A new preprint from Moon, Johnson and seven co-authors puts a number on something the AI brainstorming industry has been waving away. In a sample of 1,600 college admissions essays, applicants whose essays extended deeper into AI's blind spots had higher SAT scores (β = .11, 95% HDI [.04, .18]).
Why this reframes the brainstorm
The study ran across 1,600 college admissions essays, 481 alternative-use bundles and 1,018 interpersonal-advice responses. The human ideas sitting in regions AI output rarely covered were as good or better than the AI's matches on nine preregistered outcomes.
The exec reflex, when an AI brainstorm returns five options, is to pick the "most original" three. That pick runs on a signal that works for humans and barely works for AI. The EDGE-creativity slope is .67 for humans and .05 for AI essays. For alternative-use ideas the slope reverses to -.11. An AI option that reads as original on an automated score is often sitting in a region densely occupied by other AI outputs. The signal only tracks creativity when it is reading a human.
Flag the limits. This is a preprint, so it has not been peer-reviewed. The AI side uses GPT-5-mini across all three tasks. The admissions essays come from matriculated students in pre-ChatGPT cohorts. EDGE measures how far an idea sits from AI output in semantic space, so it is a proxy for creativity rather than a direct measure. The finding replicates across three independent task families and nine preregistered outcomes, which is stronger than one task tested on multiple models.
A room where I watched this happen
A head of strategy at a professional services firm in Zurich. An offsite on next year's service portfolio. Someone on her team had prepped a ChatGPT brainstorm with five new service lines. By the time I walked in, the room had been arguing for forty minutes about which three to greenlight.
A senior associate, two years in, put up a sixth idea. It did not fit any of the five frames. The room's reflex was to rephrase her option until it looked like a variant of option two.
I asked them to leave it unframed. I asked them to describe why option six did not fit. They described it well. They picked it.
Moon's data puts a number on what happened in that room. The AI covered the predictable regions of the service-line idea space by design. The work the firm is paid for sits in the regions AI did not cover. The sixth option was where the margin lived.
"In essays, ideas that fell further outside the space of AI-generated responses ... received higher creativity evaluations."
Next time an AI brainstorm comes to your room with five options, watch for the sixth. Keep it where it is.