---
id: "lib-2026-10-10-too-late-to-prune-generative-ai-the-screen-and"
type: "library"
title: "Too Late to Prune: Generative AI, the Screen, and Team Creative Decisions"
lang: "EN"
org: "arXiv (CAS, Zhejiang U.)"
source_type: "Paper"
date: "2026-10-10"
week: "2026-W42"
url: "https://nosov.ai/library/lib-2026-10-10-too-late-to-prune-generative-ai-the-screen-and/"
original_url: "https://arxiv.org/abs/2610.14984"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Chinese Academy of Sciences and Zhejiang University economists ran a preregistered experiment randomising 148 teams to five AI-access schedules, with a formal model, expert ratings and platform logs that trace the idea funnel."
summary_basis: "full-text"
original_minutes: "75"
tags: ["work and ai", "teams", "creativity", "deployment timing", "experiments"]
---

# Too Late to Prune: Generative AI, the Screen, and Team Creative Decisions

Tier A (Strong evidence): Chinese Academy of Sciences and Zhejiang University economists ran a preregistered experiment randomising 148 teams to five AI-access schedules, with a formal model, expert ratings and platform logs that trace the idea funnel.

Summary based on the full text.

## In short

An arXiv paper (10 October 2026) by Difang Huang, Shumin Yu, Fadong Chen and Weihua Zhou. When AI makes options cheap, the bottleneck becomes the team's screen, the point where it compares ideas and commits to a few. In a product-design experiment, giving teams the same assistant for the same 15 minutes helped a lot if access came before that screen and did nothing if it came after. Late options skipped comparison, went into the deliverable unscreened and added breadth, not quality. LLM judges missed the timing effect.

## Key numbers

- 0.62 standard deviations: drop in expert-rated creativity when the same 15 minutes of AI access moved from search to selection
- 148: three-person teams (444 participants) randomly assigned to five access schedules
- fivefold: growth of the idea pool at the selection gate with early access (about 4 to about 20 ideas)
- 29% ... 11%: share of selection-stage queries spent generating new options, late-access vs continuous-access teams
- 73%: more proposals submitted by late-access teams than teams without AI, with no gain in rated creativity

## What it reports

- Early access beat late access by 0.36 points (Hedges' g = 0.62, p = 0.018) and beat no AI by 0.49 points (g = 0.83). Late access showed no gain over no AI; continuous access did no better than early access.
- Use rates were similar: 14.9 queries per team early vs 13.8 late.
- Early access grew the pool at the gate from about 4 to about 20 ideas, which teams pruned to fewer than three proposals.
- Only 5% of late-access teams' submitted text existed before the gate, vs 22% without AI and 71% with early access.
- Late-access teams spent 29% of selection queries generating new options (vs 11% with continuous access); late output more often went into submissions nearly verbatim and never appeared in team chat.
- Crediting breadth changes the verdict: adding 0.1 points per proposal makes late access beat no AI. A panel of ten frontier LLM judges detected the benefit of access but not the cost of late timing.

## How they got it

Preregistered lab experiment: 40-minute product-design task with 15-minute search, 15-minute selection and 10-minute review stages. Five conditions: early, late, continuous, none, and early with a warning against overreliance. All used DeepSeek V3.2. Four expert raters scored every proposal; platform logs reconstructed each team's idea funnel. A search-theory model of team decisions generates the predictions and a calibration links pool size to the early-access gain.

## Limits

- Participants were students at one large public university in China, working in short timed sessions.
- One task type (product design) and one model.
- Deadlines and credit-for-breadth rules are set by the experiment and drive part of the mechanism.
- Rated creativity is the main outcome; commercial value is not measured.

## Why it matters for companies

Leaving AI "on" everywhere is a design choice with a cost. The value comes from AI widening the option set before a team decides; options that arrive after the decision tend to slip into deliverables without review. Teams should use AI heavily in early ideation, then route any late AI output through the same screen or restrict late use to developing chosen proposals. Firms using LLMs to judge AI-assisted work should know those judges may reward volume and miss process failures.

Original: https://arxiv.org/abs/2610.14984