The Human--AI Collaboration Frontier in Scientific Research

This data is related to the paper The Human--AI Collaboration Frontier in Scientific Research by Markus Leippold, University of Zurich Abstract:Generative AI now assists writing, coding, analysis, and peer review, raising a basicquestion for scientific production: does AI complement human judgment or substitute forit? We formalize this tradeoff as a \emph{Human--AI Collaboration Frontier} in whichworkflow design determines whether AI reliance raises or lowers research quality, withthe two regimes separated by a threshold equal to the human--AI quality gap. We studythis frontier in an ICLR 2026 OpenReview analysis sample of 19,381 submissions and75,800 reviews with AI-content measures for both papers and reviews. Paper AI content ismonotonically associated with lower reviewer-assessed quality, with no evidence of aninterior optimum; component scores show that Soundness declines more steeply thanPresentation, consistent with substitution of fluent drafting for substantiveverification. The evaluation filter does not appear to offset this pattern. Comparingreviews of the \emph{same} paper, AI-flagged reviews are more lenient ($+0.39$ ratingpoints) and less variable (24\% lower variance), consistent with reduced screeningprecision rather than robust targeted favoritism. Decision-stage analyses indicate thatacceptance differences operate primarily through review scores, with limited evidence ofa compensating editorial filter. Because paper AI content is chosen by authors, theproduction-side estimates are interpreted as equilibrium associations, not randomizedcausal effects. The actionable margin is verification-centered workflow design.