Google DeepMind Expands UK AISI Partnership Into Foundational AI Safety Research

Google DeepMind has signed a new Memorandum of Understanding with the UK AI Security Institute (AISI), expanding a relationship that began at the institute's launch in November 2023 from model testing into foundational security and safety research.

The agreement gives AISI something it has sought since its creation under the UK government: sharing access to Google DeepMind's proprietary models, data and ideas. The partnership also commits both organizations to joint reports and publications, collaborative security and safety research that combines the two teams' expertise, and technical discussions on complex safety problems.

The stakes are straightforward. Google DeepMind argues AI could help treat disease, accelerate scientific discovery, create economic prosperity and tackle climate change, but only if safety and responsibility sit, in the company's words, "at the heart of development." AISI's stated goal is to equip governments, industry and wider society with a scientific understanding of advanced AI risks, along with potential solutions and mitigations. Google DeepMind says it is "deeply committed" to that goal.

Three research areas

The new work focuses on areas where Google DeepMind says its interdisciplinary teams can make systems safer.

Monitoring AI reasoning. The teams will develop techniques to monitor an AI system's "thinking", its chain-of-thought (CoT). This builds on Google DeepMind's own research and on recent joint work with AISI, OpenAI, Anthropic and other partners, including the paper "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety." CoT monitoring shows how a system produces its answers and complements interpretability research.

Social and emotional impacts. The partners will investigate the ethics of socioaffective misalignment, the potential for models to behave in ways that conflict with human wellbeing even while technically following instructions correctly. Google DeepMind says earlier internal work helped define this area of AI safety.

Economic systems. The teams will simulate real-world tasks across different environments to study AI's potential economic impact. Experts will score and validate these tasks, then categorize them along dimensions such as complexity and representativeness to help predict long-term labour market effects.

Broader safety apparatus

The AISI partnership sits inside a larger Google DeepMind safety strategy that includes foresight research, safety training alongside capability development, rigorous model testing, and better tools and frameworks for risk mitigation. The company's Responsibility and Safety Council monitors emerging risk across teams, reviews ethics and safety assessments and implements technical and policy mitigations.

Google DeepMind also works with external evaluators including Apollo Research, Vaultis and Dreadnode to test its models, including Gemini 3, which the company calls its "most intelligent and secure model to date." It is a founding member of the Frontier Model Forum and the Partnership on AI.

The company says it hopes the expanded partnership will produce more robust approaches to AI safety that benefit not only the two organizations but the wider industry and everyone who interacts with AI systems.