PeopleofSuperintelligence

Researcher

Erik Jones

Erik Jones researches ways to audit language models and reduce misuse. He earned degrees in mathematics and computer science at Stanford before completing a computer-science PhD at Berkeley. His work used cognitive biases and discrete optimization to find failures, explored small models’ reasoning through Orca 2, and used synthetic tasks to reduce hallucinations. In 2025, he served on Anthropic’s Safeguards Research Team; as of October 2026, he is a safety researcher at Anthropic, coauthoring work on Claude’s conversation monitoring.

Timeline

  1. 2022

    Cognitive biases expose model failures

    In 2022, Jones and Jacob Steinhardt publish “Capturing Failures of Large Language Models via Human Cognitive Biases.” The paper uses cognitive biases to form and test hypotheses about recurring failures in model behavior.

  2. March 2023

    Discrete optimization for audits

    In March 2023, he coauthors “Automatically Auditing Large Language Models via Discrete Optimization.” Its ARCA algorithm searches prompts and outputs for target behaviors to expose unexpected failures before deployment.

  3. October 2023

    Synthetic tasks for hallucinations

    In October 2023, he coauthors “Teaching Language Models to Hallucinate Less with Synthetic Tasks.” SynTra tunes a model’s system message on a synthetic task, then reduces hallucinations on real summarization tasks.

  4. November 2023

    Orca 2 teaches reasoning

    In November 2023, he coauthors Orca 2, a Microsoft Research paper on training small language models to choose reasoning strategies. The paper reports performance close to models five to ten times larger on complex reasoning tests.

  5. February 2025–April 2025

    Anthropic’s safeguards research team

    In February 2025, he is a member of Anthropic’s new Safeguards Research Team, which works on jailbreak robustness, automated red-teaming, and monitoring for misuse and misalignment. By April 2025, he is a researcher at Anthropic focused on AI safety and robustness in large language models.

  6. May 2025

    Scalable AI safety audits

    Before Berkeley, he earns a B.S. in math and an M.S. in computer science at Stanford. In May 2025, his Berkeley PhD dissertation, “Scalable Auditing for AI Safety,” develops evaluation systems including ARCA, MultiMon, and TED to uncover rare and unexpected failures; Anca Dragan and Jacob Steinhardt advise the work.

  7. As of October 2026

    Monitoring Claude for misuse

    As of October 2026, he is a safety researcher at Anthropic. In October 2026, he coauthors “Scaling Misuse Detection with Self-Assessment,” a Safeguards post that introduces a method for Claude to assess each conversation for potential misuse after it responds.

  1. Google had the most people shaping AI every year from 2014 to 2025Our estimates of which organizations employed the most people shaping AI from 1961 to 2025, with a year-by-year chart.
  2. The United States employs an estimated 53% of people in our data with a known employment countryOur estimates of where AI careers are concentrated now and how countries’ shares of dated positions changed over time.
  3. 46% of living people in our data with known birth and employment countries work outside their birth countryOur estimate of cross-border careers in AI, from country of birth and the country of each current organization.
  4. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionOur estimates of where the people shaping AI studied and where their careers led, from education records in cited biographies.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

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

  1. Capturing Failures of Large Language Models via Human Cognitive Biases arxiv.org. Checked 19 October 2026.
  2. Automatically Auditing Large Language Models via Discrete Optimization arxiv.org. Checked 19 October 2026.
  3. Teaching Language Models to Hallucinate Less with Synthetic Tasks arxiv.org. Checked 19 October 2026.
  4. Orca - Microsoft Research: Publications microsoft.com. Checked 19 October 2026.
  5. Orca 2: Teaching Small Language Models How to Reason arxiv.org. Checked 19 October 2026.
  6. Introducing Anthropic's Safeguards Research Team red.anthropic.com. Checked 19 October 2026.
  7. Erik JONES | Singapore Conference on AI scai.gov.sg. Checked 19 October 2026.
  8. Scalable Auditing for AI Safety | EECS at UC Berkeley www2.eecs.berkeley.edu. Checked 19 October 2026.
  9. Erik Jones people.eecs.berkeley.edu. Checked 19 October 2026.
  10. Scaling Misuse Detection with Self-Assessment safeguards.anthropic.com. Checked 19 October 2026.