Humans Beat LLMs on Open Creative Tasks, Imagination Accounts for 22-45% of Score
ValerioCapraro · x · 2026-10-12
A new paper by Charness and Grieco in The Economic Journal compared three versions of ChatGPT against 738 human participants on two creative tasks, yielding 4,428 blind evaluations on a 1-10 scale where raters didn't know whether answers came from humans or AI.
Key findings:
- Closed task (write a story using ten specified words like "house", "cow", "planet"): LLMs outperform humans on average.
- Open task (invent anything you can imagine and explain why): humans significantly outscore all three LLMs.
The authors model imagination as the capacity to generate ideas beyond what instructions prescribe, and estimate it accounts for 22%–45% of humans' creative scores on the open task. The takeaway: humans stay more creative on open problems by jumping beyond prescribed boundaries.
More from AGI Musings
- PopVax's Broadly Protective COVID Vaccine Enters First-in-Human Trial, Will Be Open-Sourced — rishabh16_ · 2026-10-12
- AI won't kill math research — it will select for Hardy-style expositors, argue mathematicians — nitarshan · 2026-10-12
- Classic Hall & Jones paper predicts AI will keep pushing health spending share up — soumitrashukla9 · 2026-10-12
- Grothendieck vs Deligne: what the Weil conjectures saga says about progress today — littmath · 2026-10-12
- Byung-Chul Han's 'attentional fasting': what algorithmic feeds took from us — floguo · 2026-10-12
- a16z podcast: recursive self-improvement is happening, but closed loops don't guarantee takeoff — a16z · 2026-10-12