Additional Resources

Research Output

Working Papers

  • Do AI Referees See What Human Referees See? Evidence from Economics Journal Submissions

    This paper studies whether AI referee reports identify the substantive concerns raised by human referees in economics journal review. I compare one frozen AI referee report with the anonymized human referee reports written for the same submission-stage manuscript in 42 donated manuscript-review bundles, containing 108 human reports. A source-blind measurement pipeline turns every report into an issue inventory, merges concerns that recur across human referees, and aligns the AI report to that human benchmark. Across six independent measurement runs, the evaluated report recovers 0.377 of recurring human concerns and 0.199 of all human-raised concerns on a case-weighted basis. Among the 23 bundles eligible for the human yardstick, AI report recovery is 0.426, while the case-weighted human yardstick is 0.361; the mean run-level ratio is 1.200. The results suggest that AI feedback can reproduce part of the shared referee signal early and cheaply, making it useful for pre-submission revision and as a complement to peer review, while still missing many recurring concerns and leaving human expertise essential.

  • Delegate Pricing Decisions to an Algorithm? Experimental Evidence , with Hans-Theo Normann, Nina Rulié, and Tobias Werner.

    In a repeated Bertrand laboratory experiment, participants are assigned to one of three treatments: a baseline with manual pricing only, an outsourcing treatment where they can delegate pricing to a Q-learning algorithm, or a recommendation treatment where they receive overrideable algorithmic price suggestions. Delegation rates are substantial and higher when override is available. However, endogenous adoption combined with human override and undercutting erodes prices over time: by the final supergame, both the outsourcing and recommendation treatments exhibit lower price levels than the no-algorithm baseline, where market prices increase across supergames due to learning.

  • Sequential Selling of Personal Data: Experimental Evidence from Facebook, Instagram, and TikTok , with Jannika Schad and Benjamin Schröder.

    In an incentive-compatible online experiment we elicit willingness to sell and reservation prices for selling complete GDPR data exports from Facebook, Instagram, and TikTok across up to three sequential rounds. Treatments vary buyer identity (academic researchers vs. commercial firms) and buyer continuity (single vs. multiple acquirers). The initial disclosure decision strongly predicts subsequent selling behavior. Within subjects, willingness to sell increases slightly across rounds while reservation prices decline, with the effect being driven by participants who were willing to sell in the first round.

  • Determinants of Wastewater Charges in North Rhine-Westphalia , with Justus Haucap.