Beth Barnes

Timeline

1. By 2022

   Forecasting AI progress

   At DeepMind, she worked with its chief scientist on scaling laws for forecasting deep learning progress. At OpenAI, she helped develop safety targets and evaluated scalable oversight and code models for misalignment before release.

2. 2022

   Building independent evals

   ARC hired her to start ARC Evals, an effort to develop independent evaluations of cutting-edge AI models.

3. March 2023

   Testing autonomous capabilities

   Her update described ARC’s tests of whether models could autonomously gain resources and evade human oversight. Researchers evaluated the models in controlled settings and supervised actions that could be dangerous.

4. December 2023

   METR takes shape

   ARC Evals left its incubation at the Alignment Research Center and became a standalone nonprofit called METR, short for Model Evaluation and Threat Research. She continued to lead the team.

5. July 2024

   Explaining evaluation methods

   In an interview, she described METR’s work as threat modeling and evaluation creation, focused then on capabilities, with future evaluations on control or alignment as possible directions.

6. October 2024

   Scores linked to risk

   At a public talk, she said evaluation scores should reflect how dangerous models are and show when mitigations are needed.

7. November 2024

   Mapping rogue replication

   With Josh Clymer and Hjalmar Wijk, she wrote METR’s Rogue Replication Threat Model, examining how AI agents could acquire resources, evade shutdown, and adapt to novel challenges.

8. As of October 2026

   Monitoring model actions

   She is METR’s founder and CEO, leading a technical team that evaluates generative AI models. In September 2026, she coauthored a research note on a live per-action monitor that pauses evaluations when an action could cause harm and holds flagged actions for human review.