Google DeepMind and UK AI Security Institute Expand Partnership

Google DeepMind and the UK AI Security Institute (AISI) have entered into a new Memorandum of Understanding to expand their collaboration from model testing to foundational security and safety research. This partnership aims to provide governments, industry, and society with a scientific understanding of advanced AI risks and their corresponding mitigations.

Foundational Research Collaboration

Google DeepMind is broadening its collaboration with the UK AISI to move beyond initial testing of capable models. The expanded partnership includes the following technical and operational commitments:

  • Proprietary Access: Sharing access to proprietary models, data, and ideas to accelerate research progress.
  • Joint Publications: Producing joint reports and publications to share findings with the broader research community.
  • Collaborative Research: Combining team expertise for deeper security and safety research.
  • Technical Dialogue: Engaging in technical discussions to address complex safety challenges.

Key Technical Research Areas

The joint research efforts focus on three critical areas of the safety landscape:

Monitoring AI Reasoning Processes

Researchers will develop techniques to monitor an AI system’s "thinking" or chain-of-thought (CoT). This work builds on previous Google DeepMind research and collaborations with AISI, OpenAI, and Anthropic. CoT monitoring is intended to complement interpretability research by providing a deeper understanding of how AI systems produce answers.

Understanding Social and Emotional Impacts

The partnership will investigate the ethical implications of "socioaffective misalignment," where AI models may behave in ways that do not align with human wellbeing, even when following instructions correctly. This research builds on existing Google DeepMind work that has helped define this area of AI safety.

Evaluating Economic Systems

Google DeepMind and AISI will explore the potential impact of AI on economic systems by simulating real-world tasks across various environments. These tasks will be validated and scored by experts to be categorized by complexity and representativeness to help predict long-term labor market impacts.

Integrated Safety Strategy

This partnership with the UK AISI is part of a broader safety strategy that includes foresight research, safety training integrated with capability development, and the development of of tools and frameworks to mitigate risk.

Google DeepMind utilizes internal governance through its Responsibility and Safety Council to monitor emerging risks and implement technical and policy mitigations. Additionally, the company partners with external experts such as Apollo Research, Vaultis, and Dreadnode to conduct testing and evaluation of models, including Gemini 3, described as their most intelligent and secure model to date.

Sources