Proximal AI Collaboration Lowers Idea Selection Quality. Implementation Depends on Reliance.

A new experimental study from Groningen and Kozminski lands an awkward finding on the desks of every team racing to embed AI into creative workflows. When 221 employees collaborated tightly with a GPT-4o-based agent - what the researchers call "proximal" collaboration - their idea selection got worse. The effect on implementation was murkier, but not in a reassuring way.

Jesús Mascareño, Burkhard Wörtler, Aleksandra Przegalińska, and Leon Ciechanowski ran a between-subjects experiment in April-May 2023 across organizations in Germany and the Netherlands. One variable was manipulated: how tightly humans worked alongside the AI during an entrepreneurial idea task. High-proximity participants picked lower-quality ideas (b = -0.48, p < .01). The headline isn't that AI hurts creativity. It's that closeness changes what humans do with it.

What "Proximal" Actually Means

Proximal collaboration isn't AI in the next room. It's AI in the loop - generating prompts, claiming equal contribution, telling participants "Let's put our minds together." The low-proximity condition framed the AI as an assistant: "Ask the agent to generate ideas, but you should contribute more." Same model. Same task. Different framing of who's doing what.

The manipulation worked hard: F(1, 215) = 256.88, η²p = .54. Participants felt the difference, and their behavior shifted with it.

Distributed Cognition, Distributed Accountability

The authors lean on Distributed Cognition theory - Hutchins, 1995 - to explain it. Spread thinking and decision-making across a human-AI pair, and two things degrade. Individual accountability dilutes. Attentional focus on the harder evaluative work slips. Idea selection requires sustained engagement. Idea implementation requires turning a choice into concrete steps. Both are vulnerable when cognition gets fragmented across a partner that doesn't carry consequences.

Originality Is a Buffer, Not a Cure

Here's where it gets sharper. The negative effect on idea selection didn't hit equally. It hit hardest when ideas were unoriginal - one SD below the mean, the slope was -1.15. At high originality, the effect disappeared entirely (b = 0.26, n.s.).

Read that again. If your team is generating genuinely novel ideas, proximal AI doesn't hurt selection. If your team is generating decent-but-derivative ideas - which is most teams, most of the time - proximal AI makes selection meaningfully worse.

That's a problem. Because one of the loudest sales pitches for AI tools is that they help with the unsexy middle of the funnel: refining, ranking, choosing among the merely-good. Exactly the zone where this study says they hurt most.

Reliance Helps Implementation. It Doesn't Save It.

The implementation result is the one that doesn't summarize cleanly. The main effect of proximal collaboration on idea implementation wasn't statistically significant (b = -0.31, p = .10). Hypothesis 1b, in the paper's language, was not supported.

But conditional on reliance, the picture changed. Reliance here is concrete: the percentage of AI-generated content participants kept verbatim in their plans. When reliance was low, proximal collaboration tanked implementation hard (b = -1.33). At high reliance, the negative effect shrank but stayed negative (b = -0.60, p < .01). It didn't flip.

The other finding worth holding: implementation was higher overall when reliance was high, regardless of proximity. People who copied more of the AI's plan ended up with more concrete plans. Whether that's a good thing depends on whether you'd rather have a coherent plan you didn't really make, or an incoherent one you did.

What This Doesn't Settle

The study collapses a continuum of AI proximity into two conditions. It uses GPT-4o on a single entrepreneurial task. It runs in two countries with a young convenience sample, mean age 28.8. The authors flag all of this. None of it invalidates the core finding, but the magnitude in your context is unknown.

The sharper question is the one nobody rolling out copilots is asking. If your team's ideas are mostly incremental and you're embedding AI tightly into evaluation, this research says selection quality drops. And if your team relies less on the AI than designers assume - because critical engagement often looks like that - implementation drops too. Will your dashboards catch the difference? Or will you just see "AI adoption: high" and call it a win?

Sources: Mascareño, J., Wörtler, B., Przegalińska, A., & Ciechanowski, L., "When proximal collaboration with AI hinders innovation," Computers in Human Behavior Reports (2026), DOI 10.1016/j.chbr.2026.101054; Hutchins, E., Cognition in the Wild (1995); Sowa, Przegalińska & Ciechanowski (2021) on cobots in knowledge work; Rietzschel et al. (2010) on idea selection; Buçinca et al. (2021) on AI overreliance.