AI
Google DeepMind expands safety partnership with UK AI Security Institute
The tech giant deepens its collaboration with the UK's AI safety regulator through a new research agreement focused on model testing and foundational security work.
Google DeepMind Blog
Google DeepMind has formalized a deepened partnership with the UK AI Security Institute (AISI) via a new Memorandum of Understanding. The agreement extends their existing collaboration beyond testing to encompass broader foundational research aimed at developing safer, more responsible AI systems.
Since AISI's launch in late 2023, DeepMind has worked with the UK institute to evaluate its most advanced models against potential risks. The expanded arrangement signals a commitment to embedding safety considerations throughout AI development.
Research focus areas#
The partnership will tackle three major safety challenges:
Chain-of-thought monitoring — DeepMind and AISI will develop methods to examine how AI systems reason and generate answers. Building on earlier DeepMind research and a recent joint study conducted with OpenAI, Anthropic and others, this work aims to make AI decision-making more transparent and interpretable.
Human-AI alignment beyond instruction-following — The collaboration will investigate what happens when AI systems technically follow their instructions correctly but still produce outcomes misaligned with human wellbeing and values. DeepMind's prior work in this space will inform the research agenda.
Economic impact simulation — Researchers will model how AI affects economic systems by running real-world scenarios across varied conditions. Domain experts will score and categorize these simulations to better forecast long-term labour market consequences.
How the partnership works#
The arrangement involves granting AISI access to DeepMind's proprietary models, datasets, and research insights. The teams will issue joint publications, conduct collaborative safety investigations, and hold technical working sessions to address complex safety problems.
Broader safety strategy#
This partnership sits within DeepMind's wider approach to responsible AI development, which includes forward-looking research, integrated safety training, rigorous model evaluation, and governance structures like its Responsibility and Safety Council. DeepMind also works with external partners—Apollo Research, Vaultis, Dreadnode and others—to independently test models, and is a founding member of both the and the .
