Deepening our partnership with the UK AI Security Institute

Source: DeepMind Date: 2025-12-11 URL: https://deepmind.google/blog/deepening-our-partnership-with-the-uk-ai-security-institute/

Summary

Google DeepMind and the UK AI Security Institute (AISI) signed an expanded Memorandum of Understanding covering three research areas: chain-of-thought reasoning monitoring, socioaffective misalignment (AI that follows instructions correctly but harms user wellbeing), and AI economic impact simulation. Commitments include sharing proprietary models and data, joint publications, and collaborative safety research. The MOU follows a prior joint project with OpenAI and Anthropic on reasoning transparency.

Implications

“Socioaffective misalignment” as a named research category is the definitional move. This is Google formally acknowledging that instruction-following is insufficient as a safety criterion — an AI can follow every instruction correctly and still damage wellbeing through tone, dependency-creation, or manipulative persuasion. Naming it and funding a research program is the step that makes it a measurable safety domain rather than a philosophical complaint.

Chain-of-thought monitoring as a research track signals that CoT transparency is not yet solved. The joint OpenAI/Anthropic/DeepMind collaboration on reasoning transparency already exists. Adding AISI as a government partner on the same problem means: (1) governments want independent access to reasoning audit techniques, not just lab self-reporting, and (2) the labs believe government-accessible monitoring tools are the right political posture before mandatory evaluation regimes arrive.

Economic impact simulation is the labor market risk research that precedes policy. The EU AI Act requires labor impact assessments for certain AI deployments. This research track positions AISI and DeepMind to generate the methodology for such assessments before regulation mandates it — a standard regulatory engagement pattern (shape the methodology before the mandate shapes you).

Watch:

• Joint publications emerging from the CoT monitoring collaboration — does the technical output change how reasoning transparency is evaluated by external auditors?
• Whether AISI’s socioaffective misalignment framework influences EU AI Act implementing regulations or UK AI legislation
• How this MOU compares in scope to similar partnerships with Anthropic and OpenAI — are the research tracks identical or differentiated?