The empirical literature on AI and creativity has converged on a finding that sounds like a contradiction and is not. The finding has two parts.
Individual creativity goes up. When a single person uses AI to brainstorm, draft, or iterate, their measured output on standard creativity instruments — fluency, flexibility, elaboration, originality — improves. They produce more ideas, more varied ideas, more developed ideas, and ideas that are statistically less common in the population. This effect is robust across study designs and replicates well.
Collective creativity goes down. When many people in a population use AI on the same tasks, the aggregate diversity of outputs falls. The same suggestions, the same structures, the same combinations of ideas appear repeatedly across users who never communicated with each other. The collective shrinks even as each individual contribution grows.
Both findings are real. Both can be true at the same time.
How both are true
The mechanism is straightforward once stated. Each user’s creative score is measured against the average user without AI. Add AI to one user, and they beat the average. Add AI to everyone, and the average shifts; what looked novel against the old baseline looks ordinary against the new one. The novelty is not an illusion at the individual level; it is an illusion at the collective level.
The deeper mechanism: AI raises the floor of creativity by making mediocre ideas accessible to people who could not produce them without help. It does not raise the ceiling. The ceiling is set by the rare cases where someone combines training-corpus material in a way the corpus does not predict — and those cases are not what the model was optimized to produce. The model was optimized to produce plausible, well-formed material. Plausibility is not novelty. Well-formedness is not surprise.
A pre-AI distribution of creative output had a wide spread: many bad ideas, a few brilliant ones, and a long thin tail of unique combinations contributed by the community of practitioners. An AI-saturated distribution has a higher mean, a smaller variance, and a shorter tail. The pieces in the tail are the ones that move a field.
What the studies actually show
The Doshi-Hauser methodology is the cleanest illustration. Participants are asked to generate creative responses to standardized prompts (story beginnings, business-idea generation, etc.). Half use an AI partner; half work alone. The individual-level scoring favors the AI users — more ideas, more elaborated, more varied per-person. The collective-level scoring — how different the AI users’ responses are from each other — favors the unassisted users. The unassisted group produced fewer, rougher ideas; those ideas covered more semantic territory.
A field of practitioners using the same AI partner risks reproducing this pattern at scale. Each practitioner’s individual work improves; the field as a whole loses some of the variety that drives breakthrough. The cost is hardest to see in any single comparison and largest in the aggregate.
The optimist’s reply
A reasonable optimistic reading of the same data: the individual-level gains are real, and many people who could not previously contribute creatively now can. The collective-level loss is partly an artifact of measurement — the old baseline contained a lot of low-quality variety, and what AI has done is substitute medium-quality consistency. Whether that’s a net gain depends on what you wanted from the field.
This reading has merit for many domains. A pre-AI workforce produced lots of mediocre marketing copy; an AI-assisted workforce produces consistently medium marketing copy with less variance. For marketing copy, that is fine.
For fields where the long tail is the point — research, art, criticism, philosophy — the same trade is more troubling. The fields advance through outliers. Compressing the distribution toward the mean does not preserve the field’s engine.
What can be done
Three responses, sketched briefly:
Use AI for the floor, not the ceiling. Lean on AI for first drafts, synthesis, and routine work; protect the late stages of creative work — the moments where genuine novelty emerges — from default AI use. This is a habit that has to be built, against the grain of the tools.
Maintain non-AI practice. A practitioner who only ever produces with AI is a practitioner whose unassisted creative skills are atrophying (see B.09). Periodic unassisted practice is part of staying capable in the long tail.
Diversify model use. A field where every practitioner uses the same model will converge faster than a field where they use different models. This is weak medicine but real.
The encyclopedia’s framing, after the standardization section’s larger argument: AI is not the enemy of creativity; uniform AI use is. The distinction matters. The first is a tool question; the second is a demographic one.