When Perfection Became Suspicious
In the age of AI, perfection is no longer proof of care. Sometimes it is proof of nothing at all.
/* There is a moment when you land on a page so polished it could be a screensaver, or scroll through a brand feed arranged with suspiciously perfect taste, and something registers before a conscious thought forms.
Not quite distrust. Not quite discomfort. Something quieter.
A half-second of off.
The brain files it away and moves on, but that half-second is doing a lot of work. Understanding what happens there changes how you think about design, trust, and what makes anything believable in an environment where almost everything can be made to look perfect.
Said, for the record, by a pixel-perfect nitpicker who names layers in Figma and Photoshop.*/
The Valley Left the Robot Lab
In 1970, Japanese roboticist Masahiro Mori described a phenomenon he called the Uncanny Valley.
As robots became more humanlike in appearance, people generally felt more comfortable with them — until a point just before full human resemblance. There, comfort suddenly collapsed into something closer to revulsion.
Too human, but not quite.
The gap between almost-right and actually-right produced a response that was visceral, automatic, and largely beyond conscious control.
Robots, meanwhile, kept improving. The hair got better. The skin got better. The movements got better. Facial expressions became less wooden, then less obviously wrong. The line of the uncanny moved, as it always does when exposure and technology recalibrate the eye.
Still unsettling, often impressive, increasingly familiar.
Mori was describing robots. The valley did not stay there.
It now lives in landing pages. In brand Instagram feeds. In ads that feel like ads despite trying very hard not to. In founder content that follows the vulnerability arc so precisely that the vulnerability itself becomes suspect.
The trigger is no longer a face that is almost human. It is an experience that is almost real: too smooth, too coherent, too conveniently flawless. The kind of perfection that no genuinely human-made thing usually achieves without leaving fingerprints somewhere.
Users do not think: this is the Uncanny Valley.
They think: this feels fake. Or: this is trying too hard. Or they simply click away without articulating why.
The valley moved from faces to brands. From robotics to digital experience. And understanding why requires going further back than Mori — into evolutionary biology, of all places, where credibility has always had a cost.
The Birds Had a Point
In 1975, Israeli biologist Amotz Zahavi proposed something most of his colleagues initially rejected: the Handicap Principle.
Watching Arabian babblers — small desert birds — perform apparently reckless displays in front of predators, standing conspicuously in the open rather than fleeing, Zahavi concluded that the danger was the point.
The display worked because it was costly.
A bird fit enough to survive making itself a target was broadcasting something difficult to fake: genetic quality sufficient to absorb the risk.
This is Costly Signalling Theory. Its core logic is simple and radical at once: a signal is only credible when it is too expensive to fake.
Cheap signals are ignored or distrusted because anyone can produce them, regardless of quality. Only when a signal imposes a real cost — energy, risk, vulnerability, irreversibility — does it carry honest information.
The peacock’s tail is the classic case study in terrible practicality and excellent marketing. Metabolically expensive, physically cumbersome, actively dangerous. Everything about it should reduce survival probability. Instead, it functions as proof of quality precisely because anything less fit could not afford to carry it.
This logic runs through human signalling everywhere once you see it.
The degree that is credible not only because of what it teaches, but because of what completing it costs. The restaurant with an open kitchen, where the work cannot hide behind the menu. The maker who leaves construction details visible because the work can survive scrutiny. The expert whose credibility comes less from the claim than from years of visible decisions that can be checked against outcomes.
And it runs through design.
Before AI, a perfectly produced brand — flawless photography, immaculate typography, pixel-perfect consistency across every touchpoint — was expensive to produce. Expensive in time, in skilled labour, in attention, in money. The perfection was credible because it was costly.
It told you something real about the organisation behind it: they invested. They cared enough to bear the cost.
That is what made it a signal.
The Night Perfection Stopped Working
Then the cost collapsed.
Not gradually. Within a relatively short window, the production cost of perfection dropped from meaningful to almost negligible.
Flawless imagery: generated. Perfect grammar at scale: standard. Brand voice kept hyper-consistent across thousands of touchpoints: automated. Visual polish that once required weeks of skilled work: hours, or less.
This was a signal devaluation event.
The signal itself did not change. The cost of producing it did. And when a previously costly signal becomes cheap to fake, it loses its information value almost immediately.
The market recalibrates.
What used to say we invested, we care now says anyone could have made this. Which means it says almost nothing.
Perfection stopped working not because audiences suddenly became impossibly sophisticated, though they did become sharper. It stopped working because the economics of production changed so completely that the signal lost its structural basis.
Baumol’s cost disease gives the economic half of the story: when automation makes many things cheaper, human labour becomes relatively more expensive. Zahavi gives the signalling half: that expense is not just inefficiency. It is what makes a signal believable.
The Uncanny Valley response to over-polished digital experiences is the neurological expression of this collapse.
The brain runs an old assumption — highly polished things require real investment, and therefore indicate real quality — against new sensory evidence. The assumption no longer reliably holds. The result is not a neatly formed argument, but a feeling. Not consciously. Below that.
Something is off.
The discomfort is the prediction error, unresolved, filed as distrust.
Then Authenticity Got a Template
Naturally, the internet did what the internet always does when one signal stops working: it ran screaming in the opposite direction.
If perfection signals low effort, imperfection must signal human investment.
Real photos instead of stock. Slightly informal copy. Uneven testimonial lengths. Founder vulnerability. The messy-real reel instead of the produced commercial.
And it worked.
User-generated content outperformed studio perfection. Founder-led brands outperformed faceless logos. The uglier ad won. Not because ugliness performs, but because believability beats beauty when trust is the conversion variable.
Here is the problem: it worked visibly enough to become a strategy.
And strategies get systematised. And systematised signals get produced at scale. And signals produced at scale stop being costly. And signals that stop being costly stop being credible.
The vulnerability post is now a genre. It has conventions, templates, recognisable arcs.
The failure story with a redemption arc. The I almost quit narrative. The behind-the-scenes moment lit like an awards ceremony. The deliberate imperfection that is much more deliberate than imperfect.
Erving Goffman mapped the deeper mechanism through his dramaturgical theory in 1959, decades before social media made it acute. Social interaction, he argued, is never free from performance.
We always manage impressions. We always move between frontstage and backstage selves.
Social media collapsed the backstage. Authenticity-content culture then turned the backstage into another frontstage.
Which creates the second-order Uncanny Valley: authentic content that is almost authentic, but registers as performed. The vulnerability that is slightly too well-structured. The candid moment with suspiciously good light. The imperfection that arrives exactly where imperfection would be most effective.
Users cannot always articulate what they are detecting. They detect it anyway.
That half-second of off fires again, for the opposite reason. The signal has been co-opted, and the brain knows it — even when it cannot explain how.
What Still Costs Something
If perfection collapses as a signal, and performed imperfection starts collapsing behind it, the question becomes obvious: what remains credible?
Costly signalling logic gives a clear answer: the signals that hold are not only expensive to produce, but expensive to sustain falsely over time.
A single authentic moment can be fabricated. A single vulnerability post can be engineered. What is much harder to fabricate is a body of evidence that accumulates consistently and specifically over months and years.
Three things hold better than most.
Specificity of knowledge. Generic content about a problem is easy to produce at scale. Content that demonstrates genuine understanding of the specific, embarrassing, situational version of the problem — the one only someone who has actually lived it would know — is not.
AI can approximate the average description of a problem. Real expertise produces the specific one. The specificity is the proof of work.
Consequential decisions made visibly. When a brand or a person has clearly chosen something and excluded something else — a colour palette that deliberately will not appeal to everyone, copy that assumes a specific reader and ignores the rest, a product decision that costs short-term revenue for long-term coherence — the choice itself becomes a costly signal. It demonstrates that a mind was present, making real decisions with real stakes.
AI tends towards the broadly plausible. Human judgement can still afford to omit. The courage of omission is expensive.
A longitudinal record. A decade of decisions that cohere. A body of work where later choices are legible in terms of earlier ones. A trail of consequential moves that tells a consistent story without needing to announce one.
This is structurally difficult to fake because the cost is cumulative and compounding. You cannot manufacture a ten-year record in an afternoon. The record is the proof, and the proof is time.
These things do not produce a visual style or a content format. They produce a posture — one that generates credible signals as a byproduct of actually being what it claims to be.
Which is the only stable solution to an arms race in which every specific signal eventually gets turned into a template.
The Believability Tightrope
None of this means perfection is irrelevant.
The other extreme — deliberate roughness, proudly unpolished — is its own kind of miscalibration. Incompetence does not build trust either.
Good digital work now has to navigate a thin line: enough polish to signal competence, enough humanity to signal reality. The optimal zone sits just before visual perfection — close enough to indicate investment, short of the smoothness that now registers as synthetic.
Where that line sits is not fixed. It shifts with audience, context, category, and time.
In 2016, it sat somewhere different than it does now. In two years, it will sit somewhere different again — because normalisation cycles keep shortening, signals keep turning into templates, and the arms race has no terminal state.
Which means finding the line is not a design decision made once. It is a judgement exercised continuously — calibrated against a specific audience’s current assumptions, adjusted as those assumptions shift, and updated as yesterday’s proof starts to look like today’s performance.
The tool that produces the content cannot make that judgement.
The judgement has to precede the tool.
Believability is not a style. It is an ongoing read of what the room currently trusts — and the discipline to produce something that actually earns that trust, rather than merely performing the signals that used to.
The work is less about choosing a look than about keeping your balance without wobbling: enough polish to signal competence, enough humanity to feel real, enough restraint not to turn credibility into a performance — and not a movement more than the audience will forgive.
When everything can be made polished, polish stops being proof.
What becomes expensive is not beauty, but believability.
That said, I am not saying I would never buy a humanoid robot. I am only saying I would need to trust its calibration.
© Alexandra Mark
© Alexandra Mark