Research map / Large language models / AI and society
Position: AI Should Not Be An Imitation Game: Centaur Evaluations
Andreas Haupt, Erik Brynjolfsson
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing ML should increasingly use centaur evaluations, where humans and AI jointly solve tasks, to promote human augmentation and capture human-centered desiderata.
Problem
Current ML evaluations test systems in isolation from humans, leading to saturated benchmarks, poor proxies for human-centered desiderata, and a development bias toward human replacement rather than augmentation.
Approach
Defines centaur benchmarks with three components (human selection, interface, scoring) and argues for jointly human-AI evaluations, including a production-function framing in which the marginal value of human time measures human augmentation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine a human with an AI system for coding | human-AI coding team · functional HTTP server coding | speed | Improved |
| provide machine-learning-powered coding recommendations to coders | coders using Visual Studio Code · coding in Visual Studio Code | speed | Improved |
| equip workers with a machine-learning system for support | human workers · work tasks studied by Cui et al. (2024) | productivity | Improved |
| combine human decisions with machine-learning predictions | human-AI systems · several socially relevant decision settings | performance | Hurt |
| use human judges rather than counterfactual algorithm-only decisions | judges and algorithms · judicial decisions | performance | Hurt |
| combine radiologists with screening algorithms | radiologists and screening algorithms · radiology screening | screening performance | Hurt |
| combine humans with AI rather than use algorithms alone | human-AI systems · fairness evaluations | fairness | Hurt |
| use human-AI collaboration instead of human-only or AI-only systems | human-AI systems · studies published in 2020 to 2023 | human augmentation | No clear effect |
Limitations
- Centaur evaluations are expensive and may cost rapid model iteration
- Human-AI studies can be noisy, brittle, and require hundreds to thousands of human participants
- Human augmentation does not always hold; centaurs sometimes underperform humans or AI alone
Experimental setup
Builds on: Turing Test Turing Trap (Brynjolfsson 2022) learning-to-defer
Closest papers by meaning
- Position: Challenges and Future Directions of Data-Centric AI Alignment (ICML 2025 · similarity 0.81)
- Position: AI Evaluation Should Learn from How We Test Humans (ICML 2025 · similarity 0.81)
- Position: Human Baselines in Model Evaluations Need Rigor and Transparency (With Recommendations & Reporting Checklist) (ICML 2025 · similarity 0.80)
- AGI-Elo: How Far Are We From Mastering A Task? (NeurIPS 2025 · similarity 0.80)
- Confounding-Robust Deferral Policy Learning (AAAI 2025 · similarity 0.80)
- RE-Bench: Evaluating Frontier AI R&D Capabilities of Language Model Agents against Human Experts (ICML 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
