Generative AI Beats Average Human Creativity in 100,000-Person Study

A large-scale study from the Université de Montréal — co-authored by Turing Award winner Yoshua Bengio — has been recirculating heavily today after fresh coverage placed it in the context of where frontier models now sit on creative benchmarks. The finding: generative AI systems can beat the average human on standardized creativity tests, though the most creative humans still hold a meaningful lead.

The study compared the outputs of leading large language models — including ChatGPT, Claude, and Gemini — against responses from more than 100,000 human participants. The dataset is one of the largest direct AI-vs-human creativity comparisons published so far, and it lands at a moment when AI's role in creative work is being actively re-litigated across publishing, design, music, and film.

How the researchers measured creativity

The test the team used is a well-validated psychological assessment called the Divergent Association Task (DAT). Participants are asked to generate ten words that are as semantically unrelated to each other as possible. The score is based on how distant those words are in meaning, with originality and conceptual range counting heavily.

It's a deliberately narrow test — closer to "can you reach across categories?" than "can you write a novel?" — and it has the advantage of being scoreable without subjective judgment. That makes it well-suited for comparing AI outputs to a large human sample.

Frontier models, when prompted to take the test, scored above the average human respondent. Some of the strongest models edged into the upper-middle of the human distribution.

Where humans still win

The headline finding came with an important caveat. The top 10% of human participants — the most creative respondents in the dataset — still outperformed every AI model tested. That gap widened further on richer tasks like poetry, storytelling, and idea generation that requires emotional or contextual depth.

In other words: AI has caught up to median creativity, not peak creativity. The "creative class" — writers, designers, and artists doing the work that other humans find original and resonant — sits at the tail of the distribution AI hasn't yet reached.

The researchers framed this as good news for working creatives rather than bad. The most plausible read of the data is that AI raises the floor on creative output without raising the ceiling. Average outputs get better. The best outputs still come from people.

Why this matters

The result lands inside an industry debate that has gotten increasingly heated. Hollywood guilds, the music industry, news publishers, and visual artists have all spent the last year fighting over whether AI is a tool that augments creative workers or a substitute that replaces them.

The Montreal study points to a more nuanced answer. For commodity creative work — the kind where "good enough" beats "exceptional" — AI is now competitive with or better than the average professional output. That's the slice of the market where price competition will be most intense.

For high-end creative work — the kind where someone is hired precisely because their taste, voice, or originality is distinctive — the gap remains real. The best human creators are still doing things models cannot.

That's a different story from "AI replaces creatives" or "AI can't be creative." It's "AI compresses the middle of the market."

What this means for businesses

For marketing teams, content shops, design studios, and other organizations that produce creative output at volume, the implication is concrete: tasks that used to require a mid-level creative can now plausibly be done with AI assistance and a senior reviewer. The cost structure shifts.

For individual creatives, the read is that distinctiveness becomes the moat. Average is now a commodity. Voice, taste, and a recognizable point of view are what's defensible.

Several commentators have pointed out that this matches what's already happening in stock photography, copywriting, and basic graphic design — categories where AI output has visibly eaten into the freelance market while top-tier work continues to command premium prices.

Expert reactions

Bengio, one of the field's most credentialed voices, has emphasized that the result should not be read as "AI is creative" in the deep sense. The Divergent Association Task is a narrow proxy. Real creative work involves intent, lived experience, and an audience — none of which the test captures.

Other researchers have pushed back gently on the methodology, noting that AI models effectively trained on the kinds of word-association data that produce high DAT scores might be expected to do well on the test for reasons that don't generalize to "real" creativity.

What's next

Follow-up work from the team is reportedly looking at longer-form creative tasks — short fiction, poetry, and idea generation in specific domains — where the AI-human gap is expected to widen back in humans' favor. Those results would help quantify exactly where AI does and doesn't catch up.

In the meantime, the practical effect on the market is already visible. AI-augmented creative workflows are becoming standard inside agencies and in-house teams, with humans focusing on direction, editing, and the final 10% of polish.

Bottom line

The Montreal study won't end the AI-and-creativity debate, but it adds the most empirical data point yet: AI has caught average human creativity on at least one well-validated measure, and it hasn't caught the best humans. For everyone who works in creative fields, the question isn't whether AI changes the job — it's where you sit in the distribution.