Caltech
University of Toronto
Princeton University
Saurav KadavathMember of Technical Staff · AnthropicResearcher↗YB
Researcher
Yuntao Bai
AnthropicMember of Technical Staff
Bai helped develop Constitutional AI, which trains assistants to critique and revise their answers using written principles, then learn from AI-generated feedback.[1] He also co-authored Anthropic’s early work on learning from human preferences to make assistants more helpful and less likely to produce harmful responses.[2]
Citation Trend
Total citations to scholarly work, including research papers, technical reports, books and theses, as recorded by Google Scholar. Each point is a dated snapshot.
citations28,961As of Nov 3, '26
citations+7,066Since last snapshot, Jun 15, '26
Selected contributions
Constitutional AI: Harmlessness from AI Feedback
With Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie and others ·Trained AI assistants to critique and improve their answers using a written set of principles.
Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
With Andy L. Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort and others ·Training AI assistants to be more helpful and less harmful using human feedback.
References
- Constitutional AI: Harmlessness from AI Feedbackarxiv.org
Bai is first author; self-critique and revision guided by principles followed by reinforcement learning from AI feedback.
- Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedbackarxiv.org
Bai is first author; human preference modeling and reinforcement learning used to train helpful and harmless assistants.
Based on public sources. If accuracy is important, verify details using the links or references.
For corrections, updates, or requests to remove information, contact us.
Reviewed August 15, 2026 · Checked for accuracy and new information.
