The idea nobody could place

Coaching an innovation project this summer semester, one question had been sitting with me the whole time: whose idea was this, actually, and how much of it came from the model? It came out in a discussion round with the team: everyone said it was theirs at first, but the more we talked it through, the more they started to wonder whether that was actually true.

I wrote about the convergence side of this earlier this year: FH Vorarlberg teams kept arriving at the same clean, competent MVP, good enough to be hard to criticise, rarely strange enough to be disruptive. Generative models are good at producing plausible next steps, and when different teams keep asking for those next steps, very different starting points can begin to converge on similarly reasonable answers.

A panel discussion on model collapse recently made me look at that observation differently. Model collapse itself is a training-time phenomenon, when models are increasingly trained on model-generated data. What stuck with me wasn’t the mechanism, but the image: generation feeding generation until variation gets smoothed away.

I used to think of that as something that only happened inside a model (somewhere in a training run). What I was watching in the coaching room wasn’t literally model collapse, though the parallel made me uncomfortable: people feeding an idea back and forth through a model until nobody could really say anymore which parts they had actually worked through themselves.

And at that point, the question is no longer whether AI helped with the idea. It is whether anyone still owns it.