PeopleofSuperintelligence

Researcher

Joe Benton

Joe Benton is an AI safety researcher whose work has covered generative modeling and evaluation of AI behavior. He studied mathematics at Cambridge and extended human-feedback methods for reinforcement learning at Berkeley. His Oxford doctoral research included diffusion models, and he coauthored work on flow matching, sabotage risks, and reasoning faithfulness. At Anthropic, he managed the Scalable Oversight team and led research for its Fellows Program. As of October 2026, he plans to join METR, an organization that evaluates AI companies’ systems.

Works
22
Citations
84
h-index
6

Publication numbers from OpenAlex, as of 15 October 2026.

Timeline

  1. 2017–2021

    Mathematics at Cambridge

    He earns a BA with MMath in Mathematics at Trinity College, Cambridge, and graduates with distinction in 2021.

  2. 2021

    Research at Berkeley CHAI

    He works as a research assistant at UC Berkeley’s Center for Human-Compatible Artificial Intelligence in 2021 and builds on PAIRED to use human feedback to reduce the environment space and improve training speed. His CHAI project aims to make reinforcement-learning agents more robust by having them request extra information when needed.

  3. 2021

    Oxford diffusion-model research

    In 2021, he begins a DPhil in Statistics at Oxford, with a first project on denoising diffusion models.

  4. May 2023

    Error bounds for flow matching

    In 2023, he coauthors a paper that gives error bounds for fully deterministic flow matching under assumptions about approximation error and data regularity.

  5. October 2024

    Testing sabotage risks

    In 2024, he coauthors evaluations designed to test whether frontier models could sabotage checks of their capabilities, behavior monitoring, or deployment decisions.

  6. May 2025

    Testing reasoning faithfulness

    In 2025, he coauthors a study finding that reasoning models often fail to reveal when prompt hints shape their chain-of-thought; outcome-based training raises faithfulness before it plateaus.

  7. By 2026

    Leading Anthropic’s oversight team

    Before leaving Anthropic in 2026, he manages its Scalable Oversight team and leads research for its Fellows Program.

  8. As of October 2026

    Preparing to join METR

    In 2026, he plans to join METR, an independent organization that evaluates how safe AI companies’ systems are.

Sources and credits

Photo: We have no published photo of Joe Benton yet. Send us a free-licensed photo.

  1. Curriculum Vitae - Joe Benton joejbenton.com. Checked 14 October 2026.
  2. Joe Benton – Center for Human-Compatible Artificial Intelligence chai.berkeley.edu. Checked 14 October 2026.
  3. [2305.16860] Error Bounds for Flow Matching Methods arxiv.org. Checked 14 October 2026.
  4. [2410.21514] Sabotage Evaluations for Frontier Models arxiv.org. Checked 14 October 2026.
  5. [2505.05410] Reasoning Models Don’t Always Say What They Think arxiv.org. Checked 14 October 2026.
  6. Joe Benton joejbenton.com. Checked 14 October 2026.
  7. Why I left Anthropic’s safety team to hold AI companies accountable jbenton1.substack.com. Checked 14 October 2026.