Why AI Is Making Design Converge on the Same Look

By Anamol Rajbhandari

In 1997, a small piece of software called Auto-Tune arrived in recording studios, built to do something narrow and technical, which was to nudge a slightly flat or sharp vocal back onto the note it had been reaching for. For an individual singer, the effect was close to magical, as someone who could mostly carry a tune but never quite land it could suddenly come out sounding clean and professional. Someone who could barely sing at all could be smoothed into something passable, so that almost overnight the range of what any single voice could pull off grew a great deal wider, the hardest and least forgiving part of singing having been quietly handed off to a machine.

As the tool spread through the industry, a large share of pop vocals began to converge on one another, all of them buffed toward the same glassy, centered, faultlessly on-pitch ideal. The quality that had made Auto-Tune so useful in the first place, its pull toward a single correct standard, was never going to confine itself to one voice. It worked on every voice at once, drawing each singer closer to the same target until the targets themselves grew difficult to tell apart, so that the lift and the flattening turned out to be one motion seen from two ends. Whatever raised each voice to the standard was also what drained the difference out of the whole.

Design is standing in that line, and both halves of the pattern are already visible in it. If we were to ask a designer who has spent a season working with the new generative AI tools whether they can now make more than they could before, the answer without hesitation would be a “yes”. They can do a great deal more, faster and across a range their own training never gave them. And if we ask the same designer whether the work coming out of the field as a whole is growing more varied, there is usually a pause followed by a slightly uneasy admission that everything seems to be starting to look like everything else. 

Both of those answers are honest accounts of what the designer is actually seeing, and the reason they sound like a contradiction when set beside each other is that they are describing a single event from opposite ends of it. The freedom each designer feels is what flattening the field and learning to hold those two facts in the same hand at once may be the least examined large change now underway in creative work.

The same thing has already run its course once through human faces. A decade of filters and photo-editing apps produced what one widely read essay named Instagram Face, a look that was poreless, high-cheekboned, faintly feline, and eerily consistent from one young influencer to the next. It was the machinery of the singers all over again, in that everyone was handed the power to sculpt their own image toward a standard that used to require money and surgery, and as everyone reached for that standard, the faces slid quietly toward one another. 

What a person feels in that situation is only their own face improving, while the thing no one can feel from inside a mirror is an entire generation converging on one, similar face. The gap between the private gain and the collective loss makes it difficult to notice while it is happening. These two domains have now travelled the full arc of it, which is reason enough to contemplate why the same thing happens to every craft that lets a machine of this kind stand between the maker and the thing being made.

Why the machine returns the average

The pull toward a single standard is built into the way any of these tools works, and it is there long before anyone reaches the question of taste. A generative model takes in an enormous body of existing work and forms from it a statistical picture of what such work tends to be, and then, asked to produce something, it samples from that picture, and the likeliest sample it can draw is always the crowded middle of the distribution, the average of everything it once absorbed. Producing the most probable next thing is the entire mechanism, and the most probable thing is, by definition, the average one.

It helps to picture a thousand people each asked to sketch a chair, and then every one of those sketches laid over the last until only the composite remains. The shape that shows through belongs to no one; it is a kind of ghost of a chair, the most chair-like chair it is possible to draw. It has every private quirk of every contributor cancelled quietly against every other. We run that same overlay across nearly everything a field has ever produced, and the result is roughly what the model is reaching into each time it answers a request. So it has all the feel of a considered choice drawn from the exact center of everything, averaged.

For most people most of the time, this may feel like a genuine gift and it would be dishonest to pretend otherwise. The average of all the design work ever poured into a model is a good deal better than what almost anyone could produce alone. A beginner is handed a level of competence that used to take a few years to earn. An overstretched professional is spared the tedious middle stretch of the work, and the reach of any single person expands dramatically the moment they sit down with the tool. The difficulty is only that the center everyone is being lifted toward happens to be a single location, so that raising a novice raises them toward it and quickening an expert nudges the expert toward it too. 

Thus, ten thousand people carrying ten thousand different sensibilities, all reaching into the same model on the same afternoon, are all guided gently to the same crowded middle. Each of them experiences only the lift, and none of them can perceive the crowd gathering around them. It is because the convergence of an entire field is not something that can be seen from inside one's own chair.

Instead of feeling like a limitation, it rather feels, if anything, like flight. Which is precisely why it meets so little resistance. Since sitting down with one of these tools a person finds that everything opens outward leads to a request producing one result and a second request producing a different one. The whole sensation being that of a door swinging wide onto an enormous and well-lit room, with no way of seeing that every other door opens onto the same room or that the tool is steadily guiding everyone toward the same bright patch at its center. It shows most plainly in what appears unbidden, in the way that prompting five different tools for something as ordinary as a dashboard tends to return similar, if not the same, small handful of layouts, the ones sitting in the densest part of the training data.

And because that middle is in fact good enough, the trap closes without ever announcing itself as a trap. Since leaving the average takes real effort, the effort of discarding the first answer and then the fifth, of wrestling with the prompt, of insisting on a stubborn opinion and throwing out whatever comes back merely competent, and on a deadline, with a perfectly decent option already glowing on the screen at no cost- the decent option tends to win. It is easy to argue that this seems lazy. But it is simply the arithmetic of an ordinary working day, and when that same reasonable choice is repeated across an entire profession, the field converges without a single person having ever decided that it should. From the vantage of any one desk this looks unmistakably like freedom, and it is only when the work is set side by side, across teams and companies and whole countries, that the convergence becomes visible as something closer to weather, a condition everyone turns out to be living inside without having chosen it.

As none of it began with AI, giving the technology all the credit rather flatters it, because design had been converging for a long time before any model could draw. Products imitate whatever already works, since copying the proven thing lowers risk and the lowering of risk is most of what a company exists to do. The result was plain long before generative tools appeared: dashboards resembling cousins across wholly unrelated firms, onboarding flows that reduce to the same five screens everywhere, settings pages so alike they could be swapped between products without anyone noticing. Best practice has always been a shared center under a more flattering name. What the machine changes is the speed and the completeness of the movement toward it. Because the older kind of copying still required a human being to notice a pattern, admire it, and choose to follow it—that requirement left lag in the system, and gaps, and room for a stubborn team to wander off somewhere strange, whereas the new version arrives inside the tool itself, switched on by default and working on everyone at once.

The averaging reaches the mind

There is good reason to think the averaging reaches well past the surface of the screen and works its way back into the person using the tool. Cognitive scientists have a name for the underlying mechanism, cognitive offloading, which describes the way a mental capacity tends to weaken once the job it used to do has been handed to a device. Pocket calculators dulled the ability to do arithmetic in the head and navigation apps have quietly eroded the sense most people once had of how the streets of a city connect to one another. 

Generative AI tools carry this further than any device before them because the faculty now being handed over is the act of making itself. It is the very material from which judgment is built, so that the tool shapes the output and then, in the same motion, reaches back to absorb the effort that would have shaped it, leaving the part of a person that once did that work to soften from disuse.

From there, the loop closes on the person almost completely. What makes anyone's work recognizably their own is a private accumulation of odd particulars. The things they happen to have seen and the taste they still carry over years of getting things wrong- that accumulation is drawn upon only when someone makes a thing themselves. But it is passed over entirely when they accept what the machine offers instead. When one leans on the average for long enough, the person stops generating deviations.

The heaviest version of this cost settles on whoever is only now arriving in the field. The designers who came of age before these tools built their taste by the slow and graceless route, producing some thousands of mediocre things and working out, gradually and often painfully, why each of them was mediocre. People did not enjoy that stretch and it was also where a sensibility was forged, because the quirks that eventually harden into a personal style are found only in the friction of doing the work badly and repairing it by hand. 

A newcomer today can skip that friction almost entirely, since the tool delivers competent work on the first attempt and accepting it is the intelligent, exhausted, wholly reasonable thing to do.  Those designers begin on a floor no earlier generation ever stood on, while the ceiling they learn to aim toward is the tool's ceiling, and the tool's ceiling is the average. Newer designers may therefore never develop the particular muscle that can feel the average as average and choose deliberately to leave it. Having never once been pinned beneath it long enough to want out, and since a field renews its supply of deviation only through people who learned somewhere how to deviate, a generation taught instead to accept the middle will thin that supply.

What raises all of this above a complaint about taste is what happens to the output once it exists. The work these models produce does not simply disappear after it has been used. It pours outward, onto websites and into applications and across every LinkedIn post and social feed, and that flood is precisely the material on which the next generation of models will be trained. Today's average, thus, becomes part of tomorrow's raw ingredients, and tomorrow's model learns a picture of the world already pulled toward the center by whatever its predecessors made. When a model is fed a sufficient quantity of its own kind of output, a now well-documented thing happens to it. The distribution it has learned grows narrower, the rarest material at the edges is the first to vanish, and the whole system drifts harder toward its own mean with each successive generation, its quality and its variety draining away together.

Culture, left to itself, tends to replenish from those edges. When someone makes an odd thing, the odd thing catches on, and it becomes in time a new region of the map that everyone else can build upon. A system trained largely on its own averages has no such source of replenishment and ends by consuming its own tail. And design is where the effect will show itself first, because design operates at enormous scale and turns over quickly. Subsequently, its edges can be watched thinning very nearly in real time with a machine now sitting in the loop and hastening the collapse by feeding steadily on its own reflection.

Where the correction has to come from

The comforting thought is that this will pass the way every earlier panic about a flattening technology has passed. Culture has survived every tool that was once supposed to level it. The printing press was meant to standardize thought and ended up multiplying it beyond recognition. Similarly, photography was going to be the death of painting and instead handed painting its freedom. And synthesizers and drum machines were each written off in their turn, and each went on to widen music the moment a restless few worked out how to push it somewhere the inventors had never pictured.

If we look closely at how any of those rescues actually happened, it was by the people who recovered and never the tool in and of itself. The variety came back because someone brought judgment the instrument did not have, a taste that ran ahead of what the machine could manage on its own. That is the uncomfortable part, because judgment is precisely what a tool reaches for from the first blank screen quietly wears down. And so, the correction, if it comes, will not arrive by itself but from the way people decide to use the thing, and from a stubborn refusal to give away the one faculty that has always done the redeeming.

That faculty is earned the slow way, over the long apprenticeship that turns a fluent hand into an expert one by sharpening against resistance, when a hard limit blocks the easy move and forces an unfamiliar one, which is why constraint has always been a quiet engine of anything genuinely new. 

A tool built to erase both the effort and the limits leaves neither one intact. There is a well-documented price for handing a skill to a machine, which is that the skill fades in whoever stops using it the way it has faded in the pilots and physicians whose instruments came to do their thinking for them.

So the useful response is to do by hand the one important part that actually matters, which is the deciding. That means using the machine to reach a competent draft in seconds and then treating that draft as the start of the work rather than the end of it. It is turning it down often enough to stay able to tell the good from the merely acceptable, inventing the constraints the tool is so eager to dissolve, and guarding the slow and unrewarded practice that is the only thing keeping a person's taste ahead of the average it serves back. None of this is efficient, and almost everything about an ordinary working day argues against it.

That, oddly, is where the hope sits. As the average becomes free and instant for everyone at once, the only thing left with any real worth is the judgment to want more than the average can give. That judgment cannot be downloaded or prompted into being, only earned. 

Whether enough people will still have trouble earning it, now that the passable answer is always waiting and always free, is the question these AI tools have raised and cannot begin to answer themselves. A whole generation is quietly answering it already, mostly without noticing there was a question at all.