AI lets a solo founder produce a product, but it can’t help them understand whether that product is actually good…yet. That expertise, the hard-won judgment underneath it, has always been what decides who wins and who quietly (or epically) fails.

There’s a movement in the startup world that currently claims if you can’t build your product by yourself, with AI, you shouldn’t be building it at all. The “advice” suggests that a solo founder should handle coding, design, and marketing with AI and should not hire a traditional team at all.

Hiring someone to design or engineer your product, whether it’s an agency, a contractor, or an early employee, is treated as proof you don’t really understand your own product and a growing number of investors won’t back a founder who would do that.

Most founders are doing exactly what they’ve been told works, and often they don’t know any better. The idea that needing help is a weakness sets a tone that will eventually backfire because most people are not unicorns, because unicorns do not exist.

Paul Graham, who founded Y Combinator, has said for years that he prefers founders who can build the thing themselves. Carles Reina, an early backer of ElevenLabs, told CNBC his first instant red flag is a founder who “literally cannot build products.” More than a third of new startups are now solo-founded, because AI made it possible to go from idea to working product without the people who used to be required.

Buried in that whole position is an assumption that’s dangerously misguided. It assumes that because you can now produce the thing, you have the expertise to know whether the thing is any good. Those have never been the same thing.

AI collapsed the cost of making something to almost nothing, but it did nothing to impact the cost of judgment, the years of studying how people behave, what makes a product feel trustworthy.

We got product creation for free, but forgot that expertise didn’t come with it.

Being able to vibe-code an app over a weekend makes you an expert in exactly one thing. Can you guess? It’s in vibe-coding an app over a weekend.

I lean on these tools as hard as anyone, on fast, super AI-heavy schedules that pressure test the technology every day. Even at that speed, it’s obvious how often what comes out is close but not right, and how often it takes real experience and hard-won knowledge to catch the difference and make the call.

None of this is a knock on the founders who build a product themselves. Seriously, kudos to them. A founder who vibe-codes their idea into something real is doing something genuinely good, they’re getting what’s in their head out into the world, fast, and that has always been the hard part.

The trouble starts when the people funding them treat that first draft as the finished product, and think asking for help is a weakness.

AI moves the thinking rather than removing it. The people cheering the death of the expert have confused two different things, making what you asked for, and knowing whether what got made is any good.

Where they’re right

Dario Amodei, the CEO of Anthropic, the company behind Claude, has said he thinks the first billion-dollar company run by a single person could arrive as soon as 2026, and he put the odds at something like seventy to eighty percent. He has an obvious stake in that being true and his company creates tools that would make it possible, but nevertheless, the prediction could be realistic.

Venture capitalist and Harvard Business School professor, Jeff Bussgang, has described the same shift from the investing side, right down to a betting pool among his tech-CEO friends over the year the first one-person billion-dollar company shows up.

The cost of producing software, a brand, a marketing campaign, a working prototype, has fallen through the floor. A founder with the right AI stack can now do in a weekend what used to take a team a quarter.

Andreessen Horowitz says that building has gotten so cheap that founders now skip the questions they used to have to answer first. Their Speedrun team’s perspective is that the interesting question is no longer “can you build it,” because the answer is now always yes. The interesting, and perhaps more important question is “should you,” which is what it’s always been. Marty Cagan has been telling product teams for decades that the job is to ask not just whether you can build the thing, but whether you should build it at all.

When making the thing is nearly free, the discipline shifts from execution to judgment about what deserves to exist at all.

The money has unsurprisingly moved accordingly. At TechCrunch Disrupt, seed investors including Elizabeth Yin urged early-stage founders to lean on AI coding leverage as much as they can, and Renata Quintini of Renegade Partners agreed it’s the right move for finding product-market fit, rather than spending early cash on headcount they don’t yet need.

Rebecca Lynn points out that it’s never been easier to stand up a company “in demo mode,” which is exactly why she’s careful about what she funds.

The thing these investors are reacting to is that a lot of the roles that used to be mandatory at the start of a company are now optional, or at least deferrable.

Startups have a long history of getting too big, too early. The Startup Genome Report, which analyzed more than 3,200 startups, found that premature scaling, hiring too many people too soon is the single most common reason startups fail.

Starting with a small team is generally smarter, but a founder who can build the first version themselves, who understands the problem because they live it, who doesn’t need to hire a translator to turn their idea into a thing, has a real and durable advantage.

The clearest case is the technical founder building for people exactly like themselves, the engineer making tools for engineers, who is both the builder and the user and genuinely doesn’t need anyone in between.

The trend is shifting to the founder building it alone, but the reality is they often shouldn’t.

Travis Howell and Todd Hall looked at whether solo founders can hold their own against co-founded teams, and found that they can, under one condition, when the solo founder has deep, broad experience to draw on.

It works when the founder’s real expertise is in building the thing, the craft, the risks, the details that make a product good, not just in the problem it solves. A technical founder building for engineers can do it, because what they know maps onto what the product needs.

Knowing an industry cold, or a market, or a customer, is valuable, but it is not the same as knowing how to build something people want to use.

Right now, the industry keeps confusing the two. Thinking a tool will fix it is just foolish.

Telling a founder that reaching out for help, from someone who could genuinely add qualified perspective or experience, is a red flag gets it backwards. It punishes a person for not personally holding every bit of knowledge a company needs, which is a bar no real human should be held to.

Eight people in a conference room

On a Saturday morning in 1999, five of us met at iXL’s offices in Manhattan’s Union Square to stage a quiet coup. We picked the day on purpose. The place was empty, no account directors, really nobody around to hear what we were about to do. Just us gathered around a table voluntarily sequestered in a conference room to discuss a project we thought we could save by making it smaller.

iXL was one of those agencies that got enormous by swallowing other agencies, which is what was common during dot-com 1.0. Agency size wasn’t a virtue or a sin, it was just the reality of the industry. The project we were on had at least forty people attached to it, with one of the largest budgets in the company, and it was sinking under its own weight.

There was too much process, too many people who had to be looped in before decisions could be made, a launch date that kept on sliding.

This wasn’t a new phenomenon (not that I knew that at the time). In 1975, computer scientist Fred Brooks watched IBM try to rescue a late software project by adding programmers, and watched it get later. He turned it into Brooks’s Law, which says that adding people late to a project just makes it even later.

It’s a brutal, mathematical reality. As you add people, each new person you add has to be brought up to speed, which pulls the people who already know what they’re doing away from the work. Every new person adds a communication line to every other person, so the coordination cost climbs faster than the output does.

Forty people don’t move four times faster than ten. They move slower, they cause blockers left and right, but they feel busier doing it.

We wanted to cut the team to eight, with severe focus. Eight specialists, each the best in the building at one specific thing. The cast list we wanted:

• A front-end developer
• A couple of back-end developers
• A project manager
• An information architect (which is what you called a UX designer back then)
• An art director
• A creative director
• A designer

The premise wasn’t to fire the other forty members of the team. They’d have plenty of work to do on what was next on the roadmap, but to actually ship this thing, we needed the smallest possible group of people who each could use their expertise to collaborate and move this forward smarter and faster.

Years later, Amazon would make this famous as the two-pizza team, Jeff Bezos’s rule that no team should be bigger than two pizzas can feed, somewhere around ten people, because that’s the size that stays fast.

I was twenty-three. I was nowhere near the most senior person in that room, though I was somehow the most senior information architect on the project in all of New York, which should tell you how little a title and actual expertise have to do with each other.

I was also, for the record, a bit of a troublemaker, which is how you end up in an empty office on a Saturday trying to shrink your own project by eighty percent. We wrote the plan, we walked our account director through it, and somehow he said yes.

Then we had to go do it, and we did.

That little team became a precision instrument the company reached for whenever something hard had to get built fast, and I’m confident it made them a lot of money.

As for the giants that were busy getting bigger, most of them did not make it. iXL shrank drastically. MarchFirst, the largest of these roll-ups, a nine-thousand-person, six-billion-dollar web consultancy, burned through a billion and a half dollars in fourteen months and filed the biggest bankruptcy in the industry’s history.

The whole sector contracted by seventy to eighty percent. Being huge turned out to be the most expensive way to be fragile.

That is all to say, the idea of small, nimble teams is rarely a bad idea, so when a venture capitalist says a founder doesn’t need forty people, I can’t argue with that.

The thing to understand is the strategy of what we did in that room, because that is not the argument for the solo founder show being encouraged today.

We cut forty people down to eight experts. The pitch today is to cut it down to one founder and a tool. Those may sound similar, but one of them concentrates expertise, while the other removes it and hopes the software quietly fills the hole where it used to be.

Patrick Neeman argues that in the age of AI, when anyone can stand up the same model and the same features, the thing that sets your product apart is the experience and the judgment built around it. You own what differentiates you, the workflows, the design, the taste, the deep understanding of your user, and you buy or borrow the commodity underneath.

The mistake in the solo-founder pitch is assuming the tool handed you the part that differentiates. All it really handed you was a commodity.

Ethan Mollick, who is about as bullish on AI at work as anyone, insists that AI rewards expertise. It lets the people who already know what good looks like judge the output, catch the confident mistakes, and push the tool past its defaults, while everyone else takes whatever the machine hands them.

The AI builds you a slick dashboard and never thinks to ask what it looks like before there’s any data in it, or when something breaks. It gives you a clean screen with light gray text on white, and someone who knows catches that a lot of your users can’t read it. That stuff only gets caught when someone with real experience is deciding whether it’s actually done or even takes the time to ask the AI to review it.

Tao An, a researcher at Hawaii Pacific University, lays out the split cleanly, saying that on routine work, AI lifts a beginner up to a competent floor, which is real and useful. On the hard, judgment-heavy work, however, the kind that decides whether a product is actually good, it does the opposite. It widens the gap between the person who knows what they’re doing and the person who doesn’t, because only one of them can tell when the confident-looking answer is wrong.

The tool amplifies the expertise in the room. It does not supply it.

At the end of the day, this comes down to one thing: the expertise has to be in the room. An employee, a contractor, a cofounder, a partner, or you, if you genuinely have it. Assemble it however you want, and keep the team as small as the work allows. The tool will make almost anything you ask it to. What it cannot do is tell you whether the thing is any good. That part has always belonged to the people who know, and it still does.

The thing you can’t prompt

In 1986, Steve Jobs needed a logo for NeXT, the company he started after being pushed out of Apple. He hired Paul Rand, probably the most famous graphic designer alive, and asked him for a few options to choose from. Rand said no. He would solve the problem, he told Jobs, hand back a single answer and Jobs could take it or leave it.

He charged a hundred thousand dollars. He came back with one design, a black cube tilted at twenty-eight degrees, presented in a bound book that walked Jobs through every decision, why the word, why the lowercase “e,” why the angle, until the logo felt less like a choice and more like the only possible conclusion.

Jobs, who genuinely had plenty of taste himself, paid six figures for something he couldn’t produce on his own. Whatever was in his head, he could have generated a thousand tilted cubes before lunch, but the reasoning underneath that piece of design derived from years of accumulated judgment that let Rand know, with total confidence, which single answer was right and why.

This is the gap the whole “just build it yourself” argument ignores. Producing the deliverable (whatever it may be) and knowing whether the deliverable is good are completely different skills.

Right now, you can easily generate a logo, a landing page, an entire app, in an afternoon. What you cannot necessarily do is hold the judgment to look at what came out and know that the hierarchy is off, that the flow asks too much of a tired person, that the tone is wrong for the people you’re trying to reach, that the confident, finished-looking thing on your screen is quietly, expensively broken or costing you customers.

Christian Holst found that the average checkout asks for around eleven form fields when most need eight, and that nearly one in five people abandon a purchase simply because the process felt too long. In testing, cutting a form from eleven fields to four lifted completions by 120 percent. The tool that built you that form has no idea it’s costing you customers.

AI makes everything it spits out look “good.” It can be clean, modern, professional, but that’s not always a perk. It’s also pulling from the same places everyone else is pulling from. When everyone builds with the same tools, trained on the same examples, everything starts to look the same.

Aurelie Radom suggests that AI may not have democratized design judgment, but it did democratize output, and when the models are all trained on the same stuff, the output space compresses and products that have nothing to do with each other start converging on the same shapes and the same layouts.

Alwin de Rooij and Michael Mose Biskjaer compared nineteen studies around this and found the effect is real and very measurable. The more people lean on the same AI, the more their work converges, which means looking finished and modern is fast becoming the least distinctive thing a product can be.

Looking finished used to mean something, odds were decent that someone who knew what they were doing added a contextual lens. That’s not true anymore. AI gives you the finished look in seconds, whether or not a single smart decision went into it.

Humans, it turns out, are built to fall for this distraction. In 1995, Masaaki Kurosu and Kaori Kashimura put a set of interface layouts in front of people and found the better-looking ones got rated as easier to use, even when they flat-out weren’t. They called it the aesthetic-usability effect.

We see something attractive and our brain quietly decides it works.

AI produces attractive by default, which means it produces trust it hasn’t earned because the thing it created looks great, you believe it’s good, but there’s no proof it actually will work.

Fabrizio Dell’Acqua, Ethan Mollick, and their colleagues studied 758 consultants and found that on the tasks AI is bad at, the ones who relied on it did worse than the ones who didn’t. The AI gave them wrong answers that looked right, and the people who couldn’t tell the difference used them anyway.

Separately, Yue Liu and colleagues studied a large batch of AI-written code and found it full of quality and maintainability problems hiding under a finished surface.

This isn’t a secret in our field. Nicole Alexandra Michaelis calls it the “taste paradox,” where we all agree that knowing what’s good is the skill that matters most now, and we’ve built a way of working that keeps people from ever learning it. The more you lean on AI to tell you what works, the less you ever figure out how to tell on your own.

Ruiqi Zhou adds a business lens to this, saying that if everyone can make the same thing, making it isn’t an advantage anymore. The only thing still worth anything is the judgment to know what’s worth making.

The receipts

This isn’t an emotional argument about how things look. The numbers back it up. The business value of expertise is one of the most heavily documented things in product, and the data got sharper once AI showed up.

Hyunso Kim, Hyo Kang, and Jaeyong Song studied more than 160,000 product launches on Product Hunt, before and after ChatGPT arrived, and found that after AI showed up, solo launches spiked, especially in categories that used to require a team, but when they looked at quality, at which products actually reached the top tier, teams still dominated, as much as ever.

AI widened the door for solo entry and did nothing to close the quality gap. More things got made. The good things still came from the people with the expertise to make them good.

The money follows the same pattern. First Round Capital studied a decade of its own portfolio and found that companies with more than one founder outperformed solo founders by 163 percent, raised more, and grew faster, an advantage they tie directly to complementary skill and the sharper decisions it produces. Their data says the opposite of their advice.

That’s not a nostalgia story about the pre-AI days. Boston Consulting Group found that the companies leading in AI, the ones it calls “future-built,” are outgrowing their competitors by 1.7 times in revenue and 3.6 times in shareholder return.

These winners used real expertise to aim the new tools and didn’t just buy the tools to replace the expertise.

Benedict Sheppard, Hugo Sarrazin, Garen Kouyoumjian, and Fabricio Dore scored hundreds of companies on how seriously they treat design, then tracked their results over five years. The ones that took design most seriously grew revenue about a third faster than the ones that didn’t, and returned more than half again as much to their shareholders.

The same research shows why those winners are so rare. More than forty percent of companies still don’t talk to their users while they’re building, and over half have no real way to tell whether their own team’s work is any good.

Taking design seriously is what separates the winners from everyone else. Taking design seriously means being able to judge whether your design is actually working. Most companies can’t or just don’t.

If most companies can’t tell good design from bad in the first place, then of course they’ll believe AI made them something good. They have no way to know it didn’t.

Judgment is not the same as one powerful person’s gut, and the answer to “the tool can’t decide for you” is not “so let the loudest opinion in the room decide instead.” In 2009, Google’s lead designer Doug Bowman quit and explained why in a post called “Goodbye, Google.” The company, he wrote, couldn’t choose between two shades of blue, so it tested forty-one.

He’d been asked to prove, with data, whether a border should be three, four, or five pixels wide. That’s the opposite mistake, and it’s just as bad. When no one on the team can just say “four,” you test everything, because nobody has the expertise, trust or standing to make the call.

Real expertise is knowing which decisions you can settle and move on from in a second and which ones actually need testing.

Build small. Just bring the expertise.

Rick Rubin has produced some of the most important records of the last forty years, across rock, rap, country, and pop, with artists such as Johnny Cash, Jay-Z, the Red Hot Chili Peppers, and Adele.

When Anderson Cooper asked him, on 60 Minutes, whether he could play instruments or work a soundboard, Rubin said no. “I have no technical ability,” he said. “And I know nothing about music.”

Cooper, reasonably, asked what he was being paid for. Rubin’s answer sums up why an expert lens matters. He said, “The confidence that I have in my taste and my ability to express what I feel has proven helpful for artists.”

He can’t compose the tracks himself, but he is one of the most sought-after people in music anyway, because composition isn’t always the point. Judgment is.

The ability to sit with a thing and know, with conviction, what it needs and what’s getting in the way is the rarest and most valuable thing in any creative endeavor. It is exactly the thing that does not reside in a tool, no matter how good the tool gets.

The people telling founders they don’t need anyone but themselves and AI are half right in a way that matters. You certainly don’t need forty people. You may not even need ten. It’s fair to say that the bloat they’re concerned about is real, and killing that is good, but they’ve taken a true thing about team size and turned it into a false thing about expertise.

Fewer people, yes. Less expertise, no.

Right now, AI gives you the output of expertise without the expertise. It hands you the artifact a skilled person would have made, without the judgment that skilled person would use to know whether the artifact is any good.

That gap is real today, and it is real for the foreseeable future, though I won’t pretend to know how long “foreseeable” lasts. Maybe the models close it, maybe judgment turns out to be the last thing to automate.

Either way, you have to build in the world you’re actually in, and in this one, AI can make the thing, but it cannot tell you if the thing is good.

Founders actually doing this are the first to admit where it breaks. Anik Devaughn, who built an AI-native commerce platform largely by himself, said that founders who understand their business deeply can move faster and stay smaller, but the ones who can’t evaluate the work will just automate bad decisions.

Basically, if you’re a founder talking to an investor who says they won’t back anyone who hires help, the investor is missing the point.

A founder who can tell the difference between what’s working and what isn’t is a far safer bet than one who can only tell you the demo they put together looks great.

This is not an argument for hiring an agency, or hiring anyone in particular. It’s an argument for embracing the expertise that has to be there to create success. It can be an employee or a contractor, a cofounder or a partner, or you, if you’ve genuinely put in the years.

NYU Professor J.P. Eggers ran an experiment where MBA students used AI agents to build startups from scratch. The AI was good at discrete tasks and fast brainstorming, he found, but it couldn’t replace the judgment that comes from having real specialists in the room.

His line for what happens without them sums up a lot, saying “You’re kind of taking it on faith that what the AI is producing is pretty good.”

Assemble the smallest team you can. Keep it lean enough to move fast, and make sure someone on it can tell what’s right from what’s broken. AI can build anything. Knowing if it’s right takes a human.

AI or not, it all falls apart without someone who actually knows what they’re doing.

Anyone who treats a founder’s willingness to bring in that person as weakness has it backwards. A founder who knows what they don’t know, and goes and gets the help, is the one worth betting on.

References and further reading

The belief that you should build it yourself

• Diana Bocco, AI Is Letting Solo Founders Build Startups Without Hiring (Inc.). The solo-founder-with-AI movement in the founders’ own words, including the line that founders who can’t evaluate the work just automate bad decisions.
• Paul Graham, How to Start a Startup. The Y Combinator view that the best ideas are the ones founders can build themselves.
• Carles Reina (CNBC). The ElevenLabs backer on why a founder who “literally cannot build products” is an instant red flag.
• The State of Solo Founding (Carta data). More than a third of new startups are now solo-founded.

Where they’re right

• Dario Amodei (Inc.). The Anthropic CEO’s prediction of the first one-person billion-dollar company by 2026.
• Jeff Bussgang, AI and the Dawn of the 10x Founder. A VC and HBS professor on the same shift, and on AI multiplying expertise rather than replacing it.
• a16z Speedrun, The Idiot Index for Code. Building got so cheap that the real question moved from “can you?” to “should you?”
• Elizabeth Yin and Renata Quintini (TechCrunch). Seed investors on leaning into AI coding leverage at the earliest stage.
• Rebecca Lynn (via MLnotes). On how easy it is to stand up a company “in demo mode,” and why that makes her cautious.
• Startup Genome, Premature Scaling. Across 3,200+ startups, scaling too fast, hiring too many too soon, is the most common reason startups fail.
• Travis Howell and Todd Hall, Lone Genius or Lonely Fool? (Strategic Management Journal). Solo founders can hold their own, but only when they personally carry deep, broad expertise.
• Marty Cagan, Product Management: Start Here (SVPG). Decades of arguing that the job is to ask not just whether you can build it, but whether you should.

Small teams and the strategy of focus

• iXL and MarchFirst (Unicorn Burn). The dot-com roll-up giants that grew huge and collapsed; MarchFirst burned through $1.5 billion in fourteen months.
• Fred Brooks, Brooks’s Law. Adding people to a late project makes it later, because coordination cost climbs faster than output.
• The Two-Pizza Team. Jeff Bezos’s rule that the fastest teams are small enough to feed with two pizzas.
• Patrick Neeman, Buy Versus Build Is Now a UX Decision. Own the experience and judgment that set your product apart; buy or borrow the commodity underneath.
• Ethan Mollick, The Cybernetic Teammate. AI rewards expertise: the people who know what good looks like get the most out of it.
• Tao An (arXiv). On routine work AI lifts beginners up; on hard, judgment-heavy work it widens the gap between those who know and those who don’t.

Why output is not expertise

• Christian Holst, Baymard Institute. The average checkout asks for more form fields than it needs; trimming them can lift completions by 120 percent.
• Masaaki Kurosu and Kaori Kashimura, the aesthetic-usability effect. People rate better-looking interfaces as more usable, even when they aren’t.
• Aurelie Radom, AI Made Everyone a Creator, Not a Designer. AI democratized output, not design judgment, so products that have nothing to do with each other keep converging on the same shapes and layouts.
• Alwin de Rooij and Michael Mose Biskjaer, Does Generative AI Make Us Think Alike? A review of nineteen studies finding a real, measurable homogenization effect: the more people lean on the same AI, the more their work converges.
• Fabrizio Dell’Acqua, Ethan Mollick, and colleagues (Harvard Business School). On tasks AI is bad at, consultants who leaned on it did worse, because it produced wrong answers that looked right.
• Yue Liu and colleagues (arXiv). A large study of AI-written code found it full of quality and maintainability problems beneath a finished surface.
• Nicole Alexandra Michaelis, The Taste Paradox. Everyone agrees judgment is the skill that matters most, and the current way of working keeps people from ever building it.
• Ruiqi Zhou, Taste Is the New Moat, and That’s the Problem. If everyone can produce the output, the only thing left with value is the judgment to direct it.

The receipts

• Hyunso Kim, Hyo Kang, and Jaeyong Song (arXiv). Across 160,000+ Product Hunt launches, AI boosted solo entry but quality still concentrated in teams.
• First Round Capital, The 10-Year Project. Companies with more than one founder outperformed solo founders by 163 percent.
• Boston Consulting Group, Build for the Future. Companies leading in AI outgrow competitors by 1.7x in revenue and 3.6x in shareholder return.
• Benedict Sheppard, Hugo Sarrazin, Garen Kouyoumjian, and Fabricio Dore, The Business Value of Design (McKinsey). Companies that take design seriously grow faster, and most companies have no objective way to judge their own work.
• Doug Bowman, Goodbye, Google. The designer who left after being asked to prove, with data, which of forty-one shades of blue to use, the opposite failure: measurement used to avoid judgment.

The close

• Rick Rubin (60 Minutes). The producer who can’t play an instrument, and whose entire value is the confidence of his taste.
• J.P. Eggers (Fortune). The NYU experiment where AI agents built startups, and the judgment that was missing without real specialists.

A founder with a tool is not a founder with a clue was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

In 2014, Ethan Zuckerman, a renowned media scholar and Internet activist, apologised for his role in inventing the Internet’s most hated marketing tool: the pop-up ad.

All of us have screwed up situations in our lives so badly that we’ve been forced to explain our actions by reminding everyone of our good intentions. It’s obvious now that what we did was a fiasco, so let me remind you that what we wanted to do was something brave and noble.

In the late 1990s, Zuckerman led a team at tripod.com — a popular service which allowed users to launch their own websites for free. Zuckerman had tried countless ways to fund the growing business until concurring that the only way to keep the company afloat was to bring in advertisers. The format invented by his team — a pop-up ad — would not only boost tripod.com’s revenue but change the Internet once and for all.

With the advent of the pop-up ad came the boom of intrusive, aggressive, and persistent advertising. The industry has weaponised our lack of attention and filled every corner of the Internet with meaningless distractions and quick rewards. Pop-ups embodied everything wrong about the advertiser-centric approach: it appeared out of thin air and demanded attention like a capricious baby, no matter what was on the user’s mind or what they were occupied with. Pop-ups reduced users to targets for the omnipresent advertisers.

The Internet today is the realm of noise and chaos. Advertisers are more vicious than their predecessors, and ads themselves have transformed to appear more organic and blend into the content until it is impossible to tell them apart. Users continue to limp around this inaccessible desert, constantly bombarded with emails and notifications informing them of special offers, keeping them up to date with what their friends have been up to, and begging them to return for more mindless entertainment.

How does someone with an already disrupted attention span find the strength to concentrate when the entire world around them is vying for their attention? How do they make decisions, resist temptations, and struggle to conform in an environment that does not suit someone like them at all?

In this article, we will discuss people with ADHD, what it means to live with this condition, and why existing design frameworks for neurodivergent users address the wrong problem.

Understanding Attention Deficit and Hyperactivity Disorder

ADHD is a neurodevelopmental condition that affects a person’s ability to concentrate and control emotions and actions. Despite being commonly associated with hyperactivity (hence the name), there are several distinct types of ADHD: hyperactive and impulsive (rather self-descriptive), inattentive (characterised by distractibility and lack of attention), and combined (the most prevalent type); each manifesting differently in each individual patient. Most commonly, people (especially adults) with ADHD have problems concentrating on the task at hand, be it a work assignment or a lengthy conversation.

MRI studies of patients with ADHD show various dysfunctions in dopamine and noradrenaline transmissions. The disorder particularly impacts the prefrontal cortex, the area of the brain involved in regulating attention, motivation, and emotion, which in turn affects the patient’s ability to concentrate, control, and shift attention.

Children with attention-deficit/hyperactivity disorder show prominent abnormalities in the inferior PFC and its connections to striatal, cerebellar, and parietal regions.

ADHD is not curable but its symptoms can be reduced and managed. There are essentially two paths to regulating attention deficit: medication and therapy. Medication is commonly prescribed to children with ADHD to address the biological mechanisms of the disorder and regulate noradrenaline and dopamine transmission in the brain. Cognitive behavioural therapy is widely used to enable self-regulation and form coping strategies. This duality, however, is a mere reflection of a wider conversation around how ADHD should be viewed and addressed. Should we side with the biomedical discourse of ADHD and choose to view it primarily as a medical category, or should we focus on addressing the flaws in a system that punishes disobedience?

It is hard to deny that ADHD affects — at times, quite severely — the patient’s ability to exist and function within complex social structures that require discipline, obedience, and compliance. Children with ADHD achieve less in school than their peers and face a higher likelihood of engaging in criminal activities. ADHD is positively correlated with anxiety and depression. Adults with ADHD struggle with employment and marital problems. Suicide rates among both children and adults with attention deficit are higher on average than in the general population.

A popular interpretation of the positivist biomedical paradigm may suggest stimulants as the most appropriate answer to curbing “undesirable behaviour”. Medication’s proven ability to improve daily functioning benefits the patient as much as an institution that will gladly drug “children into docility…in the service of a tacit project of social control”. However, acknowledging that the system obsessed with producing identical, obedient workers has little tolerance for defiance does not require viewing stimulants as a tool for enforcing cooperation. Medication is not the only way to manage the symptoms of ADHD, and it is not always the most appropriate as similar, if not more powerful, effects can be achieved with individual and systemic tools.

A more balanced take suggests viewing ADHD as the product of a complex interaction between biological and social-environmental factors. It would recognise and address the biomedical mechanisms and create opportunities for people with ADHD to function, express themselves, and realise their potential in social scenarios. This take dismisses neither medication nor therapy: it views them as valid and effective tools in their own right, but offers a reframing of what is viewed as a problem in modern society.

Take, for instance, students with ADHD, who are talkative, distracted, hyperactive. Such students are only problematic if we choose to view them as such and refuse to make the necessary adjustments to leverage, not curb, their inattentiveness. A teacher who prioritises classroom discipline is a teacher who disregards the evidence that pupils with ADHD perform significantly better when given clear instructions and expectations. The teacher’s focus on obedience and order blinds them to a multitude of opportunities to maximise student performance without disrupting the learning process: for instance, by using pupils with ADHD as peer educators. By framing attention deficit as a condition associated with some deeply problematic, anti-social, disruptive behaviours, we unwittingly set pupils with ADHD up to fail in place of a far better simple solution:

Optimal pupil performance is associated with brevity and clarity of sequences of instruction, the accompaniment of verbal instructions with visual cues and the availability of resources that pupils can refer to for reminders of direction and expectations.

The same, of course, applies to other areas of life impacted by ADHD. Building sustainable relationships, pursuing meaningful careers, enjoying the simple joys of everyday living, demand adjustments — not only from the person with ADHD but from the system itself. Reasonable accommodations are not synonymous with special treatment: they are designed to empower people who are otherwise deprived of the opportunity to function in society, not to absolve them of the need to put in effort. Such accommodations create welcoming environments where people contribute to common goals in the ways that suit them best, and where their needs and challenges are acknowledged and addressed.

ADHD v Technology

The relationship between modern technology and ADHD is complex. On the one hand, smartphones, wearable devices, and other gadgets are becoming increasingly prominent in health management and emotional regulation, complementing therapeutic and medicinal treatments with handy personalised tools. To-do lists and AI chat-bots, remote symptom tracking and online testing — adults with ADHD have plenty to choose from depending on what they wish to achieve. In addition to that, the Internet has played a big role in raising awareness of ADHD and reducing stigma around the disorder. ADHD diagnoses have increased in recent years — which, considering that women are far more likely to remain undiagnosed than men, is a positive development. Technology makes it possible for people to discover ADHD, better understand its symptoms, exchange experiences, and receive treatment.

On the other hand, the digital world is a ruthless, fast-paced, overwhelming space that is neither welcoming nor kind to people with ADHD. The rates of “Internet addiction” (which stood at 14% of children and 20% of adolescents in 2006 and has since increased) are astronomical, and neurodivergent individuals who are known to have issues with substance abuse and addiction are among the groups most affected by the phenomenon. In people with ADHD, Internet addiction is positively correlated with inattention, hyperactivity, and impulsivity. Social media is especially harmful: engineered to maximise engagement, these platforms use all sorts of tools to make it nearly impossible to quit or even reduce screen time. Excessive use of social media in particular opens the door to a range of side effects, such as low self-esteem, feelings of inadequacy and underachievement, and increased social isolation.

However, it wouldn’t be fair to blame it all on social media alone. Most modern applications are designed for engagement and employ a wide range of tools to attract and retain users. One such tool is gamification — a popular technique that uses game mechanics in non-gaming settings to reward certain behaviours and entice continuous and regular use. As research demonstrates, rewards make even the most routine tasks far more exciting, allowing users to do chores, learn languages, and socialise with ease. For people with ADHD, however, gamification is a double-edged sword: while it certainly helps make learning more enjoyable, it also exacerbates ADHD symptoms. Instant gratification makes stimulation easy to achieve, forcing the user to seek more and turning mindful gamified experiences into never-ending quests for novelty. It is worth noting that because ADHD comprises at least two major domains — inattention and hyperactivity-impulsivity — the same technology that severely influences one may have little impact on the other. For example, this 2026 study found a positive correlation between the use of short video platforms and inattention but not hyperactivity-impulsivity.

The digital age has revolutionised how we create, distribute, and consume knowledge, making it far more accessible and affordable than ever before. The influx of new information and the ever-increasing speed at which new content is produced have transformed the way we live, amplified our anxieties and feelings of overwhelm. The abundance of options has further contributed to turning attention into an especially scarce and valuable resource. Today’s digital world is a terrifying place that takes users’ time and mental wellness in exchange for occasional dopamine hits. It is not designed with the neurodivergent user in mind, but it profits from their struggles: from encouraging fast, mindless interactions to enabling impulsive decisions.

Wouldn’t it be great if we could do something about this?

Why designing for ADHD is so hard

When it comes to designing for users with ADHD, recommendations commonly include reducing cognitive load, avoiding distractions, and relieving users of the need to rely on memory — solid tips that are little more than general UX principles, adjusted ever so slightly for the neurodivergent audience. These principles cannot replace a comprehensive design framework. As a complex disorder, ADHD poses multiple challenges: from social disability and academic and occupational problems to physical and psychological issues and risky behaviours. To design for ADHD means to address all these challenges at a fundamental level and, importantly, to understand why addressing them is essential in the first place.

In the book “Against Technoableism: Rethinking Who Needs Improvement”, Ashley Shew describes how most solutions are designed for people with disabilities but never with them: “[universal] design employs and imagines our bodies […] while rarely (if ever) simply asking us what we need”. What is worse, many of these solutions are built to correct or conceal disabilities rather than accommodate them. Our products, services, and environments are not simply built for the able-bodied — they force people with disabilities to adjust, to adopt particular ways of using and existing in them with the help of technology, to pretend that their disabilities are fixed or cured. Let’s medicate an inattentive student. Let’s use therapy to subdue them. Let’s agree on a design principle like “avoid clutter” in theory but allow a neurotypical middle manager to break it as they and their overlords please.

Take Bionic Reading — a method developed by Renato Casutt, a Swiss typographic designer, that emboldens the first letters of each word in a sentence to improve readability by guiding the reader’s eyes. Bionic fonts are sold for $79 per package of 10, and the company offers many other paid apps and products as well. Bionic Reading is a premium product with good advertising. It spread like wildfire in 2022, often positioned as the best possible solution for people with ADHD, dyslexia, and cognitive impairments.

Does Bionic Reading actually work? The company’s original testing method involving 12 participants (which, to the founder’s own admission, produced mixed results) does not withstand scrutiny. A real scientific test involving eye tracking on a large cohort of participants revealed that “the selective bolding of characters within words does not facilitate more efficient word identification or enhance overall reading performance in skilled adult readers”. Other tests show similar results. Bionic Reading claims that it keeps the reader’s mind focused, but in reality, it is a commercial product that tries to force a neurodivergent brain to act normally. As Daniel Lavelle at The Guardian asks: “Maybe having a wandering mind isn’t such a bad thing?” Would a better way to increase comprehension be to encourage neurodivergent people to read slowly?

So, what would it really take to design better products for people with ADHD? The process begins, unsurprisingly, with inclusion. A 2025 paper, “Towards Inclusive Guidelines for Web Design for Adults with ADHD”, describes co-designing such a framework with a panel of neurodivergent and neurotypical users. The recommendations produced by the panel varied from general UX tips (avoid complex menus, limit the number of options, keep navigation similar to other products) to ADHD-specific guidance:

• Avoid stock photos and decorative illustrations;
• Avoid components that may be confused with ads;
• Restrict scrolling to vertical only;
• Avoid non-standard hover effects and decorative animations;
• Indicate external links and buttons.

The paper assessed these guidelines, by having two designers with ADHD put together a prototype. Nearly all suggestions yielded promising results: users with ADHD found it significantly easier to navigate the prototype designed according to these principles and discover relevant information. In the neurotypical cohort, the results were relatively moderate, although non-ADHD users, too, benefited from simplified navigation, clear information hierarchy, and other improvements. Interestingly, the only design choice that was not well received was to use the Bionic Reading method, which reportedly reduced readability in non-ADHD participants.

Does that mean that checking the boxes above will make your product ADHD-friendly?

The answer is no.

Although structured layouts and simplified navigation definitely make products easier to digest, and rewards and incentives indeed keep users with ADHD engaged, these principles alone do not amount to an ADHD-friendly framework. By narrowing the recommendations down to a short list of checkboxes, we design for people with ADHD instead of designing with them. We attempt to make them “normal” rather than empowering them to be themselves.

Product-focused approach addresses the wrong problem. It simultaneously sands down a square peg and turns a round hole into a rectangle: both parts somewhat fit one another, and everyone pretends that they are not deformed and misshapen. It does not dare to challenge the system that produces square pegs and round holes: its job is to mitigate the damage, not ask questions. Its job is to make a pop-up ad less annoying, not to find a sustainable alternative. It is to have a phone number people in wheelchairs can call for assistance instead of building accessible urban environment.

Meanwhile, designing with ADHD is the process of co-creation, of mutual learning and exchange between neurodivergent and neurotypical partners. It entails, first and foremost, transforming the environment in which the collaboration occurs. A design framework for ADHD must therefore go beyond product considerations to inform product creation itself: from establishing healthy routines and enabling productive collaboration and reflection to protecting critical decisions against corporate interference.

Under this framework, a product team could add rewards to encourage daily usage, but cannot leverage anxiety and the fear of missing out that such rewards can create. This framework would suggest the team to apply the same principles to their work as well: set short-term, achievable goals, engage and reward team members, and respect each other’s time, space, and ways of working. This framework would ensure that the product is co-designed with neurodivergent users, not tested against some vaguely defined criteria. This framework would never give rise to another pop-up ad.

Is it such a surprise that this framework does not exist?

Pop-up ads are dead. The disdain for them was so strong that by 2004, when 6.2% of all ads on the Web were pop-ups, at least 20% of all Internet users had a pop-up blocker installed. Advertisers and tech corporations they funded did not cave in until new and improved formats came along. Pop-ups did not die because the public hated them. They died because something better was on the horizon.

Pop-ups are dead, but their story, their demise, and their complicated legacy make me dread the future even more than I already do. Corporations wish to cultivate users who conform, consumers who neither demand systemic changes nor require small accommodations. In the absence of regulation and external pressure, what could possibly motivate tech giants to consider people with disabilities, let alone involve them in making important decisions?

Pop-ups are dead, but people who pushed them into existence remain in charge: not the Ethan Zuckermans who inadvertently created monsters but decision-makers who breathed life into these creatures and set them loose. So long as they remain in power, so long as growth takes priority over customer experience and care, there will be no room for human-centric thinking or accessibility. So long as they continue to reduce all users to identical numbers in a spreadsheet, any considerations that do not immediately translate into revenue will remain an afterthought, forever postponed, abandoned at the bottom of the backlog.

Two years ago, a young, moderately stressed-out PhD student wrote an article about pursuing a doctoral degree while living with ADHD. In that article, the author discussed the struggles that people with ADHD face in academia, how bureaucracy and formalities are designed, as if intentionally, to crush all hope and excitement in neurodivergent people, and how simple accommodations and small adjustments may significantly impact someone’s career. That student was optimistic about the future — his own as a neurodivergent researcher, his peers’, and the tech industry’s.

In the time since, he “toughened up”. He filled out countless ethics forms, amendment requests, and research milestone reports. He completed standardised tests, passed exams, and handed in assignments. He faced every deadline, every requirement, every expectation head-on. There was nothing that could alleviate the pressure. The system itself was not interested in changing, so he changed his ways of doing things. He succeeded at some (not all) endeavours, though in the process, lost his passion for many things he had once adored. He became timid, dull, complacent.

Back then, he was excited about things to come. Today, I no longer see him in the mirror.

What to read

• HealthDirect. Attention deficit hyperactivity disorder (ADHD).
• ADDA. Types of ADHD in Adults: Understanding the Differences.
• Paul Cooper, 2008. Like Alligators Bobbing for Poodles? A Critical Discussion of Education, ADHD and the Biopsychosocial Perspective.
• Paul Cooper, Fintan O’Regan. Educating Children with AD/HD. A Teacher’s Manual.
• Louise C. Druedahl & Sofia Kälvemark Sporrong, 2019. A Foucauldian perspective on treating ADHD with medicine.
• Gyöngyi Fekete & Andrés Lucero, 2019. P(L)AY ATTENTION! Co-designing for and with Children with Attention Deficit Hyperactivity Disorder (ADHD).
• Ashley Shew, 2023. Against Technoableism: Rethinking Who Needs Improvement.
• David Skidmore. Inclusion: The Dynamic of School Development.
• Margaret D Weiss et al., 2011. The screens culture: impact on ADHD.
• Linda Wheeler, 2010. Critique of the article by Visser and Jehan (2009): ‘ADHD: a scientific fact or a factual opinion? A critique of the veracity of Attention Deficit Hyperactivity Disorder’

About the author

Kristian Mikhel is a PhD Candidate at the University of Tasmania, a product designer and a founder of Paper and Pain, a design agency for good, an accessibility advocate, and the author of the free newsletter, “The Accessibility Apprentice”.

Design v ADHD was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

Exposure to artificial intelligence runs through the tasks you do each week, not the role on your badge. Here’s how to measure yours and lower the risk.

My mom doesn’t understand my job, and it isn’t her fault. The closest I’ve ever come to explaining it is “I design how it works,” which is true and tells her almost nothing about my day.

It also works best when I’m wearing a black turtleneck and holding an iPhone, channeling my inner Steve Jobs.

That’s the curse of knowledge, and I’ve never directly defined my role in many of the places I’ve worked. Over decades I’ve managed design, product management, and engineering, and more than once I started without a job description. It wasn’t about my job title, it was about judgement applied through my experience building software.

Artificial intelligence doesn’t care about your job title or mine. It’s probabilistically predicting an outcome without looking at your LinkedIn.

In GPTs are GPTs, published in 2023, researchers from OpenAI and the University of Pennsylvania scored occupations task by task. They estimated that 80% of U.S. workers could see at least a tenth of their tasks affected, and about 19% could see at least half.

Those averages hide what matters for one person. Your exposure depends on how much of your week can be specified briefly and checked quickly, and seniority protects you only when it changes that mix.

For her work, she’s not exposed even though she’s in the models. From Claude:

The main trace of Jean Neeman in the art world is a single, well-documented piece: a teapot titled “Tea on the Go,” made in 1994 by Jean Neeman (American, b. 1948) from silver, copper, brass, and cocobolo wood, measuring about 8.4 x 7.75 x 4.5 inches. That puts her work alongside some serious company.

She doesn’t have to measure her risk because she’s an artist. I measure mine all the time so I can add value that AI can’t add, and that’s judgement.

The five steps below show how to measure yours, each with a prompt to run in Claude.

List Your Actual Tasks

Start with a week, not a job description.

List everything you spent more than half an hour on: the usability readout, the stakeholder review, the status update nobody read. Note the hours, how often each task repeats, and who reviews the result.

Also leave room for the white space between tasks.

Unblocking a teammate, the quick answer that saves someone a day, the thinking before a hard review: none of it shows up in a job description or on a calendar. It also goes unseen when people judge what you do, whether that’s a manager at review time or a researcher scoring tasks for exposure.

It’s often where you absorb complexity that, as Tesler’s Law puts it, can be moved but never removed.

Write it down anyway.

For a checklist, O*NET OnLine, the Labor Department’s occupational database, lists 31 tasks for Web and Digital Interface Designers. In Putting Tasks to the Test: Human Capital, Job Tasks and Wages, a 2009 paper, economists David Autor and Michael Handel found that tasks vary substantially within occupations, enough to predict wage gaps among people who hold the same job.

In Toil and Technology, written for the International Monetary Fund in 2015, economist James Bessen showed how automated teller machines took over dispensing cash and taking deposits. The tellers needed to run an average urban branch fell from 20 to 13 between 1988 and 2004, but banks opened 43% more urban branches, so the jobs held.

The title stayed the same while the job underneath it was rebuilt.

My own title changed in March, from director to senior principal designer. Neither title would tell you which of my tasks a model can do, but my calendar would.

Try It in Claude

Run these prompts in one Claude conversation. The first one drafts your week from your role, so you correct a list instead of writing one, and you can check its tasks against O*NET OnLine.

To map my exposure to artificial intelligence, don’t read my calendar or connected accounts, or ask me to describe my week. If you don’t know my title, ask once. Name the O*NET occupation you matched, and say if you’re recalling its tasks from memory. Draft 15 to 20 tasks from it, plus the undocumented work between them, for a normal week. Give hours, frequency, and reviewer for each, marked as guesses. Then have me correct it, including whether I manage anyone.

Score Each Task’s Exposure

Difficulty is a poor guide to exposure. In a 2023 experiment for Navigating the Jagged Technological Frontier, a team led by Fabrizio Dell’Acqua gave 758 Boston Consulting Group consultants realistic work.

With the model, they did far better on 18 tasks inside its abilities but were 19 percentage points less likely to get the right answer on one task outside them. Coauthor Ethan Mollick calls this the jagged frontier in Centaurs and Cyborgs on the Jagged Frontier: tasks that look equally hard can sit on opposite sides of an invisible wall.

Score each task with two questions instead. I call it the spec-and-check test.

• Specify the result. Ask whether you could describe a finished version in two sentences that someone else could follow.
• Check the output. Ask whether someone could confirm it’s right in less time than it took to make.

A task that’s quick to specify and quick to check is exposed, whatever your title says.

Twenty variations of an empty state pass both questions. Deciding which of three problems to fund this quarter fails both, and for now that decision stays yours.

Then weight the list by hours, because the Pareto Principle usually holds and a few tasks fill most of the week. A task that passes both and fills two days a week is where your risk sits.

When leadership reviews a pie chart of where a team’s hours go, you don’t want to be on the wrong side of it.

Passing the test means a model could do the task, not that one already does. In Labor market impacts of AI: A new measure and early evidence, from its own March 2026 usage data, Anthropic found Claude covering about 33% of computer and math tasks, against 94% rated feasible in GPTs are GPTs.

So the prompt below has Claude try one.

Try It in Claude

Paste this after the task table. It ends with Claude attempting your most exposed task, so you judge real output instead of a forecast.

Score each task in my table: could a finished version be described in two sentences, and checked faster than it took to make? Answer yes, partly, or no for each. Then say honestly if you could do each task today, fully, partly, or not yet, and why. Mark tasks with two yeses as exposed, sort by hours, and tell me what share of my week is exposed. Offer to try my top exposed task so I can judge it.

Check Your Career Stage

In the August 2026 revision of Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford’s Digital Economy Lab found no economy-wide job displacement in payroll data. Employment for workers ages 22 to 25 in exposed occupations, though, sits 19% below where it would be had it kept pace with less-exposed peers, mostly because companies are hiring fewer young people.

Experienced workers show no comparable gap, and the authors call these early indicators, not proof of cause.

Entry-level work is often the most specified on a team, handed down with clear instructions and reviewed by someone senior, which is where a model does best. Experience protects through judgment and context that rarely fit in a two-sentence spec.

A senior title over a week of status decks is junior exposure with a bigger paycheck.

If you’re early in your career, ask to own one small decision from start to finish. If you’re senior, hand your juniors judgment calls along with tasks.

Management will also change because there will be smaller teams.

Melvin Conway observed in 1968 that organizations build systems that copy their communication structures, which I cover in Conway’s Law. My extension is that much of management exists to carry that communication, so a smaller team needs less of it. Four people share six possible pairs, while 10 people share 45.

The flattening predates these tools. For The Manager Mass Exodus: How SMBs Are Flattening the Org Chart, Gusto studied payroll at about 8,500 small and midsize businesses. The share of their workers who manage anyone fell from just under 25% in 2019 to about 15% by late 2024. If you manage, move your hours from relaying information toward the decisions and coaching a smaller team still needs.

Try It in Claude

This prompt compares your task list against your experience. Correct it if it gets your background wrong.

Based on my title and experience, split my tasks into entry-level work (well specified and reviewed by others) and judgment work only I can do. If you don’t know my years of experience, ask once. Give the share of hours in each and, if I manage, how much goes to relaying information. Name the knowledge I bring that you could supply without my judgment. Tell me plainly if my week looks junior for my title, and suggest one decision I could own this quarter.

Deepen Your Domain Expertise

Knowing your field pays off more than knowing your tools. For Agentic coding and persistent returns to expertise, published in June 2026, Anthropic studied about 400,000 Claude Code sessions.

People made about 70% of the planning decisions — that’s human agency — and the agent made about 80% of the execution decisions.

The researchers rated expertise task by task and said it differs from job title: a senior engineer asking a first question about the Rust language counts as a beginner at Rust. Expert sessions succeeded about twice as often as novice ones, and a working grasp of the domain captured most of that gain, the same early-steep curve as the Power Law of Practice.

The person who understands the problem and not the model gets more out than the person who understands the model and not the problem.

Knowledge on its own doesn’t protect you anymore: If what you know can be pulled from a large language model without judgment to apply it, you’re at risk.

Companies don’t need you to tell them something they already know, they need the person who knows which answer fits this customer and this quarter, and the person who acts as connective tissue between people, building trust and getting decisions made rather than relaying status.

If you’re neither, your expertise is a search ot an AI Assistant result.

For designers, the domain is the user’s world and the business around it. Get closer to your users’ domain than your product manager is: read the regulations your customers follow, sit with support for a day, and learn how the money moves.

Try It in Claude

Answer the way you would with a new manager, without polishing. Short, unsure answers show the gaps fastest.

If you don’t know what my company sells, ask once. Tell me in one sentence who my product serves, and let me correct it. Ask me eight to 10 questions, one at a time, about their regulations, money, metrics, and workflow; “not sure” is a fine answer. Rank my knowledge gaps by how much they’d change my design decisions, suggest one way to close the top gap in 30 days, and revise the decision you suggested earlier.

Rebuild Your Portfolio

Employers expect 39% of workers’ core skills to change by 2030, according to the World Economic Forum’s The Future of Jobs Report 2025. A portfolio of polished final screens puts the fastest-changing part of your job up front and hides the part that holds its value.

Rebuild it around decisions: the problem you chose and the ones you turned down, the tradeoff and why, and what happened after launch. Show one case where you directed a model, caught its mistake, and shipped the fix.

Show the work a model couldn’t have done without you.

Look sideways too.

O*NET lists 10 related occupations for the same designer category, from software developers to graphic designers, and jobs overlap where their tasks do. As I wrote in We’ve Seen This Movie Before: Desktop Publishing Already Showed Us How AI Prototyping Plays Out, typesetting effectively ended as a profession while graphic design grew into web, motion, and product design. They all adjusted, and we’re seeing this move now too.

Whole economies have also made this move before.

The agricultural revolution let fewer people grow more food, and the industrial revolution absorbed the workers it freed. In 1900, 41% of American workers farmed, and by 2000 fewer than 2% did, according to The 20th Century Transformation of U.S. Agriculture and Farm Policy from the USDA’s Economic Research Service.

The adjustment wasn’t gentle, though.

In Engels’ Pause, economic historian Robert Allen found that British wages stalled for roughly half a century in the early industrial era while output per worker kept climbing.

People adjusted, but it took a generation, and the ones who fared best moved toward the new work instead of waiting for it to reach them.

The good news is that employers now have software to map those overlaps, something Henry Ford never had. When Ford ended the Model T in 1927, about 60,000 workers were laid off, according to Last Day of Model T Production at Ford from HISTORY. Skills platforms such as Skills Cloud, from my employer Workday, now relate the skills across a workforce.

In the report’s workforce strategies chapter, 85% of employers plan to upskill and 51% plan to move staff into growing roles. Organizations are looking to reinvest and reinvent.

Intentions aren’t programs, though, as Dan Maccarone argues in What the AI era owes you for UX Collective Editors. You own your career, not your company.

It’s up to you to figure out where the hockey puck is going, the advice Wayne Gretzky got from his father, Walter, and your task list shows which way it’s moving.

Try It in Claude

Run this on your strongest case study first. Then reuse it on the rest, and keep the list of adjacent jobs it produces.

Ask me for my strongest case study. Rewrite it around my decisions: the problem, the options I turned down, the tradeoff, how I used and corrected artificial intelligence tools (or that I didn’t), and the outcome. Ask for missing facts one at a time instead of inventing them. Then list four or five jobs whose tasks overlap most with mine, what I’d need to add for each, and which carry more exposure than mine, using the earlier test.

Conclusion

The tellers’ story has a second chapter, and the branch boom that saved their jobs didn’t last. The Bureau of Labor Statistics outlook for Tellers projects employment falling 13% from 2025 to 2035, as mobile deposits and online banking keep customers away from the window. The first wave of automation changed what tellers did, and the second is taking away where they did it.That’s why the task list is a recurring exercise. Models improve unevenly, and a task that failed the spec-and-check test last spring can pass it next spring. Redo the list every six months and move a few more hours toward the work a model can’t finish without you.

There will always need to be a human in the loop, the person who commits to what the system proposes, as my Human-in-the-Loop entry puts it. Beyond that, there has to be human agency: someone who decides rather than someone who clicks approve.

My mom won’t be able to explain my job, whatever the next title is. A model won’t need the title to find my exposed work, and it can point out where it is.

Run the first prompt next Monday, before the next reorganization or the next model release decides the mix for you and how you can explain it to your mom.

Action items

• Draft your task list. Run the first prompt and correct the table it builds from your role, or log every task longer than 30 minutes for five days by hand, using the O*NET task list for your occupation to catch what you missed.
• Score each task. Mark whether a finished version can be specified in two sentences and checked faster than it was made, and treat tasks that pass both as your exposure.
• Trade one hour. Each week, move an hour from your most exposed task to a planning decision, a stakeholder conversation, or time with users.
• Rewrite one case study. Rebuild your strongest portfolio piece around the decisions you made, including any evaluation criteria you wrote and what happened after launch, and note which adjacent job its tasks would also qualify you for.

AI doesn’t care about your job title was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

Agents at Anthropic now log six hours for every human hour, and reading their output would eat half a week. The job is to create judgement systems for what deserves attention.

On October 16, four researchers at Anthropic published How employment is changing at Anthropic, a study of how its own agents changed the work.

Anthropic is one of the companies at the front of this shift, running its own work on the models it builds, so the study is an early read on where other companies are heading.

And honestly, it looks like previous studies in new clothing as I have linked below, just accelerated. Read on.

Here are the early results: By August 2026, agents directed by employees were logging six hours for every hour their people worked.

Agent sessions spent about 80% of their time idle, waiting for a person to answer. Even for employees who don’t write code, reading everything their agents produced would take more than half the workweek.

That’s the reading tax, and every team with agents will pay it.

None of this is new — we have decades of research and many, many principles like Automation Bias that we can apply to solve these problems. Now it’s up to us to apply them.

Underneath the numbers, the nature of work is changing yet again, and this time it’s turning everyone into a manager instead of a maker of widgets. That makes this a story about human in the loop and who keeps the agency. A person who can’t read what an agent did is still in the loop on paper, while the agent makes the calls.

That puts the problem on ground designers, product managers, and researchers already work.

Here are seven steps for keeping people in charge of that loop.

Measure Reading Time

Herbert Simon saw this coming in 1971. In Designing Organizations for an Information-Rich World, he wrote that “a wealth of information creates a poverty of attention,” with most of the cost landing on whoever receives it. He meant data banks, and the line fits agents better.

Gloria Mark at UC Irvine has measured how thin it already is.

Her studies from 2016 to 2020 found people stayed on one screen for 47 seconds on average. Her first study, published in 2004, found two and a half minutes, according to the University of California’s How to sharpen your attention and meet your goals in 2024.

Every agent you add is a new sender in an inbox that was already overflowing with more information.

Anthropic measured the agent side the way a designer would, converting each session’s output into words read at 250 words a minute.

Even non-technical groups crossed half a workweek of reading.

Most days I build now, with a coding agent doing much of the work, and my limit is how much I can read and judge. So treat reading time like page load time, a budget a team sets and defends.

What’s lost in here is that a lot of work goes unreviewed, no human in the loop. Four hundred and twenty four hours is a lot more than the 168 a week currently has.

If one workflow’s weekly agent output, divided by 250, comes to more hours than anyone has, the workflow has a design problem.

Rank Agent Requests

One Anthropic engineer compared managing agents to babysitting. A child who needs a juice box opened and a child with a cut finger both arrive at full volume, and so do the agents.

Hospitals learned what that costs.

Joint Commission’s 2013 Sentinel Event Alert on medical device alarm safety estimated that 85% to 99% of alarm signals don’t require clinical intervention. Staff stop hearing them which is alarm fatigue in spades. Its database logged 80 alarm-related deaths between January 2009 and June 2012.

An agent that flags everything at the same volume has trained its human to ignore it.

Hospitals answered with alarm management: priorities, quiet defaults for routine readings, and a named person who responds. Mark Weiser and John Seely Brown at Xerox PARC described the goal in The Coming Age of Calm Technology in 1996, as technology that moves from the periphery of attention to the center and back.

Most agent interfaces skip the periphery and drop every request in the center.

Make agents justify a severity level for each request, keep the routine ones in the periphery as a digest, and bring only decisions a person must make to the center.

At Anthropic, agents mostly post in channels set aside for Claude-human work, which keeps the threads where people talk readable.

Design the Review Step

Review is where people pay the reading tax, and it’s where a human in the loop either keeps agency or hands it over. At Anthropic, agent-written code was removed or reverted at the same rates as code the same engineers wrote by hand, though the authors caution that agent errors may be harder to spot.

Perception is the bigger risk. In Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, a 2025 trial, METR found that 16 experienced developers took 19% longer with artificial intelligence tools after expecting to save 24%.

If people can’t tell how long review takes, they can’t tell when they’ve stopped doing it.

Andrej Karpathy made the same point in Andrej Karpathy: Software Is Changing (Again), his June 2025 keynote for Y Combinator. Models generate and people verify, he said, so the loop between them should run as fast as possible, with the model kept “on the leash.”

My Automation Bias entry on UX Guidelines covers what happens when that cycle slackens: people approve what the machine proposes and miss what it never flagged.

A review screen should show what changed next to what existed, what the agent checked and skipped, and where each claim came from.

Put the riskiest change first, and log review time per item so the team can see the tax.

Write Decisions Down

A manager’s decisions only count if the agents can see them. Agents reason from whatever people leave in the open. One Anthropic employee’s name stayed on a months-old document about company swag, so agents kept sending them requests for T-shirts and mugs long after the job changed hands.

Shopify’s chief executive, Tobias Lütke, gave this work its popular name in June 2025. He preferred “context engineering” to prompt engineering and called it the core skill, as Simon Willison recorded in Context engineering.

Anthropic’s engineers add, in Effective context engineering for AI agents, that a model’s context is a finite attention budget. Every loop runs on two budgets, the agent’s context and the person’s reading time.

The cheapest way to protect a person’s attention is to write the decision down once, where every agent can find it.

At Anthropic, written standing instructions for agents grew from a few hundred to 8,500 in a year. Meeting time grew 20% for the median employee and agents see none of a meeting unless someone writes it up.

For design and research teams, this is familiar work. Decision logs, research repositories, and design rationale used to be documentation few people opened.

Now agents read every page, which raises the bar for keeping it current and owned.

Manage the Agents

At Anthropic, a fifth of employees surveyed in April 2026 said their role had shifted from doing the work to managing Claudes, and interviewees described individual contributors making VP-level decisions. Peter Drucker saw the first version of this in The Effective Executive in 1967, writing that “every knowledge worker in a modern organization is an ‘executive’” if they’re responsible for a contribution.

Agents extend that to anyone with a login, and Microsoft’s The 2025 Annual Work Trend Index: The Frontier Firm is born has a name for the role: agent boss.

I call it the work about the work, and the report calls the supporting work meta-work: specifications for multi-agent runs, loops where one agent reviews another, and pipeline dashboards.

It grew from 10% of employees’ interactive sessions in fall 2025 to 22% in spring 2026. Ethan Mollick argues in Management as AI superpower that this is ordinary management. What’s scarce, he writes, is “knowing what to ask for.”

Every one of those homemade dashboards is a product nobody designed.

Christopher Noessel argued for designing this layer in Designing Agentive Technology in 2017. Anthropic’s content team shows where it leads: the writers behind the greetings on Claude.ai taught agents to write that copy, then moved into writing prompts and evaluations.

Designers and researchers already have a manager’s skills here; the job is to claim this layer before it hardens into internal scripts.

Hire for Judgment

At Anthropic, an automated review of interview scorecards found that mentions of candidates’ creative use of artificial intelligence tools roughly tripled for technical roles and quadrupled for non-technical ones between 2024 and 2026. For technical candidates, comments on autonomy and judgment rose the most, while domain expertise shrank as a share of the assessment. In effect, Anthropic is hiring managers of agents.

Kent Beck, who created extreme programming, saw the split early. In April 2023 he wrote 90% of My Skills Are Now Worth $0, with a subtitle saying the other 10% were worth 1,000 times more.

A portfolio of polished screens shows what a model can now produce, and a record of what you rejected shows judgment.

The World Economic Forum’s Future of Jobs Report 2025 found that seven in 10 employers consider analytical thinking a core skill, the top of its list. For anyone in design, research, or product, that changes what goes in a portfolio and what happens in an interview.

Show the decision trail: what the agent proposed, what you cut, and why the result held up with users.

In interviews, hand candidates a page of agent output with three planted problems, and watch which ones they catch first.

Rebuild Entry-Level Roles

The hardest number in the report concerns people never hired. Junior employees’ share of Anthropic’s technical departments slipped from 21% to 14% since early 2025, and junior technical hires fell from 27% to 9%.

Anthropic isn’t alone.

In Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Erik Brynjolfsson and two Stanford colleagues found a 13% relative decline in employment for 22- to 25-year-olds in the most exposed occupations. Not every executive agrees.

The Register’s AWS CEO says using AI to replace junior staff is ‘Dumbest thing I’ve ever heard’ covered Matt Garman’s argument from August 2025. Juniors cost the least and use the tools the most, he said, and without them nobody will have learned anything in 10 years.

The skill to supervise an agent comes from years of doing the work it now does.

Lisanne Bainbridge warned of this in Ironies of Automation in 1983: automate most of a job, and people keep its hardest part, the monitoring, without the practice that built their judgment. Entry-level roles in design and research can be rebuilt around review, with a junior making the first call on agent output and a senior explaining every override.

That’s an apprenticeship, and it costs less than hiring a senior who never had one.

Conclusion

In 1950, the United Auto Workers and General Motors signed a five-year contract that Fortune magazine called the Treaty of Detroit. Walter Reuther’s union gave management room to run the plants, and workers got cost-of-living raises, an annual wage increase, pensions, and health benefits.

The report reaches for that same settlement and names Anthropic’s version of it on three points:

• Broad stock ownership
• An early warning system for jobs at risk
• A priority on moving people into new roles inside the company

Most of us don’t set compensation policy. We do set what agents show people, when they interrupt, and how review works. Those choices decide whether everyone’s new job as a manager comes with real agency or just a sign-off button.

Right now whoever builds the dashboard is making them, and designers, researchers, and product managers should be.

The study’s framing is careful. Anthropic is one unusual company, growing fast and building the models it uses, so its practices won’t transfer cleanly to anyone else. The direction is still clear enough to act on this quarter. Output keeps getting cheaper while human attention stays fixed, and the people who decide how it gets spent will shape what these teams become.

Action items

• Measure one workflow. Count the words its agents hand a person in a week, divide by 250, and compare the hours with what anyone has.
• Tier agent requests. Give every request a severity the agent must justify, batch the routine ones into a daily digest, and keep interruptions for decisions only a person can make.
• Log decisions publicly. After any meeting that changes direction, post the decision and its owner where agents and colleagues both read.
• Pair juniors on review. Let a junior make the first call on agent output, with a senior explaining every override.

Resources

• Wilmers, Catipon, Massenkoff, and McCrory, How employment is changing at Anthropic, Anthropic, 2026
• Herbert Simon, Designing Organizations for an Information-Rich World, 1971
• University of California, How to sharpen your attention and meet your goals in 2024 (Gloria Mark’s research)
• The Joint Commission, Sentinel Event Alert 50: Medical device alarm safety in hospitals, 2013
• Mark Weiser and John Seely Brown, The Coming Age of Calm Technology, Xerox PARC, 1996
• METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 2025
• Andrej Karpathy, Andrej Karpathy: Software Is Changing (Again), Y Combinator, 2025
• Patrick Neeman, Automation Bias, UX Guidelines
• Simon Willison, Context engineering, 2025 (quoting Tobias Lütke)
• Anthropic, Effective context engineering for AI agents, 2025
• Peter Drucker, The Effective Executive, 1967 (quoted from a Fast Company excerpt)
• Microsoft, The 2025 Annual Work Trend Index: The Frontier Firm is born, 2025
• Ethan Mollick, Management as AI superpower, One Useful Thing, 2026
• Christopher Noessel, Designing Agentive Technology, Rosenfeld Media, 2017
• Kent Beck, 90% of My Skills Are Now Worth $0, 2023
• World Economic Forum, The Future of Jobs Report 2025, 2025
• Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford, 2025
• The Register, AWS CEO says using AI to replace junior staff is ‘Dumbest thing I’ve ever heard’, 2025
• Lisanne Bainbridge, Ironies of Automation, Automatica, 1983

Human attention is the new old bottleneck. We have to design for it even more. was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

AI companies have spent the past year hiring forward-deployed engineers: engineers who sit inside a customer’s team, hear the problem firsthand, and ship against it. One count of 1,000 listings found the role grew 1,165% in a single year. This year I did a version of that job for my own newsroom, without the title, and it changed how I think about listening.

Being heard runs on two clocks. Listening should be slow. Response should be fast.

Nobody feels heard by a report →

Editor picks

• Borrowed slopes →
  When newsrooms reach for the oldest kit in digital design.
  By Wira Indra Kusuma
• What I told the CFO when he asked why design matters →
  How to show the impact of design outcomes.
  By Arin Bhowmick
• Designing for an idea before it exists →
  On the useful parts that don’t quite fit.
  By Hiroshi Sato / 佐藤 仁

The UX Collective is an independent design publication that elevates unheard design voices and helps designers think more critically about their work.

On M.F. Husain and the meanings of orange →

Make me think

• Worth building →
  “Before, a project like this needed justifying. I’d have asked whether I’d use it enough to make the time worthwhile. If I expected to spend weeks on it, I’d have started thinking about other users, pricing, how it could pay for itself.”
• Design owns coherence →
  “Established functions are defined not only by their activities, but also the decision rights they own to something that the business considers important. Finance owns capital allocation. Legal owns risk. Marketing owns demand. Sales owns revenue. Engineering owns reliability. What UX/Design typically owns feels like small potatoes — consistency, usability, accessibility. None of which addresses meaningful business value at scale.”
• Grieving the loss of details →
  “Before the term “vibecoding” was coined and the industry shifted to valuing programmers primarily for designing architecture, I proudly called myself a coder as opposed to an engineer. In my eyes, this highlighted the way I focused on the details, performance optimization, knowing the intricacies of the language, and being able to explain how things work, as opposed to juggling Java-esque abstractions.”

Cool finds

Energy icons: a free, open-source library →

Prompt motion: videos made with Claude Opus 5.5 →

The convergence of design →

Careers & more

• Getting good at your job can stop you getting better →
  Many experienced people plateau out of sight.
  By Dora Czerna
• The AI experience doesn’t care about your org chart →
  Our ways of working haven’t caught up.
  By Kate Agena, PhD
• Taste is not a design methodology →
  Our favorite answer to AI-generated sameness.
  By Patrick Neeman

Support the newsletter

If you find our content helpful, here’s how you can support us:

• Check out this week’s sponsor and support their work too
• Forward this email to a friend and invite them to subscribe
• Sponsor an edition

Why design (still) matters, Taste is not a methodology, Borrowed slopes was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

Casey Harrell is a climate activist. For years, ALS had been slowly taking his speech, until even the people closest to him struggled to understand what he was saying.

In July 2023, surgeons at UC Davis implanted tiny sensor arrays in the part of his brain that controls speech. AI learned to read the signals his brain sends when he tries to talk, turned them into words, and spoke those words aloud in a voice rebuilt from old recordings of him.

The first time he heard it, he said: “It feels a lot like me.” People who hadn’t heard his voice in a long time cried.

I’ve always believed that what can destroy us may also save us. Power can be wielded in many directions. It’s never just black and white.

Right now, most people who aren’t tech bros or tech sisses are either avoiding AI altogether or working from a narrow view of what’s actually happening (and what’s possible). Many are scared. The consensus seems to be: AI is coming for us, and that’s probably gonna make everything worse.

After a global pandemic and a fresh round of nuclear anxiety, AI feels like the next unprecedented thing. The one that might finally be the nail in humanity’s coffin.

But what if it’s not?

‘Good’ AI is a choice.

What intrigues me about AI is exactly this power: the potential for harm is enormous, which means the potential for good is too. Yin and yang.

So what if it all goes right?

Humans have always been driven by what if. It’s how we’ve crossed oceans, tamed fire, and ended up, well, here.

Exactly here. Collective anxiety, crisis, tech that seems to be racing away from us. It’s no coincidence.

But that potential. The opportunity. The blank canvas. The thrill of not knowing exactly what will come out of something is as scary as it is fascinating. And it’s what we’ve always been pushing towards.

This exciting sweet spot between what’s possible now and what it might make possible is what got me into designing products for a living.

And if that work has taught me one thing, it’s that technology is never neutral. Every product is a stack of decisions: who it serves, what it rewards, what it refuses to do. A feed that keeps you scrolling at 2 am is a design choice. So is a tool that hands you back your evening.

And the good AI stories, while harder to find in an endless feed of doom and gloom, are not lucky accidents. In every one of them, someone decided who it was for, put the right people in the room, and kept humans in charge.

Good is a choice someone made. What if we make more of those?

Yes.

What if?

What if it cures cancer?

AI is already catching cancer earlier.

Here in Sweden, Lund University ran the MASAI trial. Around 106,000 women in the Swedish screening programme were split in two: half had their mammograms read the traditional way, half with AI support. The AI group saw 29% more cancers detected. The final results showed 12% fewer cancers slipping through between screenings. Fewer of the aggressive kind, too.

As a carrier of the breast cancer gene, with at least one mammogram a year since I turned 30, I’ve spent nights worrying, not about my own risk, but about my daughters’. What if AI solves breast cancer?

Zoom out from one disease to all of biology. For 50 years, working out the 3D shape of a single protein could take a scientist years. In 2020, DeepMind’s AlphaFold learned to predict those shapes from a sequence alone. Two years later, its team released predicted structures for more than 200 million proteins. It has since been used by over three million researchers in more than 190 countries. The creators have won the Nobel Prize.

And then there are the superbugs. At MIT, researchers trained an AI to scan more than 100 million chemical compounds for ones that could kill bacteria. It found halicin, which killed strains resistant to every known antibiotic in lab tests, and later abaucin. Halicin is named after HAL, the murderous computer in 2001: A Space Odyssey. A reference to its significance.

The machine we were taught to fear is finding essential medicine.

What if interfaces fit every body and every mind?

For all of computing history, people have had to adapt to the machine.

AI is starting to flip that. Now.

Be My Eyes started in Denmark, founded by a blind furniture craftsman, as an app connecting blind people with sighted volunteers over video. In 2023, it added Be My AI: point your phone at a menu, a medicine label, or your outfit, and get a detailed description back.

Casey Harrell, whose story opened this piece, is the far end of the same idea: an interface that adapts so completely it meets you inside your own body — and maybe even mind.

So while most AI applications may feel like a boost for mass production, the opposite is also true. For the first time ever, we can create interfaces that shape themselves to us. Our eyes, our voices, our attention spans, our bad days. We can finally kill the norm and build custom at scale.

What if it gives artists better access to tools?

A friend told me about how low-income filmmakers in his community can now tell their stories due to accessible generative AI for film. And while “made without AI” is more meaningful, showcasing their storytelling skills may be the way to access funding that was limited to privileged groups.

I also see this in my life. It’s now easier for me to reformat my writing to make it easier to distribute, type off my handwriting to an online doc, and automate the organization layer of my creative process to allow me more time to actually create.

Last month I built an entire app to help me keep track of my ideas as voice notes. It auto-categorizes them and surfaces them back to me based on what I’m doing and where I am. It’s been a game changer for my process and one of the most fun things I’ve done lately: the app’s design is inspired by my kids’ squishy obsession, so the visuals and haptics are satisfying to watch, and tinkering with them was as fun as working on my usual pottery.

What if AI gives back to Earth?

Rainforest Connection hangs solar-powered listening devices in the canopy. AI listens for chainsaws, trucks and gunshots, and pings rangers’ phones the moment it hears one. The rangers from the Indigenous Tembé community use it to defend their land. The network now covers more than 736,000 hectares in 37 countries. What if we can protect the Earth before it’s too late?

Floods affect around 250 million people a year. Google’s Flood Hub uses AI to forecast river floods up to seven days ahead, even where there are no river gauges, and now covers places where 700 million people live, in over 100 countries. Google says flood forecasting in Africa is now on par with Europe. I can imagine there’s even more tech that can do to protect us.

There’s more to planet Earth than nature, so…

What if it helps unite us, humans?

In 2024, DeepMind researchers built an AI mediator, named the Habermas Machine (after one of my favorite German philosophers), and tested it with more than 5,000 people debating divisive issues. People preferred its summaries to a human mediator’s, groups moved closer together, and it worked minority objections into the final conclusion so everyone felt heard. The team behind it trained Habermas to mediate, not to persuade. The job was to find where people already agreed and write it down. If you think about it, it’s the same kind of technology behind the outrage engine of our feeds, just pointed the other way.

Could something like this eventually bring us closer together? Make us comfortable discussing and disagreeing again, so we can move toward a compromise and away from extremes and isolation?

AI as a peacekeeper. A tempting thought. A machine that helps us hear each other might be a step in the right direction.

What if it doesn’t take our jobs?

For most, this is the scariest part.

And yes, jobs will disappear. The World Economic Forum’s Future of Jobs Report 2025, based on more than 1,000 employers, expects 92 million jobs to be displaced by 2030 and 170 million new ones created, a net gain of 78 million. The fastest growth is expected in frontline human work like care and education.

Forecasts are guesses, though. And even if you believe this optimistic prediction, a net gain is no comfort if the job lost is yours — or if your next one might only be available after years of retraining.

I think the truth is somewhere in the middle.

In 2024, Klarna announced that its AI assistant was doing the work of 700 customer service agents and handling two-thirds of its customer chats. About a year later, its CEO said the quality had suffered, and the company started hiring humans again.

What I’m trying to get at is this:

AI built to replace people tends to make things worse. AI built to support people tends to make things better.

MASAI didn’t fire radiologists; it gave them back 44% of their reading time. Casey Harrell didn’t need a new voice; he needed his own.

Whether AI takes our jobs or transforms them isn’t decided by the technology. It’s decided by people in meeting rooms, choosing what to automate and what to protect. Some of those people are designers. I’m one of them.

Most of those people do care.

I choose to believe that they, like me, take their responsibility personally and seriously. And that we can and will figure out a way to make AI work for good.

What if it actually frees up our time?

Last week, my friends over at Sanna, a Norwegian AI-accounting startup, announced that they were introducing a 4-day work week for their accountants.

In 2019, Finland’s then-future prime minister asked on a party panel why a four-day week or a six-hour day couldn’t be the next step, and whether eight hours was really the ultimate truth. It never became Finnish policy, but the question stuck. And we now know the answer can be yes. In the UK’s 2022 pilot, 61 companies cut hours with no cut in pay. Revenue held steady, resignations fell by 57%, and 56 of the 61 kept the shorter week.

I am hopeful that AI will finally be the technology that doesn’t turn efficiency into more work — because now the gains are so obvious, why would we not spend them on leisure, self-fulfillment, being with each other? Not doing that feels insane to me. I’m sure I’m not the only one.

I can imagine it.

A Friday that belongs to me: reading, writing, hanging with my kids, or maybe doing nothing at all. The stuff of Renaissance paintings. And maybe our future?

What if it takes us offline?

This is my most hopeful what if.

Almost every example in this piece still happens on a screen. That’s where we are. Right now. But I don’t think it’s where we’re headed. Screens add a digital layer to life that takes away from our lived experience.

I’ve always thought screens were never the point. I’m almost sure you’d agree.

We stare at our phones because that’s where the doing happens: the booking, the paying, the replying, the finding. Endless apps, each built to keep us locked in. If AI can do the doing, the reason to stare goes away. You say what you need, and get on with your day. And even further: what if you don’t have to be proactive at all; it can organize based on what it learns about you. The interface becomes secondary. Maybe, it goes away.

Think back to Be My AI. A blind user points a phone at a menu, hears what’s on it, and puts the phone back in her pocket. Think of Casey Harrell, speaking with no screen at all. The most advanced interfaces are already the ones that disappear.

Wouldn’t it be funny if AI is what finally breaks our phone addictions?

And there’s another argument.

When content becomes infinite and basically free, it stops being worth much. One-to-infinite, the logic of the feed, gets flooded with sameness. And when everything is generic, what becomes valuable again is the specific: the handmade, the local, the person who actually knows you, the things you can touch. With your heart, mind, and soul.

One-to-one, not one-to-infinite (I wrote about this before, here).

What if AI pushes our tech obsession so far that it solves it? It runs itself in the background. And we just do us.

Humans, human to human, together.

Don’t dream it, be it.

None of my musings cancel the risks. AI can (and already is) be(ing) used to surveil, manipulate, replace, and harm, and people far more qualified than me are right to sound the alarm.

But — I hope this came across in my piece — there is as much potential for good as for evil. Yin and Yang.

The future isn’t something that happens to us.

It’s something we build. And we can only build toward what we can imagine. If the only story we tell about AI is the one where it ends us, we leave the steering wheel to the people who don’t mind that ending. It’s their greatest weapon: to make us feel defeated, choiceless, numb.

So this is for the people who make things.

The designers, the writers, the engineers, the product people, the ones in the meeting rooms. We are the ones deciding, one screen, one button, one sentence at a time, what this technology rewards and what it refuses to do.

Ask who it’s for.

Put the right people in the room.

Keep humans in charge. MASAI kept a radiologist in the loop; Rainforest Connection put the tool in the hands of the Tembé rangers rather than a central authority.

Let’s build the version you’d want your mother to use at 3 am or your daughter to carry in 20 years.

And for everyone else: stay curious. Look past the headlines. Ask for the good version. Demand it. Stray from your feed. Occasionally, turn off the news.

Who knows what will happen? What can destroy us may also save us. Believing it might save us is our only chance.

So: what if it all goes right?

What if, one day, we look at the world we built with it, and say what Casey Harrell said when he heard himself again?

It feels a lot like us.

Nicole is a Content Designer turned Design Director based in Stockholm, Sweden. She potters, writes poetry, and raises little girls in a house by a meadow. You can follow her writing here or get it directly to your inbox via her publication, eggwoman. Nicole is on LinkedIn. Find her portfolio at nicoletells.com.

What if AI sets us free? was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

In the last 20 years, the tech world has transformed parts of humanity into one-dimensional digital zombies. Sadly, you can hold me personally responsible. Let me tell you why. It started in 1995, when I was 12 years old, and our household got access to the web. An internet connection was needed because my parents had an advertising agency located in an office next to our house. The World Wide Web allowed my folks to send documents to print shops without having to drive there in person.

My parents’ agency did things like branding, marketing campaigns, and product packaging designs. In the mid ’90s, my parents’ clients understood that the internet was ‘the next big thing.’ They said that ‘they had to do something with the internet.’ In other words, they wanted a website. There was no real strategy behind the need to have a website. Online presence was a goal in itself. And I could help them fulfil this goal.

As a nerdy teenager, I taught myself to write HTML and design interfaces. With my newly acquired skills, I was able to build websites for my parents’ customers. In reality, these companies would give me one of their paper brochures and tell me to ‘make this a website.’ The very first corporate websites were just elaborate business cards: a description, a photo, a list of their services, and contact details.

The interactive web

My teenage years in the ’90s, and my role as (what was then called) a webmaster, gave me much satisfaction. Over the following years, the web changed. Websites became more interactive. Webshops and social media emerged. Users could now also actively participate. They could add content and communicate with one another.

Perhaps the first interactive system I designed was a portal that enabled companies to offer health services to their staff. Through this platform, workers could get free gym memberships. Of course, on the platform, staff members could monitor their health progress. More significantly, their bosses could also scrutinise the overall fitness of their workforce. The platform enabled the company to measure the relationship between HR investments and sick leave. Where does the sweet spot lie between paying for health services and reduced costs related to staff absences? At the time, I thought I was designing a health platform, but what I was really designing was workforce optimisation.

Web Design to User Experience

Over time, my job title changed from Web Designer to User Experience Designer. I got an updated, fancier title that looks cool on a business card or LinkedIn. My role suggested I wasn’t just designing websites anymore. Somehow, I was designing experiences. I was crafting how people interact with a digital product and how they interpret the interfaces and information they are exposed to.

In other words, I was designing comprehension and behaviour. Or… was I producing something even more fundamental? Was I perhaps designing how people see themselves? Or even more elemental, was I nudging people into becoming someone different?

As the technical possibilities of the web advanced, my work became more and more concerned with what happened after a website was launched. The website itself wasn’t about the finished product. The industry began tracking what people did with it. Tools such as Google Analytics allowed companies to see how far people scrolled, how long they stayed on a page, and which actions they completed. Suddenly, the behaviour of website users could be analysed in detail.

My job shifted from making websites look and feel nice, or, as I would say at the time, cool, to optimising how people interacted with them. The importance of aesthetics was being overtaken by concepts like conversion optimisation.

The role of the metric

At the level of the tech worker, my own productivity started to be monitored. This was done through corporate frameworks like Agile and software, especially Atlassian products. I delivered story points and maintained velocity instead of just (giving my boss the illusion I was) working hard. All of a sudden, I was dealing with standups, tickets, and dashboards.

Because a website wasn’t seen as a finished product anymore, features could be added or optimised at any time, based on new insights from metrics. Time-boxed projects became ongoing trajectories without a specific end. Clients wouldn’t pay for a finished website, but would purchase an ‘envelope of time’ designers and engineers could spend on ongoing improvements. Project management turned into product management. This way, teams could quickly respond to changing requirements. Teams became agile.

Measuring the website became part of a larger move in which success was defined by performance metrics. In addition to being measured myself through story points and velocity, this calculative logic moved into the product. The metrics that guided companies became metrics inside apps. The metrics that governed workers transitioned into metrics that influenced users.

A website became something that could be continuously measured, tested and improved. This seemed totally sensible. Why wouldn’t we want to know whether our designs worked and how we could improve? These are valid questions. However, there is a downside to measuring. It defines what counts as knowledge. What can be measured becomes a particular form of authority. The things captured in Google Analytics become more important than aspects of the user that can’t be metrified.

Companies use (Agile) frameworks to measure workers. Workers build products that measure users. Users then measure themselves (and each other).

The blind spot of the metric

When I entered the tech industry, the web was still relatively young. Websites became platforms, projects turned into products, and analytics were embedded in decision-making. Measurement moved from being something that was ‘interesting’ and was used occasionally to a form of knowledge that structured every aspect of my work. Metrics became what we would see as ‘true’.

A metric tells something about reality. More significantly, it selects what part of reality becomes visible. Not only for the company and tech teams, but also for the users. Other things may still exist. Yet, they become harder to see. The number of story points you deliver shows something about your contribution, but it doesn’t take into account the support you gave your colleague or the time you spent on the phone with a client. For the user, a 10,000-step achievement might indicate a healthy person, but the metric doesn’t show what the person drank or ate that day.

Agile, product, and user metrics allow for classifying employees and products. User-targeted metrics do that too. They make it possible to categorise, rank, and compare users. An Uber rating helps the platform assess passengers and drivers. One user with a rating of 4.86 doesn’t mean much, but when millions of users are rated, it gives a lot of information. Metrics turn complex human beings into comparable profiles and enable platforms to make decisions.

Body or code

Let’s look at the user a bit more in detail to illustrate how metrics affect them. What happens when metrics become the primary way through which users understand themselves? In that case, the metric begins to compete with our real lived experience. I discovered this personally when I started using Fitbit. I had been having problems with my sleep, and I thought a health app could help me understand what on earth was going on. At first, I wanted to see on which days my sleep was worse. What was the cause of these bad nights? Maybe Fitbit could help me.

The old me would wake up and simply decide how I felt. If I had slept badly, I might take it easy in the morning. If I felt rested, I would probably do an early morning workout. My experience of my own body was the primary source of information. There was no score telling me whether I had slept well or badly.

The Fitbit-me looked at my app the moment I woke up. My night was translated into data. Sleep duration, time to fall asleep, deep sleep, REM sleep, and all that jazz are displayed on a screen. I might be told that I slept for 6 hours 39 minutes and that my sleep score was 68. Fitbit could tell me that I had less deep sleep than usual, or that my sleep was below my personal average.

When Fitbit would tell me I had a few nights of poor sleep, I would start to get nervous. I knew ‘for a fact’ I needed a good night’s sleep because I had a few bad ones. I considered Fitbit a more reliable source of information than my own perception. The metric became my reality. Obviously, going to bed feeling stressed will not lead to better sleep. I would go to bed early but lie awake for a long time.

Finding truth in a health wearable is different from being rated as an Uber driver or having work tracked in Jira. Although all three forms of knowledge rely on metrics, the metrics influence how we understand ourselves differently.

When I look at my Fitbit score, I might doubt my own bodily experience. I can wake up feeling rested, but if the app tells me that I had a bad night. I might consider the score more reliable than my own feeling. The metric changes how I interpret myself.

For a taxi driver, the Uber rating reflects how other people evaluate them. A driver may therefore adjust their behaviour to avoid a poor rating. The metric creates an incentive to behave in a certain way.

In Jira, story points and velocity make work visible within an organisation. Yet, they cannot capture everything that makes someone a valuable colleague. Helping a struggling teammate may contribute to a team’s success but is not included in story points. This way, workers may prioritise what counts as measurable output over what matters in the work itself.

The way the metric functions is thus different, depending on the context. What connects these examples is that each makes a particular aspect of life visible. Only a selected part of reality becomes meaningful. Based on this meaning, people change their behaviour. They change how they sleep, drive, or work. They understand their sleep or work through the metric.

The homo metricus

So what happens when we allow a metric to determine how we understand ourselves? Let’s go back to our work as UX designers. I have spent most of my professional life designing things that mediate how people understand and act in the world. For a long time, I thought the important question was whether my interfaces were easy to use. Whether they provided a good User Experience.

Then I started wondering whether they were doing something more fundamental. Maybe technology doesn’t just change how we do things and what we do. Maybe it also changes how we understand what/who we are.

The metrics inside our workplaces, our apps, and our wearables may seem to be nothing more than useful representations of reality. But, once metrics become part of the environments in which we act, they can also become part of how we understand ourselves.

This sounds abstract, but it isn’t. Users care about how many people liked their story, what their credit score or Airbnb rating is, or how long their Duolingo streak is. Users accept metrics imposed by the tech platform as part of their daily routine. What do you do when your follower count stagnates, you deliver fewer story points than usual, or see your Uber passenger score fall?

Eventually, we may find ourselves living according to metrics that we helped create by contributing data. I call the person produced by this environment the homo metricus.

The homo metricus is someone who understands and governs itself through algorithmically defined metrics.

Obviously, I didn’t set out to become a homo metricus. No one does. But that doesn’t mean that we aren’t a homo metricus in some way or form. We take some of the metrics that tech platforms (including Atlassian) introduce awfully seriously. (Read more about the homo metricus in this post).

The role of the designer

When I entered the tech industry, the web was a collection of pages. In fact, there was even a directory that aimed to list all the websites in the world. Over time, websites became platforms, projects became products, and analytics became embedded in decision-making. Measurement moved from something we used occasionally to a way to structure our work. The metric then travelled from the tech company to the user. Users began measuring themselves and each other.

This doesn’t mean that metrics are bad, per se. Metrics can make things visible that would otherwise be hard to see. They can help us understand our behaviour and improve what we do. However, we might occasionally forget that a metric is only a representation of a part of reality. Things not captured by the metric still exist.

Be honest with yourself... How intensely do you look at numbers? As a tech worker, how much do you care about your story points, velocity, 360-feedback score, etc. As a designer, how much do you look at churn rate, conversion numbers, and A/B tests? As a user, how much do you look at metrics provided by Uber, Instagram, Fitbit, Duolingo or any other app you use frequently?

As UX designers, one thing is to acknowledge how we are governed by (Agile) numbers. We should also be very mindful of the limited reality that metrics create. Everything that is not captured by the metric still exists. Ultimately, we, as UX designers, need to understand how the metrics we add to products change how people behave. We are not designing experiences. We are designing how people understand the world and themselves. We are designing the homo metricus.

So how can we design a UX that reduces how much the user relies on metrics without eliminating their benefits? We can’t remove the metric. It’s often a big part of the UX and helps the company to achieve its business objectives.

However, we can give the metric less authority. The Fitbit app, for instance, could show the sleep score together with the user’s own self-reported idea of how rested they feel. Jira could acknowledge the value of helping colleagues instead of treating completed story points as the only indication of contribution.

In real life, people can take off their wearable because they get tired of always seeing their steps. These options can also be added to the UX itself.

As designers, we can search for ways that give the metric more open to interpretation. This less metric-centric UX will leave room for parts of life that can’t be quantified. It may still give the metric authority, but not the only one.

This article is ultimately about how business metrics become metrics that are embedded in apps. We, as designers, should analyse where the metrics we introduce come from and what they are intended to achieve. How do they contribute to the performance metrics of our business? Is the metric the only way this objective can be achieved? Or are there alternatives?

As UX designers, we might be able to decide which metrics are dominant, and which alternatives could be available. Through our UXs and metrics, we are helping shape how people understand themselves. Let’s not make them fully developed homo metrici. What solutions or adjustments would you propose?

This article covers a very limited message of my upcoming book with the working title “Homo Metricus — How smartphone apps change what it means to be human.” In the upcoming period, I will post more insights to explore the relationship between humanity, tech platforms, and metrics.

Sources

My ideas are partially a contemporary bastardisation/reinterpretation of Michel Foucault’s later work. Other philosophical theories that cover related social problems can be found in the writings of Alain Supiot, Gilles Deleuze, and Herbert Marcuse.

UX designers are not designing interfaces, they design personalities was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.