Huang, Yu, Chen and Zhou ran a preregistered RCT with 148 three-person teams at a large Chinese public university. The same 15 minutes of model access lowered expert-rated creativity by 0.62 SD when it arrived at the selection stage instead of the search stage.

The access question is settled

Most of the executive debate I hear is still about whether to give the team AI at all. The paper quietly closes that debate and opens a louder one. When in the workflow does the model arrive.

Early-search access raised ratings 0.83g over control (p=0.002). Later-selection access produced a 0.130 gain that was not significant (p=0.43). Same minutes, same model, same prompt. Opposite sign.

What late access does looks like progress. It re-opens generation at the gate (29.1% of selection-stage queries make new options, versus 11.0% under full-process access). It piles on 73% more proposals without raising quality. It inserts assistant text into the submission without team discussion 9.6% of the time, up from 3.9%.

The bite is who caught it. The authors ran a ten-model LLM-judge panel over the same submissions. It spotted the raw access benefit. It missed the timing cost. Human raters p=0.019, model panel p=0.77.

Any Copilot analytics tile that scores AI quality with another model is systematically blind to the exact failure this paper names.

A review meeting I have watched too many times

I was coaching a head of product at a B2B SaaS company in Amsterdam last month. Her team had three concept options for a new onboarding flow. Fifteen minutes into the review, someone opened Claude and asked it to help weigh the three.

Within two minutes the team had five options on the board.

She noticed. She said so. Then the room spent the rest of the meeting discussing options four and five instead of choosing between one, two and three. The meeting ended without a decision. Everyone left energised by the new ideas.

That is Huang's finding, named. At the selection stage, the assistant re-opens generation. The gate stops being a gate. Breadth looks like progress. The team feels productive. Rated output does not move.

The practical move is to put AI access upstream of your gates. Before the shortlist. Before the brand-approval meeting. Before the pitch review. Give the model fifteen minutes at the start of search, when adding options is the job. By the time the team sits down to pick, the pool is already wide and the meeting can do what the meeting is for.

"Moving the same 15 minutes of model access from search to selection lowered expert-rated creativity by 0.62 standard deviations."

If the team is going to use AI, pick the stage. The sign of the effect depends on it.