The Imperfection Economy: Why Flawless Is Now Worthless

The Imperfection Economy: Why Flawless Is Now Worthless

In November 2024, Coca-Cola released a holiday ad. Trucks glowing through snow, small towns, the whole apparatus of manufactured warmth the company has been assembling since 1931. It was technically clean. The lighting was immaculate. And it was generated, largely, by AI — a fact the company disclosed, and a fact the internet found on its own within roughly the time it takes to pour a glass.

The backlash was not about quality. That is the part most post-mortems missed. Nobody argued the ad looked bad. They argued it looked like nothing — that a company sitting on nearly a century of accumulated human craft had chosen to spend precisely none of it. Toys “R” Us had absorbed a similar reception months earlier with its Sora-generated brand film. Same shape of reaction. Same diagnosis: the work was fine, and fine was the problem.

Something inverted that year, quietly, in the way genuinely structural things tend to. For most of commercial history, polish was expensive and therefore meaningful. A perfectly retouched image, a flawlessly kerned wordmark, a film that never wobbled — each was a receipt. It said: someone spent something here. Money, hours, talent, judgment. The polish wasn’t the value. The polish was evidence of the value, and we had all agreed to read it that way for so long that we forgot it was a proxy at all.

Then the proxy broke. Perfection became free, instant, and infinite. And the moment a signal becomes free to produce, it stops signalling anything.

This is the imperfection economy: a market in which the flaw does the work the finish used to do.

The economics of a collapsed proxy

Costly signalling theory has a tidy premise — a signal is only credible if it is expensive to fake. The peacock’s tail is honest because a sick peacock cannot grow one. The Harvard degree signals less about what you learned than about what you survived to get it.

Every craft premium in the commercial world ran on this logic. The reason a beautifully art-directed campaign commanded a fee was not aesthetics in the abstract; it was that the aesthetics were load-bearing proof of institutional seriousness. You could not fake a great photograph. You could only commission one.

Generative systems didn’t devalue perfection by making it worse. They devalued it by making it cheap to fake — which, in signalling terms, is the same as destroying it. The tail is now available to every peacock, sick or otherwise, for eleven cents. So the tail says nothing.

Markets do not tolerate a vacuum where a trust signal used to be. They reallocate. And the reallocation is already visible if you know where to look: audiences have started reading for the things generative systems cannot cheaply produce. Specificity. Consequence. Risk. The evidence of a decision that could have gone the other way.

A visible brushstroke. A vocal crack left in the master. A founder’s letter with an actual argument in it and an actual position to be wrong about. A photograph with a real place in it, on a real day, with weather that wasn’t prompted. These are not nostalgia. They are the new costly signals — expensive precisely because a machine optimising for the mean cannot risk them, and increasingly because the human producing them is choosing to leave something in that every instinct of the last twenty years said to sand out.

Flawless is now worthless not because it is ugly, but because it is uninformative.

What this does to enterprise output

Here is the uncomfortable part for anyone running a function rather than a brand.

Most large organisations have spent the last decade building machinery whose entire purpose is the elimination of variance. Brand systems. Tone-of-voice guidelines. Template libraries. Approval chains that exist to ensure nothing leaves the building that anyone could object to. This machinery was rational. When production was expensive and error was public, variance was the enemy.

Generative tooling arrived and did not disrupt that machinery. It completed it. It gave every organisation the ability to produce infinite variance-free output at near-zero marginal cost — the exact thing the machinery was built to want.

And so the strategic error of this cycle is not that enterprises will use AI badly. It’s that they will use it exactly as designed, and arrive at a destination where every piece of output they produce is indistinguishable from every competitor’s, because all of them are drawing from the same distribution and all of them have optimised for the same absence of objection.

You can already feel this in the market. Look at the last twenty B2B homepages you visited. The same gradient. The same three-column trust section. The same sentence about empowering teams to do their best work. This convergence predates AI — it was already underway when everyone was copying the same Stripe-adjacent template — but generative systems have taken it from a trend to a terminal condition. When the cost of producing the consensus artefact falls to zero, the consensus artefact is all anyone produces.

The implication for leadership is blunt: your approval process is now a competitive liability. Every gate that exists to remove the thing someone might object to is a gate that removes the only material a costly signal can be made from. An organisation that cannot ship something a reasonable person could disagree with has lost the capacity to say anything at all.

That is not an argument for sloppiness. Sloppiness is variance without judgment, and audiences read it accurately as what it is — not-caring, which is exactly the signal that got Coca-Cola in trouble. The distinction is precise and worth holding: the imperfection that carries a premium is imperfection that survived a decision. Someone saw it, understood what removing it would cost, and left it.

Why audiences got so good at this

There’s a folk assumption that detection is a technical problem — that people spot AI output by catching the artefacts. Six fingers, warped text, a hand that dissolves into the background.

That was true for about eighteen months and it isn’t anymore. The artefacts are going away. What isn’t going away is a much older and far more reliable human instrument: the ability to detect the absence of a point of view.

We have always had this. It’s the same faculty that reads a condolence card as generic before consciously parsing a word of it, that hears a politician’s non-answer as a non-answer, that knows within four seconds of a first date whether the other person is present. It operates below argument. It is pattern-matching against an enormous accumulated corpus of what it feels like when someone means something.

Generative systems are, structurally, machines for producing the centre of a distribution. That is not a flaw in their engineering; it is the engineering. And the centre of a distribution is precisely where nobody lives. Every actual human position is somewhere off-centre — that’s what makes it a position. So the output reads, to that ancient instrument, as a person with no location. Fluent, plausible, competent, and from nowhere.

This is why disclosure debates keep missing. The question is never really “was this made by AI.” The question the audience is actually asking, and has always asked, is “did anyone here decide anything.” Those correlate right now, which is why the labels feel useful. They will decorrelate, and when they do, the label will stop mattering and the underlying instrument will not.

The human premium

If perfection is no longer the proxy for investment, something has to be. The reallocated signal — the thing now carrying the weight polish used to carry — is what I’d call the human premium: the measurable value of demonstrable human judgment in the output.

I’ve argued elsewhere in this Codex that Industry 5.0’s actual thesis was never the technology. Industry 4.0 automated the process; 5.0’s premise was that the human returns to the loop not as a fallback but as the value-bearing element — the part you’re paying for. That framing, which read as slightly soft when it was formulated, is now simply the economics. When machines produce the competent baseline for free, the entire margin migrates to whatever the machine can’t do. And what it can’t do is care about a specific thing, in a specific way, at a cost.

The human premium has three components, and they’re worth separating because organisations tend to think they’re buying one when they’re buying another:

Consequence. The output is attached to someone who bears the cost of being wrong. A named author with a position. A designer who signed it. This is why the byline is quietly becoming the most valuable real estate on a page, and why anonymous “content” — the entire SEO-farm apparatus of 2015–2023 — is now worth approximately zero, not because Google punished it but because readers stopped extending it credit.

Specificity. The output contains something that could only have come from a particular vantage point on a particular day. Not “brands are grappling with AI” but the exact detail nobody would have invented. Specificity is expensive because it requires having been somewhere.

Restraint. The output declines to include what the distribution would have suggested. This is the hardest one and the most diagnostic, because it is invisible — it’s the paragraph you cut, the effect you didn’t apply, the three trend-signals you left out because they weren’t yours. Restraint is the only one of the three that cannot be performed. You can fake consequence with a fabricated byline and fake specificity with an invented anecdote. You cannot fake having not done something.

The trap that’s already sprung

Here is where I’d ask for some intellectual honesty from the room, because the obvious move is already being made and it will not work.

The moment “imperfection reads as human” became legible as a market insight, the manufacture of imperfection began. Deliberately grainy product photography. Lo-fi shot on a phone, allegedly. The vocal crack left in — engineered, in the third take, on purpose. Hand-drawn typefaces produced by a plugin. Founder letters with a strategically placed typo. The word “honestly.” Podcast ads read badly on purpose.

This is a category error, and the audience is going to eat it alive, for a reason that’s already in the argument above: a manufactured flaw is not a costly signal. It costs nothing. It risks nothing. It is, structurally, just another output from the centre of the distribution — the distribution has simply learned that the centre now includes a bit of grain.

Distressed jeans didn’t make anyone look like they’d worked. They made everyone look like they’d bought distressed jeans, and within a season that read exactly as clearly as it was.

So the strategy that follows from the imperfection economy is not “add flaws.” It’s considerably more demanding, and more expensive, and that expense is the entire point. It is: make things that could fail, attach a name to them, and don’t remove the part that makes them yours. The flaw is not the asset. The flaw is a symptom of the asset, which is a real decision made by a locatable person who could have made a different one.

Which means the imperfection economy is badly named, and I’ll wear that — it’s a description of the surface. The thing underneath is an authorship economy. The market didn’t develop a taste for rough edges. The market lost its ability to read polish as evidence of a human, went looking for a replacement signal, and found the only one that has ever actually worked, which is the visible fact of someone having decided something.

Perfection was never the product. It was the receipt. We just spent a hundred years mistaking the receipt for the meal, and now that the receipts print for free, we’re going to have to prove we cooked.

The Imperfection Economy: Why Flawless Is Now Worthless

imperfection economy · human premium · AI content authenticity · costly signalling · brand differentiation AI · authorship economy · AI slop backlash · Industry 5.0 human premium