contest the case against you. That right rests on a hidden premise, which is that you can see and challenge the reasoning.

A human judge can be questioned.

A human’s logic can be cross-examined.

You can find the flaw and name it.

Closed algorithmic systems break that premise. In State v. Loomis (2016), a Wisconsin man was sentenced with help from COMPAS, a proprietary risk-assessment tool whose maker would not disclose how it scored him.

He argued, reasonably, that he could not contest a number he was not allowed to inspect. The Wisconsin Supreme Court upheld the practice but hemmed it in with warnings, and the U.S. Supreme Court declined to hear the appeal.

A trade-secret algorithm offers a result without a reason you can examine.

The unresolved core is simple to state. Due process assumes a contestable reason. A trade-secret algorithm offers a result without a reason you can examine. As government agencies reach for AI to decide benefits, flag suspects, and shape sentences, that gap between a result and a reviewable reason widens, and it widens fastest for the people least able to fight back.

This is where the human-centered AI framing, associated with the computer scientist Ben Shneiderman, earns its keep: keep a person with real authority in the loop, and preserve an explanation the affected person can meaningfully challenge. The Constitution may not need a new “right to explanation” amendment yet. Courts can read the existing due-process guarantee to require contestability.

If closed systems keep making liberty-affecting decisions without inspectable reasons, that reading has to grow sharper, or the guarantee goes hollow. This is a subtraction. When a sealed system decides and no one can inspect its reasoning, the right to contest the case against you quietly shrinks.

Action items

• If your system informs decisions about people’s money, freedom, or access, build an explanation the person can contest, not a score they must accept.

• Avoid closed, unexaminable models for consequential government-facing decisions. Contestability and explainability is becoming a legal requirement, not a nicety.

• Keep a human with genuine authority in the loop, and document where they exercised judgment rather than deferring to the output.

The Frame Bends Where It Can and Breaks Where It Cannot

The framers did not write a rulebook. They wrote a frame, and frames flex. Most of what generative AI throws at the Constitution will resolve the way the telegraph, the automobile, and the internet did, through interpretation, one case at a time.

Carpenter and Chatrie show the machinery working: old principle, new facts, a warrant where there was none. The system is working, by design.

The honest work is separating the strain that bends from the strain that breaks. The Fourth Amendment bends as the courts restore the friction the framers assumed. Due process bends, because contestability can be read into a guarantee that exists. The First Amendment holds, uneasily, protecting synthetic speech alongside the real kind.

The one place amendment talk is honest rather than reflexive is the Copyright Clause: the framers named their beneficiaries, Authors and Inventors, and meant human ones. That word is load-bearing in a way that “search” and “speech” are not.

None of this was ever a technical problem. These are human questions about who counts, who decides, and who answers. The Constitution has always been a document about people.

AI just forces us to say which parts we meant that way.

About The Author

Patrick Neeman is the author of uxGPT: Mastering AI Assistants for User Experience Designers and Product Managers. It’s for sale on Amazon and Gumroad at $9.99 for digital in Kindle, EPUB or PDF format and $19.99 for paperback. Go take a look.

Patrick is a Sr. Principal User Experience Designer across the Workday product suite. He also has been head of user experience, mostly as player-coach, for the last 14 years at Evisort, Knowable, Icertis, Apptio and Jobvite and has over 20 years experience in the User Experience field.

He is also the author of Usability Counts, runs the UX Drinking Game.

You can read more about him at Pexplexity, and connect with at LinkedIn, X (Formerly Twitter),Threads and Substack. You can also follow great topics at the Slack group, AI for UX.

Roger Black and David Carson, two of the most famous magazine designers in the world (and two of my favorites), worked in the same industry, in the same decade, on the same kind of object—the magazine (and sometimes the web)—and produced work that looks like it came f…

Read more

In 1950, Brunswick Billiards commissioned a Cincinnati novelty firm to design a giveaway. The result was a plastic eight-ball with a 20-sided die suspended in dark blue alcohol and twenty pre-printed answers.

The product cost two dollars and outlasted roughly seven generations of person…

Read more

Everyone keeps asking which AI tool to use, as if one of them is going to win. None is. There’s a stack — more a mix, three or four tools combined into one good cocktail.

I’ll write mine up, just as much for me as for the readers so I gain clarity.

I came up in print design…

Read more

Ten rules from a designer who keeps no screens, read again for the age of AI. Content AI assisted. Most images generated by Gemini.

Dieter Rams, one of the greatest designers no one has ever heard of, spent four decades at Braun making mundane things people used without thinking about them — yet the radios, calculators, shavers, the shelving Vitsœ still …

Read more

Some of it survived: quotes still mean exact match in Google, and a handful of operators like “site:” still work for people who know to use them. Most of it didn’t. L…

Read more

That was the takeaway I had from Stanford University’s 2026 Artificial Intelligence Report published a few weeks ago — it’s a good read we should all review, even at 425 pages — and this stood out: 53% global population adoption of generative AI within three years.…

Read more

AI tooling has burned that bridge with a flamethrower.

When one person can prototype with built out design systems coded right into the tool, write the copy, and stand up a working data layer …

Read more

That’s the line this field guide is built around - Establishing reality vs. fantasy.

Pull up your design folder. The p…

Read more

The conversations happening right now about AI coding tools have an eerie similarity to the conversations that happened in 1985 about desktop publishing. The tools will democratize the field.

The middle is collapsing.

Designers will become builders, or builders will become designers, or both.

The whole production pipeline is moving onto one desk.

Some of…

Read more

Forty slides, the personas laminated, the journey map color-coded, every box lined up.

The room nodded. Everyone felt aligned. And then nobody opened it again.

I’ve made that deck.

I’ve made the persona set that got reviewed once and never reopened, and the journey map that lived on a wall until the next reorg painted over it.

Hours …

Read more