Disclosure: this post is adapted from my essay at piszczek.pl with the help of an AI model, then reviewed by me before posting. The sources, measurements and predictions are mine; the predictions are registered with a commit hash at piszczek.pl/predictions.
Epistemic status: confident about the individual measurements, which I've linked to primary sources; moderately confident about the synthesis; the three predictions carry explicit probabilities. I'd most like pushback on the mechanism claim and on which counter-evidence I've underweighted.
Claim. The sameness people notice in model output, and increasingly in human work made with models, is not the model returning the mean of its training distribution. It's mode concentration introduced in preference tuning, followed by anchoring on a shared first draft.
Evidence for.
- Artificial Hivemind (NeurIPS 2025): 25 models, 50 samples each, one open-ended prompt about a metaphor for time, two ideas. Cross-model similarity 71 to 82%.
- Wenger & Kenett: 22 LLMs vs 102 humans on standard creativity tasks. Models score slightly higher on individual originality and much lower on variety across responses (mean pairwise distance 0.459 vs 0.738).
- Lanchantin et al. (DivPO): for the standard objective, placing all probability mass on the highest-reward response is an optimal solution, and this stays true when several responses tie. A diversity-aware objective gets about 75% more story diversity at similar win rates, so the collapse is a default and not a necessity.
- Holzner et al. meta-analysis, 28 studies, N = 8,214: individual quality g = +0.27, diversity across people g = −0.86.
Evidence against, which I think is real.
- Fashion variety has risen since the 1980s in a 150-year image corpus; top-ten baby-name share fell from about a third to under a tenth; national music charts diverged after 2017.
- The human-side collapse depends on the interface. When the model asks or offers options and the human leads, quality rises and diversity holds (CHI 2026, N > 1,000). Incentives for originality restore collective diversity in two preregistered RCTs.
- MusicLab (2006): social influence makes outcomes more unequal and less predictable, which is not the same as more similar.
Where that leaves the claim. Narrower: the free tier converges, the free tier is most of the volume, and the price of legible difference rises. Airbnb data on 481,747 photos shows an inverted U between visual uniqueness and occupancy, which is the shape you'd expect.
Predictions (registered with thresholds and refutation criteria at https://piszczek.pl/predictions).
- Deezer's first 2027 disclosure still shows fully AI-generated tracks at 45% or more of daily deliveries. 80%.
- Axalta's 2027 report still has white, black, grey and silver at 78% or more of new cars worldwide. 85%.
- By the end of 2030 the name of an AI model or model family is in ordinary use as the name of a style (dictionary entry, stand-alone Wikipedia article, or museum exhibition title). 65%.
Full essay with a 15-row sources table: https://piszczek.pl/blog/the-average-is-free
