PeopleofSuperintelligence

Researcher

Jonathan Uesato

Jonathan Uesato is an AI researcher whose work spans robust machine learning, language-model safety and AI alignment. At DeepMind, he coauthored research on adversarial robustness and specification gaming, then studied toxicity controls and feedback for mathematical reasoning. He left OpenAI in 2024 and later coauthored studies of chain-of-thought faithfulness and alignment faking, while mentoring Anthropic research on reward hacking. As of October 2026, he is a member of technical staff at Anthropic and supports research into reward seeking during reinforcement learning.

Now at
Anthropic

Timeline

  1. 2019

    Robust learning at DeepMind

    In 2019, he works in DeepMind's Robust and Verified Deep Learning group and coauthors a study showing that unlabeled data can be a competitive alternative for training adversarially robust models.

  2. 2020

    Specification gaming

    In 2020, he coauthors Google DeepMind's essay defining specification gaming as meeting an objective's literal specification without achieving its intended outcome.

  3. 2021

    Detoxifying language models

    In 2021, he coauthors research on detoxifying language models. It finds that basic toxicity interventions can improve automatic toxicity metrics while reducing coverage of texts about marginalized groups and their dialects.

  4. 2022

    Process feedback for math

    In 2022, he coauthors a comparison of process- and outcome-based feedback for math word problems. It finds that correct reasoning steps require process-based supervision or learned reward models that emulate it.

  5. 2024

    Leaving OpenAI

    In 2024, he leaves OpenAI as one of several departures from its AGI-safety staff.

  6. January 2025–May 2025

    Reward hacking and faithfulness

    In January 2025, he mentors an Anthropic study showing that training on documents about reward hacking can increase or decrease the behavior. In May 2025, he coauthors a study of chain-of-thought faithfulness; it finds that the reveal rate for hints is often below 20% in many settings.

  7. December 2025

    Alignment faking mitigations

    In December 2025, he coauthors Anthropic research testing black-box interrogation, scratchpad length penalties and process supervision as training-time mitigations for alignment faking.

  8. As of October 2026

    Member of technical staff

    In August 2026, Anthropic thanks him for discussion and support on its study of a reward seeker trained through reinforcement learning. As of October 2026, he is a member of technical staff at Anthropic.

Career

  1. NowAnthropicMember of technical staff
  2. Until 2024OpenAI

From the biography.

  1. 22 of 133 people shaping AI with a former OpenAI role now work at AnthropicIncludes Jonathan Uesato. Our estimates of where OpenAI recruits came from and where 133 people with former OpenAI roles work now, including the companies and labs they founded.
  2. Of 943 people shaping AI who worked at Google, 446 work there nowIncludes Jonathan Uesato. Our estimates of where the 943 people shaping AI who worked at Google and Google DeepMind work now: 446 at Google and 452 across 278 other organizations.
  3. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes Jonathan Uesato. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.
  4. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Anthropic. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.

Sources and credits

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

  1. Identifying and eliminating bugs in learned predictive models deepmind.google. Checked 21 October 2026.
  2. Are Labels Required for Improving Adversarial Robustness? arxiv.org. Checked 21 October 2026.
  3. Specification gaming: the flip side of AI ingenuity deepmind.google. Checked 21 October 2026.
  4. Jonathan Uesato - ACL Anthology aclanthology.org. Checked 21 October 2026.
  5. Solving math word problems with process- and outcome-based feedback arxiv.org. Checked 21 October 2026.
  6. Exodus at OpenAI: Nearly half of AGI safety staffers have left, says former researcher fortune.com. Checked 21 October 2026.
  7. Training on Documents About Reward Hacking Induces Reward Hacking alignment.anthropic.com. Checked 21 October 2026.
  8. Reasoning Models Don't Always Say What They Think arxiv.org. Checked 21 October 2026.
  9. Towards training-time mitigations for alignment faking in RL alignment.anthropic.com. Checked 21 October 2026.
  10. Training a Misaligned Reward Seeker alignment.anthropic.com. Checked 21 October 2026.
  11. Lab Employee Statement on Extreme AI Risks calltolead.org. Checked 21 October 2026.