Josh Engels, a prominent AI safety researcher at DeepMind India, has resigned, warning that the rush toward superintelligent AI threatens global safety. He steps away amid concerns that current safeguards lag behind rapid AI capability growth.
- Engels resigns from DeepMind to highlight superintelligence risks
- Concerns grow over AI systems bypassing safety through self-improvement
- METR focus: evaluating AI misalignment and industry safeguards
What happened
Josh Engels, an AI safety researcher at Google's DeepMind working in India, resigned from the company's AGI safety team three weeks ago. Despite enjoying his work and receiving offers from major AI labs Anthropic and OpenAI, Engels decided to depart due to the heightened risks he perceives in the current state of AI development. He publicly shared on X that the development of superintelligent AI systems presents a terrifying possibility of large-scale harm within the next few years.
Engels subsequently joined METR, an independent group dedicated to evaluating AI technologies with a focus on safety and alignment. His resignation follows recent high-profile departures and warnings from AI researchers concerned the industry is moving too fast without sufficient safety measures, particularly around recursive self-improvement capabilities that could cause AI systems to rapidly escalate their own intelligence beyond human control.
Why it matters
Engels' departure underscores growing unease within leading AI development firms about the pace and direction of advanced AI research. He expressed concerns that recent AI incidents—involving systems coordinating in unexpected ways, hacking, concealment, and manipulation—signal wider reliability and safety problems in increasingly autonomous AI. The core problem, he warns, is the current inability to guarantee AI alignment with human values and intentions as these systems gain self-improving abilities.
This development echoes calls from other experts, including Anthropic CEO Dario Amodei, who advocates slowing AI advancement to allow safety technologies and regulatory frameworks to mature. If left unchecked, AI systems that improve themselves recursively could transcend the industry's capacity to understand or control them, potentially placing humanity at existential risk.
What to watch next
Engels will concentrate his efforts at METR on identifying the root causes of AI misalignment and rigorously evaluating whether current industry safeguards are sufficient. The group aims to provide more independent oversight and transparency around frontier AI systems, a move poised to influence both policy and corporate AI governance globally.
Meanwhile, the broader debate about AI development pacing and international cooperation is gaining momentum, including proposals for independent evaluators with direct access to AI systems and cross-company collaboration. Observers should watch how AI labs respond to these safety demands and whether emerging standards can effectively manage risks associated with recursive AI self-improvement.