The Exceptional Centaur: Why Human–AI Teams Usually Underperform Their Stronger Member, and the Narrow Conditions That Produce Synergy
Abstract: The centaur, a human paired with an artificial intelligence collaborator, is widely assumed to outperform either partner working alone. The empirical record reverses that presumption. A preregistered meta-analysis of 106 experiments demonstrates that human–AI teams perform worse on average than their stronger member, with a pooled deficit of g = −0.23. In chess, where the centaur myth began, assisted teams lost their edge to standalone engines within a decade. Field trials in radiology, software maintenance, consulting, and recruiting reveal that joint performance degrades because trust outruns algorithmic accuracy, while explanations fail to restore calibration. Synergy emerges only under narrow structural conditions, where errors diverge, tasks permit open-ended creation, and workflows reverse scrutiny through automated critics, challengers, or algorithmic triage. Because relying on the machine weakens verification skills while models continue to advance, collaborative synergy is an unstable interregnum rather than a durable architecture of work. Keywords: Centaur Teams, Human-AI Collaboration, Algorithmic Synergy, Collaborative Deficit, Automation Bias, Appropriate Reliance, Jagged Technological Frontier, Calibration of Trust, Algorithmic Triage, Cognitive Forcing Functions, Moral Crumple Zones, Metaknowledge, Deskilling, Verification Skills, Freestyle Chess, Adversarial AI, Lisanne Bainbridge, Garry Kasparov
Authors
- Felix Velasco (ORCID: https://orcid.org/0009-0001-6047-0595)
- Josie Jefferson (ORCID: https://orcid.org/0009-0006-3330-8214)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24254416
- Primary Topic
- Human-Automation Interaction and Safety
- Type
- preprint