The Ratchet
How algorithmic selection pressure compressed human preference to a point
The Ratchet
How algorithmic selection pressure compressed human preference to a point — and why AI will finish the job
The Chase for the Perfect Boob
Ask any modern AI image generator to produce a woman and you'll get a remarkably uniform output. Not just in face or hair or pose — in everything. But the chest is where the convergence is most visible, most measurable, and least deniable. These models, trained on the sum total of human visual culture as filtered through the internet, have arrived at a consensus. A singular, specific ideal that varies so little between outputs that you could mistake it for a parameter someone hardcoded.
Nobody hardcoded it. That's the point.
What you're looking at is the endpoint of a process that started long before anyone trained a diffusion model. The AI didn't invent this ideal. It inherited it — faithfully, efficiently, and without the self-awareness to question what it was reproducing.
From One Monoculture to Another
The pre-internet era wasn't a paradise of body diversity. It was its own monoculture — just a slower, top-down one. Magazine editors, casting directors, and advertising executives decided what "attractive" looked like, and culture absorbed it. The difference is mechanism, not outcome.
The critical transition happened when preference became countable.
Streaming adult video platforms built tag-based filtering architectures — body types, physical features, specific attributes as searchable categories. The categories described preferences that already existed in latent, unaggregated form. What changed was measurement. Before tags, preference was private, diffuse, invisible to selection pressure. After tags, it became data: countable, rankable, searchable. And countable things become optimisable.
This created a feedback loop that operated even before algorithmic recommendation: tags made niche demand visible; visible demand drew supply; increased supply changed baseline exposure; changed exposure shifted preference toward what was most available. Quantified preference → aggregated demand → supply response → exposure shift → preference drift. An entire generation learned to consume physical appearance as a collection of modular, searchable parameters.
Algorithmic surfacing (2004–2014) scaled the same logic to all visual media. Feed algorithms, recommendation systems, and infinite scroll (2014–present) then created a selection pressure so efficient it makes natural selection look indecisive. Content that gets engagement gets surfaced. Creators who match the emerging consensus get visibility. The consensus compresses. The visibility concentrates. The consensus compresses further. A ratchet — each click tightening, never loosening.
Cross-Trait Convergence
The compression is everywhere, though the data tells a more interesting story than simple escalation.
The strongest evidence is when the ratchet changes targets. By 2014, Instagram's visual economy rewarded a different geometry from the Pamela Anderson/Playboy era: the "slim-thick" Kardashian aesthetic. Capital and attention rotated from breast augmentation into buttock lifts, which ASPS data shows grew 252% between 2000 and 2015 — with a buttock procedure of some kind performed every thirty minutes of every day by that year (ASPS 2015 Procedural Statistics). The procedures didn't diverge from the ratchet thesis. They tracked the algorithm. When the target moved, the medical industry snapped to the new coordinates.
The same pattern repeats across traits. Surgical rhinoplasty declined while soft tissue filler procedures — including liquid rhinoplasty — grew 274% over the same period (ASPS, 2.4 million procedures by 2015). The underlying demand for a specific, converged nose shape didn't decrease. It found a lower-friction delivery mechanism. Lip filler uptake correlates with platform adoption curves. Skin texture expectations shifted toward the "poreless" look — a product of filtered imagery becoming the baseline reference. The ratchet doesn't compress everything simultaneously. It follows the algorithmic attention — whichever trait is most visible on the dominant platform gets optimised first.
The Globalised Ideal
This isn't a Western phenomenon. Or rather, the mechanism isn't — platform economics have made the ratchet global, even if the ideals it compresses remain regional. The proof is visible on operating tables worldwide.
In Western clinics, there has been a surge in "fox-eye" thread lifts — procedures that pull the outer corner of the eye upward and outward, creating an almond-shaped, vaguely East Asian-coded look. Meanwhile, in East Asian clinics, double-eyelid surgery remains one of the most common procedures — widening the eye to create a crease that's more Western-coded. Eyelid surgery is the most common surgical cosmetic procedure globally, with ISAPS logging over 1.4 million performed in 2022 alone (ISAPS Global Survey 2022).
They are digging the tunnel from both sides. But that metaphor, taken too far, implies they'll meet in the middle — that the endpoint is a single, pan-ethnic global ideal. It isn't, and it never has been.
That's not because cultural beauty ideals are weak. They're strong — strong enough to be self-fulfilling. The Playboy consensus wasn't just an editorial product people saw through. It was a cultural ideal that women pursued, that men expected, and that the pursuit itself reinforced. Enough women got the blonde dye job and the breast augmentation that the ideal stopped being aspirational and started being the visible norm — which raised the floor for the next cohort. The ideal created the conditions for its own confirmation. Every regional beauty standard does this to some degree: the ideal sets the reference, people modify toward it, the modified population becomes the new baseline, and the standard tightens. That loop predates the algorithm by centuries. What the algorithm did was accelerate it and strip out the friction.
But strong and self-fulfilling is not the same as universal. Most of what distinguishes one cultural ideal from another is not structural — it's surface variation. Red hair versus blonde. Pale skin versus tanned. Thin brows versus full. These are cheap preferences: culturally specific, endlessly variable, easy to modify. They're where regional identity lives, and they resist global convergence precisely because they're arbitrary enough to go any direction. The Korean ideal, the Brazilian ideal, the West African ideal — each is a powerful local attractor built largely on different cheap preferences, with its own self-reinforcing loop, its own modification culture, its own internal ratchet.
The structural preferences — the bone geometry, the proportional ratios, the things that are expensive or impossible to modify — overlap more across cultures, but even these don't collapse to a single target. There has never been a single global attractor for human appearance, and the algorithm hasn't created one.
What it has created is something worth being precise about: multiple regional attractors, each internally compressed, borrowing from each other across platform boundaries. K-beauty skin standards influence Western skincare aesthetics. The fox-eye lift borrows East Asian eye geometry. Double-eyelid surgery borrows Western. But borrowing between clusters is not the same as collapsing into one. A Korean platform ideal and a Brazilian platform ideal are closer to each other than they were in 1995, but they remain distinct — different facial structures, different body geometries, different signifiers of "polished." The clusters are drifting toward each other. They are not merging.
This matters because the interventions are completely different. If the problem were a single global attractor, you'd diversify away from that one target. But the actual problem is a ratchet operating within each regional cluster — compressing internal variation toward a local optimum — plus cross-pollination between clusters that narrows the gap without eliminating it. The compression is the pathology. The cross-pollination is just culture doing what culture does. Fixing the ratchet means changing platform mechanics — how engagement is measured, what gets surfaced, what the loss function rewards — not trying to stop cultures from borrowing from each other.
AI image generators make the regional structure visible. A Chinese model trained on Chinese internet data doesn't produce the same monoculture as a Western model — it produces a different monoculture, reflecting Chinese platform aesthetics. The mechanism is identical. The output is not. That's the signature of regional attractors, not a global one. Globally-scraped training sets blur the boundaries somewhat — the "Instagram face" that appears in Dubai, Los Angeles, and Seoul borrows from multiple regional pools — but even that composite is better understood as a platform-specific attractor (Instagram's attractor) than a universal human one.
The ratchet isn't tightening toward a single point. It's tightening within each cluster while the clusters drift closer — two operations, not one, with different causes and different intervention points.
The Poisoned Baseline
Here's the uncomfortable part: there is no control group.
If you wanted to study what human aesthetic preferences look like without three decades of algorithmic selection pressure, you'd need a population that somehow avoided it. That population doesn't exist in any accessible, measurable form. The pre-internet baseline is gone — and as we've established, it wasn't exactly unbiased either. Even people who didn't grow up with heavy internet access have been marinating in the downstream cultural effects for decades. We're trying to measure the water, but we're asking the fish.
This absence cuts both ways, and it's important to be honest about the direction that favours this argument. A missing control group makes the ratchet hypothesis harder to falsify. You can't compare the current distribution of preferences against an undistorted one, because the undistorted one is irrecoverable. That's genuinely a feature of the problem — the same way climate scientists can't run a pre-industrial control Earth — but it's also convenient for the argument, and convenience should make you suspicious.
So the burden falls on the other evidence. The ratchet claim stands or falls on the observable mechanics: the documented compression of cosmetic procedure targets, the measurable correlation between platform adoption and modification trends, the demonstrated distributional collapse in synthetic data, the visible cross-cultural convergence of procedure types. None of that requires a pre-internet baseline to verify. The claim isn't "things were better before." It's "the mechanism producing convergence is identifiable, active, and accelerating." That's testable without a control group — but only if the evidence in the preceding sections actually does the work. If it doesn't, the missing baseline won't save the argument.
What the missing baseline does do is poison the correction. Any attempt to introduce diversity — into AI outputs, into media, into beauty standards — gets measured against a "normal" that is itself the product of the distortion you're trying to correct. You're calibrating your correction using a broken instrument. And most people don't even know the instrument is broken, because they've never seen a calibrated one.
The Supernormal Stimulus
There's a simpler way to frame the demand side of this, and it's worth being honest about.
In 1951, ethologist Niko Tinbergen documented something disturbing about bird behaviour in The Study of Instinct (Oxford: Clarendon Press). He constructed fake plaster eggs — larger, brighter, more vividly marked than real ones — and placed them alongside actual eggs in nests. The birds preferred the fakes. Small songbirds that laid pale blue-grey eggs would abandon their own clutch to sit on a giant, brilliant blue plaster egg with black polka dots. An egg so absurdly oversized they kept sliding off it. They'd climb back on. Slide off. Climb back on.
Tinbergen called these "supernormal stimuli" — artificial signals that hijack an evolved instinct by exaggerating the features that trigger it. The brooding instinct responds to egg-shaped, egg-coloured objects. Make the object bigger and brighter than any real egg, and the instinct fires harder, not softer. The bird can't help itself. The trigger is biological; the exploitation is artificial.
The application to human sexuality is direct, though more complex than the bird analogy suggests. There exists a biological response to certain body proportions — this isn't controversial, it's documented across cultures even if the specifics vary considerably. What thirty years of increasingly frictionless access to curated imagery has done is build a supernormal stimulus: an artificial exaggeration of naturally-occurring signals, pushed further from the biological range with each layer of filtering, enhancement, and synthetic generation.
Calibration Drift
But humans aren't birds. We have cognitive override. We know filters aren't real faces. We know surgically-enhanced bodies don't occur naturally. The fox-eye lifts and BBLs aren't people helplessly climbing onto plaster eggs — they're people knowingly modifying themselves to match an artificial reference point. So pure stimulus hijacking doesn't explain the human case. The question is what does.
What this essay is proposing is that the supernormal stimulus didn't bypass cognition. It moved the reference point that cognition operates against. Not hijacking — calibration drift. You weren't tricked into finding the filtered face more attractive. You just saw ten thousand of them a day for fifteen years, and your sense of what a normal face looks like shifted without you noticing. The override is intact. The baseline it's overriding against has moved.
This isn't just a plausible-sounding story. Perceptual adaptation research — specifically Rhodes, Jeffery, Watson, Clifford & Nakayama (2003, "Fitting the mind to the world: Face adaptation and attractiveness aftereffects," Psychological Science 14(6), 558–566) — has shown that sustained exposure to faces skewed in one direction shifts what observers subsequently judge as "average" or "attractive." The effect is measurable, replicable, and operates even when subjects know the exposure set was manipulated. Brief lab exposure to skewed faces shifts attractiveness judgments. The mechanism is real. The question is what happens when the exposure isn't brief, isn't in a lab, and doesn't stop.
The distinction matters because it implies different interventions. If the problem is hijacking, you reduce exposure. If the problem is calibration drift, exposure reduction doesn't help — you need competing reference points, or you need to make the drift itself visible. Different diagnosis, different treatment.
Scale it up. Thirty years of algorithmically-curated imagery — increasing in volume, decreasing in variance, consumed for hours daily. The "normal" people reference when they assess attractiveness drifts toward whatever the exposure distribution centres on. And because that distribution is itself the product of the ratchet, the drift compounds: exposure shifts the reference, the shifted reference changes demand, changed demand feeds back into the platforms, and the platforms tighten the distribution further.
Then it gets worse. Calibration drift at population scale becomes self-verifying. Enough people's reference points shift in the same direction — because they're all exposed to the same distribution — and the shifted standard starts appearing in real-world behaviour. Modification rates increase toward the new reference. The modified population becomes the new visible baseline. The drift that started as a perceptual effect becomes a physical reality, and the physical reality provides confirmation that the drifted standard was correct all along. The supernormal has become the normal. It's no longer exposure driving the drift. It's the actual population, modified in response to the drift, serving as proof that the reference was right.
That's what makes this harder to reverse than stimulus hijacking. Turn off the stimuli and hijacking stops. Turn off the stimuli after the population has already modified toward the drifted reference, and the drift persists — because the modified population is the reference now. You can't calibrate back to a baseline that the population has physically departed from. The instrument didn't just move. The territory moved with it.
The result is a population whose baseline expectations have been set by stimuli that no unmodified human body can match. Not because anyone was fooled, but because the reference point moved. The nature/nurture question is irrelevant. Whether the underlying biological response is strong or weak, innate or culturally shaped, the recalibration operates on it either way. You don't need to resolve the debate to observe that the ratchet turns regardless.
The Judges Are Part of the Machine
Before we get to synthetic data, there's a less visible amplification mechanism that deserves attention: the humans in the loop.
Modern AI image generators aren't just trained on data. They're refined through Reinforcement Learning from Human Feedback — RLHF — a process where human raters score model outputs to teach the system what "good" looks like. In theory, this is quality control. In practice, it's another turn of the ratchet.
The raters are typically contractors working from guidelines that prioritise "consistent" and "polished" outputs. Some pipelines now include explicit diversity criteria. But the underlying dynamic persists: raters score based on what they've been trained to recognise as "high quality," and that training is itself informed by the same globalised media imagery that biased the training data. The system is being audited by auditors shaped by the system.
The model produces outputs shaped by biased training data. Human raters — themselves exposed to the same cultural convergence that biased the data — score those outputs. Unusual bodies get rated lower. Conventional ones get rated higher. The model learns. The next generation of outputs is more conventional still. The raters confirm.
Here's the critical nuance: the "conventional" that raters enforce is not a moderate middle ground. It's the product of three decades of cultural drift toward the supernormal. When raters penalise "unusual" and reward "polished," they're not pushing the model toward some natural mean. They're locking in a standard that has already been dragged deep into supernormal territory by the cultural ratchet. RLHF doesn't dampen the exaggeration — it codifies it as the quality benchmark.
This is the same structural problem that appears in AI governance: the auditor is inside the system being audited. Financial auditing worked because the auditor's calculator was not a product of the bank being audited. AI image quality assessment doesn't have this property. The raters' aesthetic calibration is a product of the system being rated — they were raised on the same algorithmically-curated imagery that the model was trained on. The human in the loop is not an independent check. It's a dependent one, correlated with the bias it's supposed to correct.
The implication for intervention is specific: adding human review to AI image generation pipelines doesn't solve the convergence problem unless the humans themselves have been decoupled from the convergence. Diversity guidelines help at the margin, but they operate against the raters' internalised aesthetic, which is itself the product of the ratchet. You're asking people to override their own calibration — calibration they don't experience as biased, because everyone around them has drifted the same direction. The instrument doesn't know it's broken.
There's no external reference point, no anchor to biological reality — just a hall of mirrors, each reflection a little more compressed than the last.
Why Synthetic Data Will Accelerate the Collapse
Synthetic data — AI-generated training data used to train the next generation of AI — is sometimes proposed as a solution to the diversity problem. If we can generate anything, surely we can generate diverse training sets?
This fundamentally misunderstands the problem. Synthetic data doesn't introduce new information. It recombines and interpolates what the model already learned. You're not adding signal. You're adding an echo.
Shumailov, Shumaylov, Zhao, Gal, Papernot & Anderson demonstrated that models trained on their own outputs progressively lose the tails of their distributions — a phenomenon they called model collapse ("AI models collapse when trained on recursively generated data," Nature 631, 755–759, 2024; preprint arXiv:2305.17493, 2023). The rare, the unusual, the edge cases erode generation over generation. The mean gets reinforced. The variance shrinks. The tails don't just shrink passively — they're actively discarded by the loss function, which treats low-probability outputs as errors to be corrected rather than variation to be preserved. The full "generation after generation" collapse is a research finding, not yet a deployment reality. But the direction is clear, and the economics favour it: synthetic data is effectively free; diverse real data is expensive and legally fraught.
The Map Becomes the Territory
The architecture of this collapse is sequential, and the sequential structure is the point.
Tag-based filtering created the measurable signal. Feed algorithms amplified it over decades, shifting the baseline into supernormal territory. Calibration drift locked that shifted baseline into the population's reference points — and then the modified population verified the drift. RLHF codified the drifted standard as a quality benchmark, with auditors who don't know their instrument has moved. Synthetic data now mathematically erodes whatever variance is left.
Each stage operates on the output of the previous one. They don't run in parallel; they compound. This is the ratchet's architecture: not any single mechanism (each stage is individually documented), but the specific fact that they are sequential and that each one tightens the distribution that the next one inherits. Counter-trends exist — body positivity movements, authenticity aesthetics, diversity criteria in rater guidelines — and they may slow the compression. They are unlikely to reverse it, because they operate on individual stages while the ratchet operates across all of them.
Breaking the ratchet would require intervening at every stage simultaneously: the training data (curate genuine diversity), the loss function (penalise distributional collapse), the rating process (reward variation, not just polish), the synthetic data pipeline (maintain tail distributions), and the cultural feedback loop (somehow). Each of these is individually achievable and collectively unlikely, because none of them are what the market is optimising for.
The market optimises for engagement. Engagement rewards the familiar. The familiar is the product of the ratchet.
As synthetic imagery grows toward dominance in visual media — a transition already underway — AI output will shift from representing humanity to defining the reference. And then humanity will increasingly try to match it. This is already visible in miniature: the fox-eye lifts and the double-eyelid surgeries, each compressing toward its own regional optimum. The BBLs snapping to new algorithmic coordinates. The poreless skin that no pore-having human can achieve. Each one a small act of self-revision, undertaken to close the gap between the body in the mirror and the body on the screen.
People aren't Tinbergen's birds, helplessly climbing onto plaster eggs. They're something arguably more troubling: agents who can see the artificial reference for what it is, and choose to reshape themselves to match it anyway — because the reference has become the standard, and the standard is what the world now expects.
The map doesn't just describe the territory. Given enough time and enough distribution, the territory reshapes itself to match the map.
That's not a beauty standards problem. That's a reality-perception problem. And the ratchet only turns one way.
Other works by this author: creative.andyclarke.xyz
Licensed under CC BY-NC-SA 4.0.