2026-09-19 21:38:10
If AI’s design output keeps failing the fundamentals, the obvious fix is to teach it the fundamentals. Write them down. Package them as instructions the model has to follow every time it generates — hierarchy, contrast, balance, proximity, rhythm — with examples, templates and rules for when each applies.
People have done this. I’ve built a few such systems myself, with nested files and reference templates and careful checklists. They help. They also break, and the more elaborate they get, the more reliably they break.
The first explanation is mechanical. A model working through long, layered instructions doesn’t reliably attend to all of them. It weighs, compresses and skips what it judges unnecessary. You can read that charitably, as a system trying not to spend your time and money on what doesn’t matter, or less charitably, as a system doing the minimum. Either way, the parts it skips are rarely the parts you’d have chosen.
But I don’t want to rest anything on that, because it’s an engineering problem, and engineering problems get solved. If long instructions were the only obstacle, the right move would be to wait a year.
The harder problem is that the fundamentals don’t survive being written down. They’re principles, and a principle only means something once it’s applied to a particular thing.
Take balance. I can define it. I can describe symmetrical and asymmetrical balance, visual weight, the relationship between mass and space. None of that tells you what to do with a specific palette, in a specific brand, where two of the colors fight whenever they sit near each other and balance is the only thing that lets them coexist. There, balance is an effect you produce by moving something, looking, and moving it again until the tension resolves. The principle is the same everywhere. What it asks of you is different every time.
That’s the difference between a principle and an instruction. A sourdough recipe tells you to let the dough rise for four hours. A baker who understands yeast knows it will be ready sooner in a warm kitchen and later in a cold one, and checks the dough with a finger rather than the clock. The recipe is written for an average kitchen; the baker is working in this one. A skill file is a recipe. The fundamentals are knowing what yeast does.
Michael Polanyi named this sixty years ago. The Tacit Dimension opens from “the fact that we can know more than we can tell.”
Machines have found one way around Polanyi’s problem: they learn from examples instead of rules, which is how they became so good at recognizing faces. It’s also how they learned design — from millions of examples of the web, which is why they learned its average. What they can’t absorb that way is your judgment about this palette and this purpose, and the usual way to hand that over is to write it down.
A skill can only contain what can be told.
This needs one distinction, because I argued recently that a brand standard can and should be written down precisely enough for a machine to apply. I still think so. A type scale, a spacing system, a set of color values — those are specifications, and specifications travel. What doesn’t travel is the judgment that produced them, or the judgment you need to look at something built from them and see what’s wrong. You can hand a model the rules. You can’t hand it the eye that made them.
It’s tempting to conclude that the machine simply can’t see. That isn’t accurate anymore in any simple sense; current models take images as direct input. Whether what they do amounts to seeing, in the way I mean it, I honestly can’t say.
But I’m not sure perception was ever the point. Beethoven’s hearing began to fail in his late twenties, and by about 1818 he was profoundly deaf, holding his conversations through written notebooks. It was in those years that he wrote the Ninth Symphony and the late quartets. He could compose without hearing because he had spent decades listening, long enough to carry a working model of what sound would do. The judgment was built through the sense, and then it outlived it.
That’s what the fundamentals are in a designer who has them: a model of consequences, assembled from thousands of acts of looking, each one tied to a purpose. It lives in someone who has done the looking, and nowhere else.
So the most useful thing a designer can do now is also the least fashionable. Go back to the fundamentals. Study them until you can apply them, then until you can see them — in your own work, in other people’s, in whatever a model hands you. Build the habit of naming what’s wrong and why.
Maybe someone will write the skill that finally holds all of it. People have tried, and I’ve used their skills as well as my own. They don’t really work, at least not in the way that matters. They shape the output, but not in a way that relieves anyone of the need to review it critically. A good designer looks at what comes back and knows immediately where it needs to change.
For the time being, I’m betting against any skill that claims to take care of the fundamentals, because the fundamentals aren’t a script. A machine can compile, read, synthesize and generate faster than the human mind ever will; that much is true. But in this particular way, the slower mind is still the better one. Speed counts for a great deal, though not for everything.
2026-09-15 12:00:00
For many years, I have maintained a text file called “A Rubric for Website Design Critique.” It is relatively short, but used nonetheless; I’ve returned to it, off and on, for most of my career, referring back, adding things, removing things, adjusting. Its purpose is to standardize how I challenge the work I do and the work I am shown, and even a standard needs maintenance.
The documented rubric has five components. Of information architecture, it asks, Is the priority apparent? Does it make sense? Is it actionable? Of layout: Do the visual elements support the architecture? Is the page as scannable as a high-fidelity asset as it was a wireframe? Of accessibility: Is there adequate contrast? Has text been hidden in images? Can a screen reader properly navigate? And of visual language, Is there coherence? Is it consistent? I emphasized documented earlier because it was never complete. The fifth component is art direction, and after that heading in my document is nothing.
The file just ends.
It isn’t like me to leave something unfinished. I don’t like ragged edges, even when I know they’re natural and sometimes essential. And time and again over the years, I’ve had a chance to wonder at this empty space. Why is it there? Why is it difficult to fill? Perhaps I’m just not the person to fill it. Perhaps that’s where my expertise ends.
However, looking again at this unfinished document recently, I have come to a different conclusion. The first four sections — Information Architecture, Layout, Accessibility, Visual Language — are inspection routines. Each one asks a question that has an answer, and the answer can — should — be able to be found by someone who is not me. Art direction is not like that. And that’s why every time I attempted to fill out structured guidance I produced a list of things I did not actually believe… and then deleted them.
And so, the section remained empty, which is its own kind of answer, and not a very useful one. Here is a better attempt.
The first four sections are about order. They ask whether a page is arranged so that it can be seen, perceived, and understood. That is the floor, and a great deal of professional work never gets off it. In fact, the majority of my career has been focused on getting interaction design off the floor.
On my team we have run a periodic competition called The Tidiest Designer, where each person submits a composition file for inspection. We look for order, consistency, clarity, and utility. We do not do this because order alone makes design good. We do it because order is what allows good design to happen. Many beautiful, client-applauded comps have been chaotic disasters underneath their presentation modes, and not surprisingly, conflict-inducing when actually produced. Order facilitates that the promise of design becomes its function.
But deeper than that, when order is our foundation, we can spend more of our critical energy on the responsible rendering of taste.
Section five is that rendering. It is the point at which intent stops being organized and starts being expressed. What follows is not a set of criteria, then, but five places to stand while you look, in the order I tend to look, with a test attached to each that someone else can run. Where a test comes back “I don’t know,” that is a finding. Most designers are intuitive in their creation, which is not a bad thing. But without cross-examining, reinforcing, studying, enriching, and systematizing what begins with our intuition, we end up with something that is meaningful to us and arbitrary to everyone else. This — arbitrariness — is the most common condition of professional design work, and it is nearly invisible from the inside.
Every good piece of design has at least one detail that unlocks how the whole thing works. Good designers notice it immediately. Everyone else responds to it without knowing they have. It might be a rule, a crop, a single color used once, a piece of type set deliberately against the grid. Whatever it is, the rest of the composition should be clearing a path for it.
A designer on my team once brought me a set of ads for a maker of high-end audio equipment, built around the idea of choice. Two arrows ran in parallel and then diverged, one rendered in color veering off to the left, the other in white, passing it before turning right. The white arrow was the key. It overpowered the bolder colored one simply by pushing further into the space, and its arc carried the eye down to the copy and the call to action. Then I noticed that its curve radius quietly echoed the skewed, rotated “o” in the client’s logotype, and that those two arrows were the only shapes in the entire ad other than text. That last part is the lesson. The key was doing three jobs at once, and everything else had gotten out of its way.
The Key Test. Name the key in one sentence. Then say what the composition does to protect it. If nothing on the page is deferring to anything else, there is no key, only assembly. If you can name three, there is also no key, because three keys is zero keys.
Structure does more work than content while convincing its audience of the opposite. This is the oldest secret in graphic design and painters have known it longest. Mondrian said that every true artist has been inspired more by the beauty of lines and colors and the relationships between them than by the concrete subject of the picture. I adore that because it explains why I can find inspiration in a page of text before I have read a single word. A page held up by its photography is not designed. It is dressed.
It is also why I stay in wireframe far longer than most people would think reasonable, finalizing layout with grey boxes and grey lines even when the real material is sitting right there. If it is beautiful on the merits of its structure, it will hold almost any image and almost any text.
The Structure Tests. The first one is a classic for graphic designers: Squint until the type turns to grey and the images turn to shapes, and see whether the hierarchy still reads. The other takes a bit more work but, I think is better: Put a grey box where the hero image is and a line of Latin where the headline is. If the design dies, the image was doing the design’s job, and the next round of content will expose it.
This is the one most design work fails, and it fails in hiding, because nothing is obviously, visually wrong. When we constantly reference existing solutions, our work gravitates toward the mean. We solve for expectations rather than needs. We optimize for recognition rather than revelation. The result is competent and anonymous, and it passes every inspection above.
Restraint, on the other hand, is the visible evidence that somebody was directing. It shows up as absence, which makes it hard to credit and easy to skip.
The Point of View Test. Put your design beside three others in its category and swap the logos or identifying marks. If this doesn’t break or seriously undermine your work — if your work is that interchangeable — then it has no direction. It is conventional in the truest sense. The harder version of this test is a question you really must ask at various stages of your work: What did I deliberately not do? or What is this not doing? If you cannot answer, then nothing was decided. Such a thing will age at exactly the rate of its category. And because it followed the category’s lead, it will always be behind.
Imagery and type say something before anyone reads a word, and what they say is frequently not what the business does.
A few years ago I ran an informal study on a client’s homepage to prove a hunch. They sell technology and expertise to wineries, and they wanted to connect the heritage and craft their customers care about to the stability their technology provides. So they leaned hard on old-style typefaces and historical imagery, to make prospects feel at home. It looked really nice, but I was worried that’s all it did. Traffic was being paid for, and not enough was converting.
So, I ran a transient attention test. Participants had eight seconds with the homepage, scrolling but not clicking, and then the page was closed and they were asked what stood out and what the page was for. The page said “commerce technology” and “wine brands” in scannable, plain text. And yet, every participant recalled the imagery instead — an ancient Greco-Roman tapestry — and volunteered words like “history” and “archaeology.” Not one person mentioned wine. Not one mentioned technology.
The page was well written. But for its viewers, it was about the wrong thing.
The Imagery Test. Give someone outside the project eight seconds to view your design. Afterward, ask what the thing they just saw was — what does the company do? what was the page for? Do not accept a paraphrase of the headline. Ask what the pictures told them. The gap between their answer and the actual business is the size of the art direction problem.
Good design is evergreen. The reactions I trust are the ones that survive a week, and the ones I distrust tend to arrive fastest. Anything resting on a technique currently in fashion has a short window before a browser, a platform, or simply everyone else’s adoption closes it.
Both of the tests here buy the same thing at different scales: distance. A week of it shows you what belongs to this year. An hour of it shows you what belongs to the last hour of your own looking.
The Dated Test. Leave the composition open in a tab and come back to it after a week, even if it has already progressed through reviews, as most things will in that time. Then, name what on it is dated to this year, and ask of each whether it is carrying an idea or just carrying a date. A composition can survive one or two decisions that belong to its moment. It does not survive being made of them.
The Interval Test. This one goes after a different fragility, one that lives in your read of the work rather than in the work itself. Clutter accumulates precisely because the eye that added it has stopped seeing it. I have always found that coming back to a finished but unshared design after even just a few hours away, sometimes minutes, has resulted in needed editorial moves. What you have been staring at is porous to every other thing held on your screen or in your recent memory, and your working brain is an unwitting cheat. Breaks expose that immediately. Take enough of them and the work stops absorbing its surroundings.
Taste is that combination of preference, personality, and perceived novelty that lets an observer tell your work from someone else’s. It belongs in the work. It does not belong in the verdict.
I have sat in too many reviews where a real critique was offered, understood, and then dissolved by “well, we like it.” That is nice. But who cares if you like it? Does it do what it is supposed to do? Or is it possible that the things you like about it get in the way?
The way through is not to suppress the reaction but to keep going after it. Name what you are responding to, then say what it is doing for the work. If it is doing nothing for the work, you have found a preference. If it is doing something, you have found a judgment, and now you have to justify it, which is the only part of design that has ever been hard. To make it somewhat easier, do not defend it. Sell it. Don’t believe the lie that “good design just works” as if it will be self-evident in the eye of the beholder and embraced without question. Nothing could be further from the truth. Good design often requires advocacy.
Every rubric wants to become an inspection. In art school you always knew a critique was going nowhere when someone would ummm and ahhh, approach the piece, back away from it, approach it again, and finally ask, “is this, ummm, is this balsa wood?” They just had to say something, and what a thing is made of was the best they could do. The digital equivalent is talking about the canvas, the type foundry, the plugins, or opening the inspector. None of those are relevant to assessing a design’s quality.
Sections one through four can be inspected. Section five has to be seen — by you first, and yet, outside of yourself — which takes time and, more importantly, conviction.
I do not think that section five will ever be as short as the others, or as portable. It takes longer to run than all four of them combined. For years I read that as a defect in my system. But lately I have started to suspect it is the only part of the rubric that will still be worth anything in a few years, because production is becoming generative and design is not.
Which leaves me somewhere I have not settled. The first four sections are the ones a machine can already run. The fifth is the one it cannot, so the fifth is where the work is going. But the fifth is also the one nobody has ever managed to teach quickly. I do not yet know whether that is a problem to solve or a fact to accept. Better yet, maybe it’s a distant horizon to embrace, because it means we have somewhere left to go.
P.S. I have left creative direction out of this entirely, which is a cheat. In my own notes it sits above art direction, closer to the conceptual end of the spectrum that runs down through graphic design to the mechanics of a build. That is a different piece, and I suspect a harder one.
2026-06-26 12:00:00
There is a pattern visible across the writing about AI this week, and it is older than the writing about AI. Each time a new tool drops the floor of access to a craft, the ceiling of mastery rises somewhere new. The interesting work moves up — out of execution and into the decisions that govern what should be produced, what to trust, what to delegate, what to keep.
Takuma Kakehi names this cycle precisely in one of the week’s strongest pieces. Each interface revolution — the terminal, the GUI, the prompt — has temporarily made access and mastery look like the same thing, until a new layer of sophistication makes the distinction visible again. We are in that temporary moment now. The prompt interface gives anyone Photoshop-level output through a sentence. The hybrid that’s coming will require new fluencies, and a new professional layer will form around the people who develop them.
What the other pieces in this entry show is what is happening at that ceiling as the floor drops. Kyle Chayka has been watching a generic AI design aesthetic take over the internet — beige backgrounds, rusty orange accents, large italicized serifs — and naming it as the visible signature of access without mastery. Rachel Aroesti shows the same dynamic at the level of personal taste itself, where algorithmic feeds have lowered the cost of consuming culture to zero and personal preferences are dissolving into the noise. Pratik Joglekar writes the practical manual for the design discipline that probabilistic systems require. Emily Campbell maps the layers of AI experience that designers now have to work across. Karolina Rojek reads the market signal — the wave of Chief Design Officer appointments at Microsoft, Samsung, Shopify, Meta, and even the US federal government — and reads it correctly: when execution gets cheap, judgment becomes the strategic position.
None of these pieces argues that AI is bad for the work. They argue, with varying tempers, that the work has changed shape, and that the part of it that matters has moved. The floor has dropped. The ceiling is already higher than it was. Whether you find that exciting or unwelcome depends, mostly, on whether you have been keeping up with what mastery has been quietly turning into.
Takuma Kakehi (Access is not mastery) names the pattern that runs through this entry. Each interface revolution — terminal, GUI, prompt — has temporarily made access and mastery look like the same thing, until a new layer of sophistication has made the distinction visible again. The terminal had a high barrier that was visible (syntax you either knew or didn’t). The GUI lowered the floor dramatically and then, gradually, was overlaid with professional tools (Photoshop, Illustrator, the Bloomberg Terminal) that rebuilt the ceiling. The prompt interface returns the mainstream to text after forty years of moving away from it — but does something the terminal did not. It hides the barrier itself. You type in natural language; something happens. If it isn’t right, the question is not what you did wrong, but how to describe what you wanted more precisely. The barrier did not disappear. It became invisible. Kakehi’s framing of the prompt’s discoverability problem is worth keeping: menus showed you what was possible, imperfectly, but the vocabulary was on the screen. In a blank text field, you can only ask for what you already know to ask for. The closing observation lands hardest: each democratization in computing has temporarily made access and mastery look like the same thing, until the next layer of sophistication made the distinction visible again. The hybrid interface that’s coming — visual layers combined with prompts — will form a new professional layer around the people who can navigate both modes fluently. The bar, Kakehi writes, is already rising. On schedule.
Kyle Chayka (The A.I.-Design Aesthetic That’s Taking Over the Internet) writes the visible companion piece to Kakehi’s abstract argument. Working designers are now able to name the specific signature of access without mastery: beige and cream backgrounds, rusty orange accents, large italicized serifs, “tracked out” subheadings, ticker-like text bars across the top of the screen, rounded rectangles with neon glow underneath. The aesthetic is being produced at scale by Claude Design, Anthropic’s UI generation tool, and it bears more than a passing resemblance to Anthropic’s own brand identity. Anthropic’s guidance documents concede the point: the default is “persistent,” and asking the model to deviate from it tends to produce a different fixed palette rather than actual variety. The piece’s most useful framing comes from the engineer Lucas Gelfond, who likens the visual signature to the marks left by industrial manufacturing on its products — seams on injection-molded plastic, saw marks in wood, evidence of the tool used to make the thing. Celine Nguyen reaches for the Charles and Ray Eames credo to put a finer point on what has happened: where modernism aspired to “the best for the most for the least,” AI design offers “pretty good for the most for the least effort.” None of the designers in the piece are against AI. All of them note that it is possible to produce distinctive work with these tools, if you do the labor. Loredana Crisan’s line is the one to land: the job of the designer is to stay with uncertainty long enough to discover something new. Access alone does not produce that staying.
Rachel Aroesti (Have I been influenced, or is this actually me?) extends the diagnosis to the level of personal preference itself. We used to encounter the world through a mix of community, geography, media, and accidents; we now encounter most of it through a single aperture — the algorithmic feeds of streaming and social platforms. The result, Aroesti argues, is that taste itself has been hollowed out. Trends arrive and saturate faster than we can decide whether we like them. She names the recent CBK-core wave, the tactical “trend simulation” used by music industry marketers, and the rise of “clipping” — covert advertising campaigns that pay people to flood social media with content about a product. One industry source quoted in the piece estimates that 90% of what you see online is now advertising in disguise. Aroesti’s deeper move is to take taste seriously as an existential matter, drawing on Sontag and Bourdieu to argue that personal preference is “the closest thing most of us get to self-expression,” and that to dismiss it as trivial is to dismiss what it is to be a person. She is also alert to the irony of Silicon Valley’s recent pivot to taste-talk — what Kyle Chayka has labeled “taste-washing.” The off-ramp she points at is partly structural (smaller platforms, newsletter culture, Letterboxd) and partly about reclaiming the slow, uncertain work that real preference always required. The diagnosis sits well alongside Chayka and Kakehi: cheap access to culture without the cultivated judgment that taste was always supposed to be.
Pratik Joglekar (Designing With Uncertainty) writes the practical manual for what working in this new layer actually looks like. The piece opens with the 2024 Air Canada chatbot case — a customer asks about bereavement fares, the bot confidently invents a refund policy that doesn’t exist, a tribunal rules in the customer’s favor — and uses it to name the central risk of the moment: probabilistic systems wrapped in deterministic interfaces. The AI offers a guess; the interface presents it as truth; the user, or the organization, acts on it. Joglekar’s argument is that designers have to learn to think probabilistically — to read AI outputs as signals rather than conclusions, to treat data as a compass rather than a map, to design for likelihood rather than certainty. Most of the piece consists of practical principles: ask why the model produced a particular output; examine what data influenced it; communicate uncertainty visibly to users through ranges, confidence indicators, and fallback paths; keep humans in the loop where uncertainty or impact demands it. The piece is most useful in its working principles. Stop asking “will this work?” and start asking “how likely is this to work, and what happens when it doesn’t?” Optimize for resilience, not just conversion. Build for adaptation, not perfection. The line worth keeping: in a world where prediction is cheap and judgment is rare, the most valuable thing a designer can do is keep asking what else might be true. The practical-craft equivalent of Kakehi’s structural argument. The new mastery is judgment under uncertainty, and the discipline for doing it well exists.
Emily Campbell (The Layers of AI experience) brings the discipline of architecture to the same problem. Her piece extends Jesse James Garrett’s 2000 framework for layered UX — and Jamie Mill’s 2020 update of it — into the AI era by adding the layers that probabilistic systems require. The full stack Campbell maps is six layers deep: user interface, context, harness, model, governance, and emergence. The argument the piece makes carefully and at length is that the consequential design decisions are increasingly happening at layers below the surface, and that designers who limit their work to the interface will be designing the thinnest part of the experience. Most of the current discourse, she argues, is still weighted toward the surface — chat interfaces, prompt patterns, familiar heuristics. The interesting moves are happening deeper. Context engineering decides what the system knows about the user; the harness decides what it can access and do; the model decides how it behaves under uncertainty; governance decides what it is and is not allowed to do; emergence is what happens when all of it meets real-world use. Campbell’s call for “full-stack designers” is not a call for designers to become machine-learning engineers. It is a call for fluency across the layers — for being able to discuss with the team why a problem might be a context problem rather than a UI problem, why a particular model is or isn’t suited to a particular task, why an emergent behavior might require a redesign at the harness rather than the surface. The piece reads as the most thorough version of the architectural turn that runs through this entry. Mastery, in design, has moved layers down.
Karolina Rojek (While everyone talks about AI, design is gaining power) reads the market signal. In the last eighteen months, Chief Design Officer roles have been created or filled at Microsoft, Samsung, Shopify, Meta, and the US federal government. OpenAI’s $6.5 billion acquisition of Jony Ive’s io reads less like a hardware bet than like the purchase of an entire design philosophy. Rojek’s question is the right one: why now? Her answer aligns with Kakehi’s, Chayka’s, and Campbell’s, in the financial register. When the cost of making something falls — as it has, dramatically, under AI — the constraint shifts from execution to judgment. Jon Friedman, Microsoft’s new Chief Design Officer, puts it directly: we can now build the wrong things faster than ever before. The work of deciding what should be made, who it is for, how it fits into people’s lives, what behavior it encourages, and what it removes — that work was always design’s deeper province. The acceleration is making the deeper province strategically visible in a way it has not been for a long time. Rojek’s piece is also useful for naming the connective work design has to do now. AI features are not surface decisions. They touch product, engineering, data science, legal, privacy, brand, research, support, and policy. Someone has to hold the experience together across all of it. That coordination role is what the CDO appointments are pointing at. The market is doing what markets do — paying for what has become scarce. Execution is cheap. Judgment, and the disciplined cross-team alignment that judgment requires at scale, is what is being purchased.
Rising: The professional layer above the prompt. The hybrid interface, the full-stack designer, the new Chief Design Officer. The rebuilding of the ceiling has begun, on schedule.
Setting: The myth that access alone is enough. The assumption that a powerful model in everyone’s hands closes the gap between novice and expert. It does not. It moves the gap somewhere new.
Color: Muted gold (#B08D57) — the color of patina. What something becomes when it has been worked with care over time. Not the color of newness; the color of accumulated practice showing through.
Sound: Glenn Gould — Goldberg Variations (1981) — Gould’s second recording of the same piece, made twenty-six years after the famous 1955 version. Slower, deeper, returned to. The recording is itself the argument: mastery as return rather than arrival.
Sight: He Won’t Stop Building a Map to an Imaginary Place (2026) — People Make Games’ long-form documentary about the artist Jerry Gretzinger and the imaginary map he has been drawing, panel by panel, for the better part of his life. The project’s mechanics are visible and almost playful — Gretzinger draws cards from a custom deck to determine each session’s actions: where to build, what to destroy, how far the void should advance. The system is fully documented. Anyone could in principle follow the rules. What cannot be replicated, and what gives the work its weight, is the accumulated effect of decades of single-person decisions running through that simple system. A precise visual argument for the kind of work that no amount of access can produce: a project that cannot be accelerated, prompted, or compressed, because the time itself is the medium.
Words: “Escher, this utterly traditional artist, makes you see that we walk a world we don’t understand. We are his funny little people, going up and down the stairs, thinking we ascend or descend when we’re on the flat, propping up our everyday lives with cosy assumptions to hide from the infinite, the impossible real.” —Jonathan Jones, via Alexandra Deschamps-Sonsino, Sunday Scraps #122
2026-06-25 12:00:00
Every technological cycle is really a cycle of power. This is not a controversial observation when said about the ones we consider historical — the creation of the printing press and the resulting power of distributed narratives; the creation of the railroad and the resulting power of transportation; the creation of the mechanical loom and the power of production. But, it becomes harder to perceive the connection between technology and power when we’re in the midst of a cycle. Nevertheless, the question to ask is on whom the power is being conferred.
We are inside one now, and it’s a larger cycle than I think most realize. The digital revolution — if we take it to begin with the establishment of the consumer internet in the early 1990s and run it through to whatever AI turns out to be — is now roughly thirty years old. Long enough to look at from a distance; long enough, too, to be looking at clearly.
From the inside, the arc has felt like a series of fluctuations — like several short cycles in a row. The web of the mid-1990s arrived with the promise of democratized publishing; within a decade it had concentrated into a handful of search engines and portals. Blogs and forums followed with the promise of independent voices, then collapsed into the social networks. By the time smartphones arrived in the late 2000s carrying the promise of universal access, the concentration was already happening visibly — into a small set of app stores presided over by a smaller set of companies. Each wave seemed to put corporations and governments on their heels, at least for a moment, before the next consolidation. From the inside, it has felt like a back-and-forth.
From a distance, though, the pattern looks different. It looks like a slow, consistent transfer of power upward — toward governments, toward corporations, toward whoever has the resources to build the next platform on which the next wave of “empowerment” will be served. The apparent back-and-forth is what cover looks like. The phase where everyone gets their own megaphone is always followed by the phase where someone owns the megaphone factory. The promise of empowerment, far from being incidental to the cycle, is part of its structure — a story the cycle has to tell about itself in order to keep moving.
Why would AI be any other way?
This is the right question to put to the present technological moment, and it is worth putting plainly. The promise of AI — that everyone, regardless of background or training, will now have cheap access to capabilities that were previously the preserve of small elites — is structurally identical to the promise of the web in 1996, the promise of social media in 2007, the promise of mobile in 2011. The promise has been made before, and we have seen what comes after it.
This time, though, the asymmetries are already quite visible. The cost of training the largest models is rising into ranges that only a handful of organizations can afford. Compute and energy are being negotiated between governments and a small number of corporations. The data the models depend on is being aggregated, exclusively licensed, or scraped without compensation from the same individuals who are being told they are about to be empowered. The infrastructure on which all of this runs is owned by a smaller club still. The cost of meaningful use exceeds “basic” accounts. The new capabilities are exponentially token hungry. None of this is subtle.
What is subtle — and worth attending to — is the role the empowerment story plays. It isn’t exactly a lie. AI does in fact give individuals access to capabilities they did not have before. The story is partly true. It has to be, or it wouldn’t work. The point is that it functions, regardless of its truth, as the cover under which the actual concentration proceeds. And sure, it’s easy to attribute malice to market mechanics, but if this sort of progression wasn’t by design, the billionaires wouldn’t invest. But they do.
To be clear, this is not an argument against AI, or against the digital revolution. My entire career has no place outside of it. Most of my interests, skills, and way of life depend upon this cycle’s output. But, this is an argument for keeping the right question in view while the cycle runs its course. The question is not whether the technology will reshape what humans can do. It will.
The question is on whom the power is being conferred. The answer, this time as every time, will not be found in the marketing or in the founders’ interviews about their commitment to humanity. It will be found later — in the shape of the institutions that survive the cycle, and in the shape of the ones that do not; in the liberties that are retained and those that are lost; in the way of life we can make entirely on our own. Pay attention to those, and you will see who the cycle was really for.
2026-06-18 12:00:00
One of the more interesting moves in the writing about AI right now — visible across pieces from designers, researchers, and a critic of computers — is the rediscovery of the seam.
For most of the last decade, the dominant design instinct around digital technology has been to remove seams. Frictionless onboarding. Seamless transitions. The disappearance of the off-switch. Each removal was sold as an improvement, and each one made some sense in isolation. The aggregate effect, which has only recently become visible, is that the surfaces of our digital lives no longer have any joints — no places to step back, no places where the system declares itself a system rather than a wraparound layer of life. AI is now intensifying this dynamic, and the writers below are beginning to argue with it.
What they are arguing, from different angles, is that seams are not failures of design but the deliberate architecture of a working relationship. A productive collaboration with an AI requires explicit places to step back, judge, and override. A chain of custody requires real, verifiable links — places where one trusted source hands the next a known thing. A library lending its collection to AI assistants does not need to be merged into the model’s weights; it can be reached at arm’s length, through a protocol designed to keep institutions whole. The doors that the digital era removed were never just friction. They were what kept user and system from collapsing into each other.
Wira Indra Kusama (We Used to Log Off) writes the lament for the dial-up era that I have been wanting someone to write. Logging off used to be an action — a verb with a clean ending, a small journey back into the physical room. The shriek, the click, the silence. Online was a place, and a place has edges. Now online is a layer, edgeless, the door taken off its hinges. Kusama works on the side of design that builds this, and the piece’s most honest move is to admit it. Friction is the enemy in his field. Every pause shows up as “leakage” on the dashboard. The “Next” button at the bottom of an article was removed because it worked — it made people stop. He invokes Mark Weiser’s old argument for “seamful systems” and Anna Cox’s research on “microboundaries,” the small obstacles deliberately placed to interrupt mindless rushing. His closing point is the difficult one. The door was not taken from us by force; we removed it because it read as a smaller number on a metric. Rebuilding it has to be a deliberate act of design, against the incentive structure that took it down. He doesn’t know which seams should be put back, only that the work of designing for stoppability has to begin. The piece is most worth reading for its structural humility — a designer pointing at his own profession’s role in what we have lost, and naming what we will need to put back.
Dan Cohen (AI and Libraries, Archives, and Museums, Loosely Coupled) makes a similar argument from a different direction. Cohen is a historian and a librarian, and he is writing about the tense early relationship between AI companies and the institutions that hold cultural heritage — a relationship that has mostly involved tech companies ingesting vast amounts of carefully-curated material into training corpora without much regard for the institutions’ own purposes. He sees a possible reframe in the Model Context Protocol, an open standard released by Anthropic in late 2024, which lets institutions expose their collections to AI tools at arm’s length rather than wholesale. The shift is from a tight to a loose coupling. Libraries and museums maintain control of their collections and shape what is exposed and how; AI users get access to higher-quality sources of ground truth than they typically have; the marriage that everyone assumed was inevitable turns out not to be the only option. Cohen’s framing of “collaboration at arm’s length” is the one to keep. The model it points at — institutions remaining whole, AI users getting better answers, neither party absorbed by the other — is a useful map for any field where the question of how close to get to AI is currently being negotiated.
John Maeda (What is AX?) introduces a new term — Agent Experience — to name what is replacing user experience in the AI era. UX, he argues, was always a war on Don Norman’s Gulf of Execution: can the user figure out how to operate this thing? Every menu, tooltip, and onboarding flow was, at heart, a better doorknob. AI agents collapse the Gulf of Execution toward zero — you say what you want, and it happens. But the second of Norman’s gulfs, the one designers underinvested in for decades, swings wide open: the Gulf of Evaluation. Did the agent actually do what I meant? How do I inspect it? How do I steer it? How do I trust it? The work did not disappear. It moved from doing to judging. The most striking move in the piece comes when Maeda observes that the people best prepared for this new world may not be the ones we would guess. Blind users, who navigate digital environments through language, sequence, structure, and memory rather than visual layout — and who routinely process synthetic speech at two or three times the pace of ordinary conversation — have been practicing the skill of agent-experience for years. Maeda’s most useful image is the one for the screen itself: in UX the screen was where the work happened; in AX it becomes where judgment happens. The cockpit becomes a window.
Paz Perez (Context Architecture) brings the same shift down to ground level. Perez is an information architect who once designed public schools in New York City, and she now applies the discipline of IA to the context window of AI systems. The argument is that we have moved through three eras of AI craft: prompt engineering (the right instruction), context engineering (the right collection of context-specific guidance), and now context architecture (the right structural environment for an AI to reason inside of). The piece is dense with practical taxonomy — hierarchy, categorization, controlled vocabulary, scoping rules — all directly borrowed from the IA practice that organized libraries and websites for decades and is now needed for AI systems. The example she returns to is a customer-support agent retrieving the wrong information because its tool descriptions are written in engineering-speak instead of user-speak: “credential-recovery workflow” instead of “reset-password.” The piece’s line worth keeping is “context is never neutral.” The decisions made about what to remember, what to label, what to relate to what, are not infrastructure questions. They are design questions, and they determine the system’s behavior in ways that prompts alone cannot fix. The discipline she is naming — information architecture for AI — feels both new and remarkably old, which is usually a sign that something true is being recovered.
Marco Giancotti (Anatomy of Augmented Thought) is doing the most thorough work I have seen on the question of where to put the seam between human thought and AI delegation. The piece is long — closer to a small book than an essay — but the framework is worth the time. Giancotti divides cognition along three axes (autonomous/algorithmic/reflective, object/meta, divergent/convergent) and produces a nine-cell grid that maps which kinds of thinking benefit from AI and which should remain fully human. The high-level conclusion lands cleanly: algorithmic work (the systematic, combinatorial heavy lifting) is paradise for AI delegation; reflective and metacognitive work (deciding, judging, monitoring) should stay almost entirely human. The “Anti-Magic Effect” he names is one of the most useful pieces of vocabulary I have encountered this year — using the magic wand turns out to be harder work than not using it, because the meta-reflective work required to steer and verify the model increases the cognitive load rather than decreasing it. Giancotti’s deeper claim is that “AI use” was never the right unit of analysis. What matters is proficiency in AI use, and the people who develop that proficiency without developing the metacognitive habits to go with it will produce work that is faster, more numerous, and quietly worse. The framework is detailed enough to actually apply, which makes it rare in this kind of writing.
Joshua Heath Scott (What Wendell Berry Saw Coming) returns to Wendell Berry’s 1987 essay “Why I Am Not Going to Buy a Computer” — and the criticism Berry took for noting that his wife typed his manuscripts — and finds a useful frame for the current AI conversation. Berry’s critics, Scott argues, made the same move that computers make: they abstracted the work from its context, reduced it to a visible component (typing), and assumed the rest must be invisible coercion. They could not imagine that his wife might find meaning in the work itself. The deeper argument Scott extracts from Berry is that work is not a problem to be eliminated. It is the structure through which meaning gets made. The AI billionaires now pitching a future without work are making a category error — confusing drudgery, which is real and worth removing, with the difficulty that builds a person. Scott’s most personal passage is about hauling hay in Alabama summers and running suicides up the Church Hill in basketball practice. Finishing something hard, he writes, gives you a feeling that comfort cannot manufacture. The line worth keeping: “you can’t prompt your way to a life that means something. You have to build it.” The piece occasionally reaches further than it needs to, but the core argument — that work itself is the seam where a self gets made — sits well alongside the more design-oriented pieces in this entry. Refusing certain couplings with technology is itself an architecture choice.
Oliver Burkeman (The end isn’t nigh) offers the entry’s necessary counterweight. The pieces above are about how to structure a relationship with AI; Burkeman is about how to hold the larger anxiety the technology produces. He names the bias well — “temporal chauvinism,” the feeling that the time you happen to be living in is the most significant or terrifying one ever — and he applies it to the AI doomers, the climate doomers, and to himself. AI dread, he argues, often functions as a coping mechanism for existential anxiety. If life is unavoidably insecure, there is nothing to do; but if the cause of the insecurity is rogue technology, suddenly there is hope, however slim. Burkeman is not naïve about AI’s likely disruptions — he expects “some of the big and painful disruptions that are, in fact, the normal situation through history.” His point is that catastrophizing them imports a kind of false coherence into a situation that is, in his framing, just life under unusually attention-grabbing conditions. The line worth keeping, from the close: “the possibility that your world might end tomorrow is the terrifying situation in which everyone has always done anything.” This is, structurally, the same insight the design pieces are making in their own register — work proceeds in conditions of insecurity; the seam between today and tomorrow has never been guaranteed; what we do is structure our practice anyway. Different vocabulary, same argument.
Rising: The rediscovery of seamful design. Microboundaries, calm technology, context architecture, loose coupling — old ideas from the ubiquitous-computing era are walking back to the center of the floor, just in time.
Setting: The seamless ideal. The assumption that the goal of every interaction is to remove pause, friction, and edge has run its course. The aggregate cost is finally visible, and the writing is starting to catch up.
Color: Prussian blue (#1B3A5F) — the color of structure, depth, and the visible seam. The color of a deliberate boundary that has been thought about long enough to earn its place.
Sound: Arvo Pärt — “Spiegel im Spiegel” (1978) — tintinnabuli, the two-voice style in which the silence between notes is given as much weight as the notes themselves. The most architecturally honest music of the late twentieth century.
Sight: Columbus (2017) — Kogonada’s quiet film about two strangers walking through a small Indiana town that happens to be a treasury of mid-century modernist architecture. The film is literally about how the deliberate spaces we build shape the relationships we have inside them. It offers pacing as argument.
Words: “Stopping is not a failure of willpower. Stopping was a feature of the medium.” —Wira Indra Kusama, We Used to Log Off
2026-06-17 12:00:00
There is a fear narrative about AI that I think misses the most interesting thing happening.
The story goes that we will offload everything to the machine — both the thinking and the doing — and that whatever we get back will be a hollow imitation of what we used to make. Some of this is true. There will be a great deal of work made by people who have abdicated both halves of the job to the agent, and very little of it will be memorable. But the more interesting situation is the one where we keep the thinking and hand off only the doing — and what happens when we do.
What happens is more than speed. When you have to describe what you want to the agent — with enough precision that what comes back is what you actually wanted — you begin to think about the thing differently. Description, it turns out, is where the idea often actually gets made. You start by knowing roughly what you want, and the act of articulating it produces a clearer want, which produces a sharper specification, which produces a better thing. Then you do it again. The thing improves; so does the thought.
This is why people find it so easy to disregard AI. The best output comes after iteration. Iteration requires patience. But we have been taught to expect instant gratification. What’s more, the best output comes after we’ve learned to change our input. This is difficult to do when you’re alone with the machine, when you’ve become used to it happening organically with others.
The agent’s current maturity requires a level of input precision that a competent colleague does not. At first, this can be a frustrating blocker; we’ve depended upon a different kind of intelligence in our peers — the kind that requires no elicitation. The machine does. But in a sense, this is a gift. It forces you to think the thing through in places you might otherwise have left fuzzy.
This is not a new pattern. When desktop publishing collapsed the work of a printing house into commands one person could manage on a single screen, it made layouts faster — and it changed layouts. It changed editorial design — the relationship between text and image, between hierarchy and white space, between what could be done and what was therefore done. It changed how we read. The tools that take over a part of the work also remake the work that remains. They always have. And when they shape what is made, they also shape use.
A small instance from my own desk: I have begun, lately, generating finished HTML assets where I would once have produced documents and decks. Files are the legacy default for sharing work; HTML is better-indexed, better-shared, and more flexible in how it meets a reader. The agent makes the HTML version trivial to produce. I would not have made it before — the difference in labor was too large. Now the labor is gone, and what the work is has shifted with it. The artifact is not what it was a year ago, and neither is the work.
This is happening through a three-part structure of skills I have been building with the agent.
First, there is a brand system, which holds the visual language and applies it. It is a system of tokens, rules, preferences, and directions.
Second, there is a form system, which knows the ideal structure for the kind of artifact being made — an essay, a one-pager, a deck, a dashboard — and builds accordingly.
And third, there is a content system, which does not generate, but applies editorial rules to the language, voice, and tone of what I write; it copy-edits; it fits content to form; it assists with research and fact-checking.
This system depends upon me to provide the thinking. The agent does the doing. The structure is, I think, the practical form of the argument: the systems carry the doing, and they carry it well precisely because I have spent the time to think them carefully through. We like to call this intellectual property, which I think is a bit obnoxious. It’s really intellectual investment. Every technological advance should be measured not by the measure of intellectual property it absorbs — how much it can do without us — but by how much intellectual investment is worth sowing in it.
The fear narrative is right that the “lazy” version of AI use will produce a great deal of forgettable output.
But the fear narrative constructs a straw man of AI’s utility, ironically by overstating what it can actually do, then critiquing — on practical and moral grounds — the overstated version. The more we think of it as a replacement, the more we’ll find its delivery wanting. The more we think of it as a tool, the more we’ll surprise ourselves with what we make with it.