Imagine a task that is simple enough to take 4 minutes. Generate 10 words. That’s it. Make them as different from each other as possible in all important ways: meaning, usage, sound in the mouth. This is not a test you might find in an IQ exam, or one that is framed as a competition to sift human talent from machine learning. This is a creativity test that psychologists have spent years validating as a true indicator of divergent thinking. Now, for the first time, researchers have run it in parallel on 100,000 people and multiple artificial intelligence models.
As with new scientific discoveries, the results are disturbing. They break down one assumption and raise something more complex in its place.
On January 21, 2026, a team led by Karim Gerbi from the University of Montreal published their findings in Scientific Reports. The core of their findings reads like a headline meant to provoke. Some of the leading AI language models, GPT-4 among them, currently outperform the average creative human. They do this on an objective and measurable scale. They are over the middle. But here’s the catch everyone has been waiting for in one way or another. That means even the best AI systems still fall short of what the most creative humans can do. The top 10%? They create even bigger gaps.
“Our research shows that some AI systems based on large-scale language models can now outperform average human creativity on well-defined tasks,” Gerbi says. “While these results may be surprising and even disturbing, our study also highlights an equally important observation: Even the best AI systems still fall short of the levels reached by the most creative humans.”
This is the perfect expression of a paradox. We have crossed the line, but we have not yet crossed it.
What is important about Gerbi’s work is not that it contains shockers, but that it is truly systematic. His team did not compare a single AI model to a small number of human volunteers. They tested multiple major language models (including GPT-4, GeminiPro, Claude 3, and GPT-3.5) on a large, age-, gender-balanced, and geographically diverse cohort. The study includes co-lead authors Antoine Bellemare Pépin and François Lespinasse, and also counts Yoshua Bengio, one of the founders of deep learning itself and currently working at Antropic, as a co-researcher.
The task they used was the Divergent Association Task (DAT). You are asked to generate 10 words, and your score is determined by semantic distance (how far apart the words are from each other within the language’s vast conceptual space). A highly creative person might suggest “galaxies, folk, freedom, algae, harmonica, quantum, nostalgia, velvet, hurricanes, photosynthesis.” The list seems to vary. Nothing resonates. Nothing can be predicted.
The measures work. Researchers verified that performance on the DAT predicted performance on other creativity tests (alternative use tasks, insight questions, and real-world creative writing). It’s not a parlor trick. It’s a true window into how creatively the human mind can function.
So what happened when Gerbi’s team used the machine to run this test?
GPT-4 won. Achieved an average score that exceeds the average human score. GeminiPro reached levels that were statistically indistinguishable from the human average. Several other models underperformed to varying degrees. But here’s the important thing. When slicing human data into segments, the moment you see people in the top half through creativity, all AI models fall below that threshold. Add together the top 10 percent of humans (approximately 10,000 participants) and you open up a gap that even GPT-4 cannot cross. The data reveals a ceiling. That’s what the machine did.
“Based on data from over 100,000 participants, we have developed a rigorous framework that allows us to compare human and AI creativity using the same tools,” Jarvi explains. The sheer scale of the work is part of what makes it so appealing. Previous research on AI creativity has yielded mixed results, with sometimes machines outperforming humans and sometimes the opposite. Often they relied on smaller samples and competing metrics. This study focuses on that claim by examining 100,000 human data points.
But that’s not all. The research team asked whether it was possible to tune the creativity of AI. can. They were able to elicit higher creativity scores from GPT-4 by adjusting temperature, a hyperparameter that controls the randomness of the model’s word sampling. At low temperatures, it produces conservative and predictable output. Higher temperatures lead to more randomness and more exploration. Pushing it up causes the model to take more risks, go beyond well-worn paths and generate more diverse associations. The highest temperature tested yielded an average score higher than 72 percent of human responses.
They also experimented with strategies. They asked the model to generate 10 words using a variety of prompts. One strategy focused on etymology (the roots and origins of words). It worked. Both GPT-3.5 and GPT-4 significantly improved their scores when explicitly instructed to change their etymology. Another asked for semantic opposites (words with opposite meanings). Not surprisingly, opposite words were semantically close, which led to poorer performance. Importantly, AI systems are sensitive to how humans frame tasks. It is not locked into a single mode. they adapt.
To test whether the DAT improvements actually translated into real-world creative work, Jarvi’s team had the model generate haiku, movie summaries, and flash fiction stories. They measured these using something called Divergent Semantic Integration, which tracks semantic distance across languages at the sentence level. They also applied this metric to haiku and plot summaries written by humans. Results: GPT-4 consistently outperforms GPT-3.5 on all three writing tasks. But humans, especially those who have sampled works from specialized sources, maintain an advantage. Haiku written by humans outperforms machines. Human movie plots involve more semantic complexity. The machine approaches, but never arrives.
Does this mean that machines will replace writers, artists, and inventors, a fear that has haunted creative professionals since the remarkable rise of generative AI?
“Although AI can now achieve human-level creativity in certain tests, we need to move past this misleading sense of competition,” Gerbi says. “Generative AI has become, above all, an incredibly powerful tool for human creativity. Generative AI will not replace creators, but it will profoundly change the way those who choose to use it imagine, explore, and create.”
The findings suggest that such concerns remain premature at this point. The creative work that sustains a career—the award-winning writing, the field-defining conceptual breakthroughs, the ideas that resonate—comes from the highest levels of human creative ability. The machines haven’t arrived there yet. They reached the median value. They have been rising towards the average. But they are not at the top.
But the study raises other questions, and Jarvi cautiously notes: “Studies like ours encourage us to rethink what it means to be creative by directly confronting the capabilities of humans and machines.” If GPT-4 can go beyond the average human divergent thinker, generate semantically complex stories, and condition, prompt, and persuade to be more original, then perhaps the traditional hierarchy of human creative abilities will be completely different than we thought. Perhaps creativity is not a single thing. Perhaps the kind of associative, combinatorial thinking that machines are currently capable of doing represents an aspect of a larger, more complex phenomenon.
Research data is publicly available. The code is on GitHub. This is the kind of work designed to be built upon, extended, and revisited as new models emerge and new questions arise. For now, one thing is clear. It means we’ve crossed an interesting threshold where machines are incomparably worse at creative tasks. But the question of what they actually are, how they work, and what their existence means for human creativity and creative professions is only just beginning.
Research link: https://www.nature.com/articles/s41598-025-25157-3
related
Learn more from NeuroEdge
Subscribe to receive the latest posts by email.
