0
0
In January 2026, a research team in Montreal published a number that’s hard to shrug off if you’ve ever thought of yourself as “the creative one.” They gave more than 100,000 people a creativity test called the Divergent Association Task — then gave the exact same test to several AI language models, including some built by the very people running the study.
The task is almost embarrassingly simple. Name ten words. The only rule: the words can’t relate to each other in any way. “Dog” and “bone” fail instantly — too connected. “Umbrella” and “thunderstorm” fail too. The farther apart your words are in meaning, the higher you score. It’s one of the most widely used shortcuts researchers have for measuring divergent thinking — the mental skill underneath brainstorming, problem-solving, and a good chunk of what we call art.
When the results came in, the machines won. On average, the AI models scored higher than the humans.
The Test 100,000 People Just Took
The study, “Divergent creativity in humans and large language models,” was published in the peer-reviewed journal Scientific Reports, out of the Universite de Montreal and Concordia University. Lead researcher Karim Jerbi worked with co-authors Antoine Bellemare-Pepin and Francois Lespinasse — and Yoshua Bengio, one of the most cited AI researchers alive, was part of the team.
The Divergent Association Task earned its trust in psychology labs because it sidesteps the biggest problem with measuring creativity: subjectivity. You can’t easily hand a panel of judges 100,000 poems and get a consistent score. But you can measure how semantically distant a person’s ten words are from each other using the same word-embedding math that powers modern search engines and, not coincidentally, large language models themselves. It’s fast, it’s free of a grader’s personal taste, and it scales — which is exactly how one team managed to test six figures’ worth of humans against a handful of AI systems in a single study.
Inside the AI Creativity Test That Beat 100,000 People
Here’s the part worth sitting with: this result isn’t really a surprise once you understand what the test measures. Divergent word association rewards pulling from the widest possible net of unrelated concepts — and a large language model’s entire existence is built on holding an almost incomprehensibly wide net of word relationships in memory at once. Where a person has to search their own limited, tired, distracted brain for a word that has nothing to do with the last nine, a model can sample from a vocabulary trained on a meaningful fraction of everything ever written.
That’s a genuine advantage on this specific task. It is not evidence that a model wanted to write something beautiful, meant anything by it, or felt anything while doing it. The Divergent Association Task measures one narrow, real, and useful sliver of creativity — the ability to break out of your own mental ruts. It was never designed to measure intention, meaning, grief turned into a song, or a five-year-old’s drawing of her dog that isn’t technically a dog but is somehow, unmistakably, about love.
This isn’t the first time this year a head-to-head study of humans against AI has landed somewhere more interesting than the headline suggested. A University of British Columbia study on loneliness found that texting an actual stranger beat an AI chatbot at easing isolation — even though the chatbot was, by most measures, the more polished conversationalist. The machine can outperform us on the metric we hand it. That’s different from replacing what we were actually looking for.
Averages Compress a Range Into a Single Number
It’s worth remembering what an average actually does: it takes a wide spread of individual results and flattens them into one figure. A study reporting that AI beat humans “on average” is not the same as a study reporting that AI beat every human. Any group of 100,000 people includes people who’ve spent decades training the exact mental habit this test measures — poets, comedians, ad copywriters, kids who never stopped playing word games — sitting right alongside people who haven’t been asked to free-associate since grade school. The headline number is real. It’s also, by definition, not the whole story of any one person who took the test.
What Creativity Was Never Built to Prove
Here’s where the story gets interesting for reasons that have nothing to do with technology. Earlier this year, the U.S. government named an $800 million AI initiative the “Genesis Mission” — and whether or not anyone announcing it meant it this way, the word carries more history than a press release can hold. Long before anyone thought to measure creativity with a word list, the opening pages of the Bible describe a Creator making a universe out of nothing and then making a person to reflect that same Creator — and the very next thing that person is asked to do is name things, tend a garden, build, shape, make. Creativity, in that framing, was never presented as a talent you had to earn or a title you had to defend against competition. It was described as evidence of who you were made to reflect in the first place — closer to a birthmark than a trophy. It’s the same idea an ancient poem got to thousands of years before anyone had a word for “design.”
That’s a strange comfort to land on in a week when a headline says a machine out-performed you at something you thought was uniquely yours. If creativity was never the score you were protecting, losing a round of it to a computer doesn’t touch the thing it was actually pointing to.
Maybe that’s worth remembering the next time you sit down to make something — a meal, a birthday card, a bad first draft of anything. The point was never to win the test. It never was.
Discussion Question
If a machine could out-create you tomorrow on every test anyone could design, would that change why you make things in the first place? Tell us in the comments.
Share This
- AI just out-scored 100,000 people on a peer-reviewed creativity test. It made me ask a bigger question than “who’s more creative” — what was creativity even for in the first place?
- A study out of Montreal just had AI beat 100,000 humans on a standard creativity benchmark. Before you panic, it’s worth knowing exactly what that test does and doesn’t measure.
- The machine won the word game. It still can’t tell you why you made the thing you made. That’s not nothing.
Common Questions About the AI Creativity Study
What test did AI beat humans on for creativity?
Researchers used the Divergent Association Task, which asks a person to name ten words that are all unrelated to each other in meaning. The wider the semantic distance between the words, the higher the creativity score. It’s a widely used, peer-reviewed measure of divergent thinking.
Did AI really beat 100,000 humans at creativity?
A January 2026 study published in Scientific Reports by researchers at the Universite de Montreal and Concordia University, including Yoshua Bengio, found that large language models scored higher than humans on average on the Divergent Association Task, after more than 100,000 people took the same test.
Does this mean AI is more creative than humans?
Not in the full sense of the word. The test measures one specific skill — pulling unrelated words from memory — not intention, meaning, emotional stakes, or lived experience. It’s a real result on a narrow, well-defined task, not a final verdict on human creativity as a whole.
Why did AI score better on this particular test?
The task rewards drawing from the widest possible range of unrelated concepts, and language models are built on vast word-relationship data that gives them an inherent advantage at exactly that kind of recall — separate from whether they understand or care about what they produce.
What does this study actually mean for how I see my own creativity?
It’s a reminder that a single benchmark, however well-designed, was never built to measure the full weight of why a person makes something. Creativity tied to meaning, memory, and identity isn’t something a word-association score can take from you.
