ChatGPT matches the top 1% of humans in creative thinking

Generative AI has always been praised for grunt work like summarizing notes or drafting emails. When it came to genuine creative spark, the consensus was that human brains still held the crown.

A study led by Dr. Erik Guzik at the University of Montana just poked a massive hole in that belief. In a classic psychological evaluation designed to measure human creativity, GPT-4 scored in the top 1% of human test-takers for original thinking.

Testing GPT-4 on protected psychological exams

The research team, which included Christian Gilde from UM Western and Christian Byrge from Vilnius University, used the Torrance Tests of Creative Thinking (TTCT). Developed decades ago, the TTCT is a gold standard in psychology that mimics real-world brainstorming tasks, such as asking participants to list unconventional uses for everyday items like bricks or basketballs.

Crucially, TTCT materials are proprietary and strictly protected. ChatGPT could not scour open web databases or cheat by regurgitating leaked answer keys.

The researchers submitted eight response sets from ChatGPT alongside answers from a control group of 24 college students. All tests were sent directly to the Scholastic Testing Service for blind grading, meaning the evaluators had no idea artificial intelligence was in the mix.

The results were then benchmarked against a national baseline of 2,700 college students who took the exam in 2016.

ChatGPT placed in the top percentile for both fluency (the raw volume of relevant ideas) and originality (how rare and novel those ideas were). It slipped slightly on flexibility—the ability to jump between entirely different conceptual categories—landing in the 97th percentile.

That marks a dramatic leap from previous tests run on GPT-3 just a year earlier, where the older model lagged well behind human peers on original answers.

When Guzik asked ChatGPT what such high marks would imply, the chatbot offered a telling reply: humans may not fully understand their own creativity yet, and new evaluation metrics might be needed to distinguish human ideas from machine outputs.

Why it matters

We think this study dismantles the lazy claim that large language models are merely bland predictive text engines incapable of novelty. When scored against established rubrics with predefined lists of common answers, GPT-4 reliably produced ideas that human graders considered exceptionally fresh.

For entrepreneurs, product designers, and marketers, this makes GPT-4 an immediate asset for early-stage ideation. If your team is stuck in conventional groupthink, running divergent thinking prompts through the model costs next to nothing and delivers top-tier breadth.

However, high TTCT scores should not be mistaken for complete creative mastery. Divergent thinking is only the starting line of innovation. Coming up with unusual uses for a basketball is easy; recognizing which ideas are technically viable, economically sound, and culturally resonant requires taste and real-world judgment.

The smart play here is not treating AI as an autonomous visionary, but using its raw divergent horsepower to shatter creative bottlenecks while keeping human judgment firmly in the editor's seat.