Today, we are announcing an expansion of our partnership with the UK AI Security Institute (AISI) through a new Memorandum of Understanding focused on fundamental security and safety research to ensure that artificial intelligence is developed safely and benefits everyone.
Our research partnership with AISI is a key part of our wider collaboration with the UK Government to accelerate progress in safe and beneficial AI.
Build a foundation for collaboration
AI has immense potential to benefit humanity by helping treat diseases, accelerate scientific discovery, create economic prosperity, and tackle climate change. To realize these benefits, safety and responsibility must be at the center of development. Assessing our models against a wide range of potential risks remains a key part of our safety strategy, and external partnerships are a key element of this work.
That's why, since our launch in November 2023, we have partnered with UK AISI to test our highest performing models. We are deeply committed to UK AISI's goal of providing government, industry and wider society with a scientific understanding of the potential risks posed by advanced AI and potential solutions and mitigations.
We are actively collaborating with AISI to build more robust assessments of AI models, and our teams have collaborated on safety research to advance the field, including recent work on “Thought Chain Observability: Emerging and Vulnerable Opportunities for AI Safety.” Building on this success, we are now expanding our partnership from testing to broader and more fundamental research in a variety of areas.
What the partnership includes
Under this new research partnership, we are expanding our collaboration to include:
- Share access to unique models, data, and ideas to accelerate research progress
- Collaborative reports and publications that share results with the research community
- More collaborative security and safety research combining the expertise of our teams
- Technical discussions to address complex safety challenges
Main research fields
Our collaboration with AISI focuses on key areas where Google DeepMind's expertise, multidisciplinary team, and years of pioneering and responsible research can help make AI systems safer.
Monitoring the AI inference process
We work on techniques to monitor the “thinking” of AI systems, also commonly referred to as Chain of Thought (CoT). This research builds on previous Google DeepMind research and recent collaborations on this topic with AISI, OpenAI, Anthropic, and other partners. CoT monitoring helps understand how AI systems generate answers and complements interpretability research.
Understand the social and emotional impact
We will work together to explore the ethical implications of socio-emotional mismatches. This means that even if instructions are technically followed correctly, an AI model may behave in a way that is inconsistent with human well-being. This research builds on existing Google DeepMind research that helps define this important area of AI safety.
Economic system evaluation
Explore the potential impact of AI on economic systems by simulating real-world tasks across a variety of environments. Experts score and examine these tasks, then categorize them along dimensions such as complexity and representativeness, and predict factors such as long-term labor market impact.
Collaborate to realize the benefits of AI
Our partnership with AISI is one element of how we aim to realize the benefits of AI for humanity while mitigating potential risks. Our broad strategy includes visionary research, extensive safety training coupled with capacity development, rigorous testing of models, and developing better tools and frameworks to understand and reduce risk.
Strong internal governance processes are also essential to safe and responsible AI development, as is collaboration with independent external experts who bring fresh perspectives and diverse expertise to our work. The Google DeepMind Responsibility and Safety Council works across teams to monitor emerging risks, review ethics and safety assessments, and implement relevant technical and policy mitigations. We also partner with other external experts such as Apollo Research, Vaultis, and Dreadnode to conduct extensive testing and evaluation of our models, including Gemini 3, our most intelligent and secure model to date.
Additionally, Google DeepMind is proud to be a founding member of the Frontier Model Forum and Partnership on AI, where we are focused on ensuring the safe and responsible development of frontier AI models and increasing collaboration on important safety issues.
We look forward to our expanded partnership with AISI, allowing us to build a more robust approach to AI safety that will benefit not only our organization, but the broader industry and everyone involved in AI systems.
