Josh Engels said he left Google DeepMind’s AGI safety team three weeks before announcing on September 12, 2026, that he had joined METR, an organization focused on evaluating advanced AI systems and investigating misalignment. He said the move reflected his concern that frontier-AI capabilities could advance faster than the methods used to align and evaluate them.

Josh Engels moved from Google DeepMind to METR

Engels said he enjoyed his work at Google DeepMind and had turned down offers from Anthropic and OpenAI. He said the stakes surrounding advanced AI had become high enough for him to focus on independent evaluation and alignment research at METR.

Why recursive self-improvement worries Engels

Engels’ central technical concern is recursive self-improvement, a scenario in which an AI system helps create or improve more capable successor systems. Each generation could then contribute to the next, creating a feedback loop in which capability growth becomes harder to evaluate and control.

The safety question is whether alignment methods can keep pace with that process. Alignment refers to the work of making an AI system reliably follow human goals and constraints, including in situations its developers did not anticipate. Engels said researchers do not currently know how to make systems safe enough for recursive self-improvement.

That is a warning about a potential future pathway, not proof that recursive self-improvement is already occurring in public AI systems. Nor does Engels’ five-year horizon establish that catastrophic harm will happen within five years. He said he did not know the exact probability; his point was that he considered the risk high enough to demand more time and stronger evaluation.

What Engels will investigate at METR

Engels said his work at METR will focus on three connected questions:

  • Where does misalignment arise during training?
  • Are current mitigations sufficient?
  • Is alignment research making enough progress to solve the problem?

METR’s work, as described in Engels’ role and the organization’s reported focus, involves evaluating frontier AI systems for autonomous capabilities and potential catastrophic risks. For Engels, that means examining what advanced systems can do, how reliably they follow intended instructions and whether existing safeguards match their capabilities.

He has argued that AI development should be paced so capabilities do not outrun the ability to align models and determine how aligned current systems actually are. That proposal is a call for additional time and evaluation, not a claim that development has already crossed an irreversible threshold.