Rohin Shah is Director of AGI Safety & Alignment at Google DeepMind, where he leads the AGI Safety & Alignment team. With Anca Dragan he wrote "The case for reasoning transparency," one of the five inaugural essays of the DeepMind Institute, published September 16, 2026 (The Case for Reasoning Transparency (Shah and Dragan, DeepMind Institute, September 2026)).
Background and role
According to his DeepMind Institute biography, Shah completed his PhD at the Center for Human-Compatible AI at UC Berkeley, where he worked on building AI systems that can learn to assist a human user "even if they don't initially know what the user wants." At Google DeepMind the team he leads aims to improve AI safety by ensuring that, as capabilities scale up, AI systems increasingly "try" to do what their designers intended; the biography lists his areas of work as amplified oversight, interpretability, frontier safety and externalized reasoning (The Case for Reasoning Transparency (Shah and Dragan, DeepMind Institute, September 2026)).
Research
Shah is one of the 41 authors of "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety" (Korbak et al., July 2025), which recommended that frontier developers "consider the impact of development decisions on CoT monitorability" (Source: arxiv.org). With Jonah Brown-Cohen and David Lindner he wrote "Quantifying the Necessity of Chain of Thought through Opaque Serial Depth" (arXiv, March 2026), which defines opaque serial depth and computes upper bounds on it for Gemma 3 models (Source: arxiv.org).
He is also a co-author, with Anson Ho, Jean-Stanislas Denain and David Atanasov of Epoch AI and Samuel Albanie of Google DeepMind, of "A Rosetta Stone for AI Benchmarks," a Google DeepMind-funded statistical framework for placing different benchmarks on a single capability scale (A Rosetta Stone for AI Benchmarks — Ho, Denain, Atanasov, Albanie, Shah (Epoch AI / Google DeepMind, 2025)). He is among those thanked for feedback in Hubinger et al.'s 2019 paper on mesa-optimization (Risks from Learned Optimization in Advanced Machine Learning Systems (Hubinger et al., 2019)).
Positions
In the September 2026 essay, Shah and Dragan argue that the legibility of reasoning models' chain of thought is a valuable but fragile safety affordance, citing the GPT-6 Astra system card's reported decrease in chain-of-thought monitorability as evidence of "a concerning downward trend." They propose measuring chain-of-thought transparency, preserving transparent architectures through a limit on opaque serial depth that "regulations or AI developers themselves could set," and auditing training rewards so that the chain of thought is not optimized to look aligned. They propose that developers who deviate from standard chain of thought demonstrate that their methods are "comparably monitorable," and state that regulation could help if commercial goals under-incentivize transparency (The Case for Reasoning Transparency (Shah and Dragan, DeepMind Institute, September 2026)). See Chain-of-Thought Monitorability.
Relationships
- related: Google DeepMind — Director, AGI Safety & Alignment
- related: The Case for Reasoning Transparency (Shah and Dragan, DeepMind Institute, September 2026) — co-author
- related: DeepMind Institute — inaugural contributor
- related: Anca Dragan — co-author and Google DeepMind safety leadership colleague
- related: Chain-of-Thought Monitorability, Opaque Serial Depth — principal research subjects
- related: A Rosetta Stone for AI Benchmarks — Ho, Denain, Atanasov, Albanie, Shah (Epoch AI / Google DeepMind, 2025) — co-author