Has artificial intelligence caught up with human creativity? A study in Scientific Reports led by Karim Jerbi, professor of psychology at Université de Montréal and associate member of Mila, with co-first authors Antoine Bellemare-Pépin and François Lespinasse and collaborators including Yoshua Bengio, Jay Olson of the University of Toronto Mississauga and Google DeepMind, offers the most systematic answer yet, comparing several large language models with more than 100,000 human participants.
The core measure was the Divergent Association Task, in which participants produce ten words as semantically distant from one another as possible. Some models, including GPT-4, outperformed the average human on this test of divergent linguistic creativity. But the average score of the most creative half of humans exceeded every AI model tested, and the top 10 per cent widened the gap further. On creative writing tasks, including haiku, film plot summaries and short stories, the most skilled human writers retained a clear advantage.
The study also showed how much AI creativity depends on human steering: raising a model's temperature setting produced more varied output, and prompts built on etymology elicited less obvious associations and higher scores. The authors conclude that AI is best understood not as a replacement for creators but as a tool that expands human exploration, with the interaction between person and machine now central to the creative process itself.