Large language models now beat the average human on creativity tests, but their stories, designs and ideas keep coming out alike, and the most creative humans stay ahead
In January 2026, a Université de Montréal team compared 100,000 people with nine AI systems. GPT-4 beat the average human on a standard creativity test; the top 10% of humans still beat every model.
Newer work keeps finding the same split. A Cornell study of 20,000 AI-written stories found nearly 90% reused one of 11 stock words, including 'Elias' and 'lighthouse'. A Duke comparison of 22 models against more than 100 people found model answers far more alike than human answers.
Why it matters
As AI matches baseline human creativity, creative industries must decide which work still needs human judgment.
Questions about this story
Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.
No questions yet — be the first to ask.
Key Indicators
0.592
LLM Design Originality
Same-prompt design pairs from different models scored 0.592, versus 0.764 for human-model pairs
21%
GPT-5 Agent Medal Rate
GPT-5 with AIDE earned a medal in only 21% of runs, despite roughly double the novelty of gold-medal humans
~90%
AI Stories With Stock Words
Share of 20,000 AI-written stories that used at least one of 11 recurring words; more than half featured a lighthouse
Images from Openverse under Creative Commons licenses.
Videos from YouTube.
Voices
Curated perspectives — historical figures and your fellow readers.
Mark Twain
(1835-1910) ·Gilded Age · wit
Fictional AI pastiche — not real quote.
"I see they've finally built a machine that can be as mediocre as most of us—though I confess it took them longer than I expected. The real trick wasn't teaching silicon to have ideas, but teaching it to have the *wrong* ideas with the same confident regularity as a committee of humans, which suggests we've been training these contraptions on congressional records."
0% found this insightful
Sign up to generate historical perspectives on this story.
A new benchmark for UI design creativity found LLM outputs more repetitive than human ones. Originality scored 0.592 for model pairs versus 0.764 for human-model pairs; creative range 0.581 versus 0.902.
Cornell Finds AI Fiction Repeats Itself
Research
Researchers prompted four language models for stories 20,000 times. Nearly 90% of outputs used at least one of 11 recurring words, including 'Elias,' 'Mara' and 'lighthouse,' and more than half featured a lighthouse. The paper was presented at EMNLP 2026 in Budapest.
August 2026
Automatic Creativity Scoring Questioned
Research
A preprint found automatic scoring systems misalign with human judgments of creative text. LLM judges systematically preferred AI-generated stories over human ones.
LLM Creative Outputs Converge Across Providers
Research
A review of three years of model releases found declining cross-provider output diversity, most sharply on the Alternate Uses Task.
July 2026
AGC-Bench Tracks General Creativity Factor Across 83 Models
Research
A benchmark of 83 frontier LLMs found a general creativity factor explaining 81.5% of score variance. Even so, the top human outperformed the best model.
June 2026
Humans Edge Out LLMs in Large-Scale Creativity Study
Research
A Nature Human Behaviour study of 9,198 humans against 215,542 LLM outputs found human creativity slightly higher on average. Humans showed greater variability and stronger top-end scores; prompting models to adopt genius personas lifted performance only to a threshold.
January 2026
Largest AI Creativity Study Published
Research
Université de Montréal team publishes comparison of 100,000 humans against nine AI systems. GPT-4 exceeds average human creativity; top 10% of humans still outperform all AI.
2026
GPT-5 Agents Show High Novelty, Low Feasibility
Research
In a competitive machine-learning setting, GPT-5 with the AIDE framework produced solutions with roughly double the historical novelty of gold-medal humans, yet only 21% of runs earned any medal.
Creativity Judgments Shift With AI Label
Research
A BMC Psychology study found that labeling a creative product as AI-made or human-made shifts how creative people judge it to be. The direction of the bias depends on the domain.
Generative AI Use Makes Human Output More Uniform
Research
A meta-analysis of 19 studies found that people who co-create with generative AI produce ideas, designs, and texts more similar to one another's.
July 2024
AI Story Ideas Lift Individuals, Narrow the Pool
Research
An online experiment found writers given story ideas from a language model produced stories rated more creative, better written and more enjoyable, especially among less creative writers. Those stories were also more similar to each other, which the authors described as a social dilemma.
February 2024
University of Arkansas Study
Research
Researchers pit 151 humans against GPT-4 across three divergent thinking tests. GPT-4 proves more original and elaborate than human participants across all measures.
August 2023
First Major AI-Human Creativity Comparison
Research
Koivisto and Grassini compare 256 humans against ChatGPT on creativity tests. AI chatbots outperform average humans, but the best humans still exceed AI.
June 2021
Divergent Association Task Published
Methodology
Researchers publish the Divergent Association Task in PNAS, offering a faster, computationally scorable creativity test that measures semantic distance between word choices.
1967
Alternative Uses Task Developed
Methodology
Psychologist J.P. Guilford creates the Alternative Uses Task, asking people to generate creative uses for everyday objects. It becomes the standard test for divergent thinking.
Scenarios
1
AI Becomes Standard Creative Collaborator
Likely
Discussed by: McKinsey, World Economic Forum analysts, creative industry researchers
AI tools become embedded in creative workflows the way spell-checkers became embedded in writing. Writers, designers, and artists routinely use language models to generate initial ideas, overcome creative blocks, and accelerate brainstorming. This lifts the baseline quality of average creative work while leaving exceptional creativity as a human differentiator. Entry-level creative roles shrink as AI handles routine ideation.
2
Creativity Tests Proven Inadequate for AI
Possible
Discussed by: Cognitive scientists, AI researchers examining methodology limitations
Further research reveals that current creativity tests measure only narrow dimensions of creative ability—semantic distance and divergent thinking—while missing judgment, cultural context, and meaningful novelty. AI appears creative because tests reward unusual word combinations, not because it generates genuinely useful or moving ideas. The field develops new benchmarks that capture what creativity actually means.
3
AI Creativity Ceiling Persists
Possible
Discussed by: Researchers noting persistent human advantage among top performers
Subsequent studies confirm that while AI continues improving on average creativity measures, the top tier of human creators maintains an unbridgeable gap. Something about exceptional human creativity—perhaps grounded experience, genuine intentionality, or cultural embeddedness—remains beyond algorithmic reach. This stabilizes a division of creative labor where AI augments but doesn't replace top creative talent.
4
AI Breaks the Creativity Ceiling
Uncertain
Discussed by: AI capability researchers, technology forecasters
Future model architectures or training approaches enable AI to match or exceed even the most creative humans on all measured dimensions. Creativity joins chess, Go, and protein folding as domains where AI surpasses human performance entirely. This triggers fundamental reconsideration of what makes human contribution valuable in creative fields.
An August 2026 preprint shows automatic scoring systems misalign with human judgments of creative text, and LLM judges systematically prefer AI-generated stories. If the field's core measurement tools are flawed, the 2026 AI-vs-human creativity scores come into question. Expect pressure for new benchmarks that reflect human judgment.
A three-year review found LLM creative outputs becoming more similar across providers. As models converge on similar training data and techniques, AI-assisted creative work may gain speed but lose variety.
7
Applied Creative Tasks Expose AI's Repetition Problem
Likely
Resolves by Q2 2027
Discussed by: Design Creativity Bench authors, competitive ML researchers
Benchmarks on applied tasks like UI design and competitive coding show LLMs produce appropriate but repetitive outputs. GPT-5 agents generate novel solutions that rarely work. If this pattern holds, AI will augment rather than replace creative work that demands feasibility and variety.
Historical Context
3 moments from history that rhyme with this story — and how they unfolded.
1 of 3
May 1997
Deep Blue Defeats Kasparov (1997)
IBM's Deep Blue chess computer defeated world champion Garry Kasparov in a six-game match—the first time a computer beat a reigning world champion under standard tournament conditions. Kasparov had won against Deep Blue the previous year.
Then
The defeat sparked debate about machine intelligence and led Kasparov to accuse IBM of cheating. IBM retired Deep Blue rather than granting a rematch.
Now
Chess engines became standard training tools for human players. Top players now routinely use AI analysis, and computer-assisted 'centaur chess' emerged as a format combining human intuition with machine calculation.
Why this matters now
Chess was once considered proof of human intellectual superiority. Its fall to machines prompted predictions that creativity would remain humanity's last bastion—a claim now under direct empirical testing.
2 of 3
March 2016
AlphaGo Defeats Lee Sedol (2016)
Google DeepMind's AlphaGo defeated Go world champion Lee Sedol 4-1 in Seoul. The game of Go, with more possible positions than atoms in the universe, was believed to require human intuition and creativity to master.
Then
Move 37 in Game 2—an unconventional shoulder hit—stunned observers and became iconic as evidence of machine creativity. Lee Sedol retired from professional Go in 2019, citing AI's unbeatable dominance.
Now
Go players adopted AI training partners. The victory accelerated investment in deep learning and reinforced that machine learning could master domains requiring apparent creativity.
Why this matters now
AlphaGo's creative, counterintuitive moves challenged the assumption that pattern recognition can't produce genuine novelty. The Montreal creativity study tests whether this extends to open-ended creative tasks.
3 of 3
June 2020
GPT-3 Demonstrates Creative Writing (2020)
OpenAI released GPT-3, a 175-billion parameter language model that generated essays, poetry, and code from brief prompts. The Guardian published an essay written entirely by GPT-3, sparking widespread discussion about AI authorship.
Then
Writers and educators debated plagiarism implications. OpenAI initially limited API access due to misuse concerns.
Now
GPT-3 catalyzed the generative AI industry. By 2024, language models became standard tools for drafting, editing, and brainstorming across industries.
Why this matters now
GPT-3's fluent output raised the question that the Montreal study now answers empirically: can AI creativity be measured, and how does it compare to human performance?
