What happens to a generative model that keeps learning from generative output: the range narrows, the rare cases disappear, and everything it makes starts to sound like everything else.
Also known as recursive training degradation, model autophagy, AI inbreeding
Model collapse is the degradation that sets in when a generative model is trained on data produced by generative models, including its own earlier output. Each training round treats the previous round’s output as if it were a fresh sample of the real world. The rare cases, the outliers, and the unusual phrasings sit at the tails of the distribution and get sampled least often, so they thin out first. What survives is the middle.
The effect compounds quietly. One generation of synthetic training data barely registers. Several generations in, the model has lost the ability to produce the unusual thing at all, because it no longer holds much evidence that the unusual thing exists. Researchers distinguish early collapse, where the tails vanish, from late collapse, where output converges into a narrow band with very little variance. Neither announces itself. The writing stays fluent and the images stay clean, which is exactly why it is hard to catch by looking.
Agencies do not train foundation models, so collapse can look like somebody else’s problem. It is not, for three reasons.
The sameness has a cause. If AI-assisted concepting keeps landing on the same visual language, the same headline rhythm, and the same three metaphors for transformation, some of that is the prompt and some of it is the distribution the model is drawing from. Knowing which is which changes what you do about it.
It sets a ceiling on AI-first creative. A model pulled toward the middle is very good at producing competent work and structurally worse at producing the strange, specific idea that wins a pitch. That is an argument for putting AI on the volume and humans on the differentiating ten percent, and it is an argument built on mechanics rather than sentiment.
It turns your own archive into an asset. A studio’s past work, its client interviews, its transcripts, and its rejected routes are human-made material that no model has averaged. Feeding that into retrieval, fine-tuning, or even a well-built prompt library is a practical hedge against drawing only from the shared middle.
A studio runs its concept sprints through a house model prompted from the brief. In March the work feels sharp. By September the art director notices that three unrelated clients (a credit union, a hiking brand, and a meal kit) have each been handed a quiet-confidence route with the same muted palette and the same sentence shape in the headline. Nothing was copied. The model simply has a center of gravity, and the sprints kept landing on it. The fix is not a cleverer prompt. It is loading the sprint with material the model has not averaged: the client’s own customer interviews, the category’s weirdest historical campaigns, and the studio’s own rejected work from two years ago.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.