Deepening our partnership with the UK AI Security Institute
Google DeepMind ● Covered by 2 sources
Google DeepMind is deepening its partnership with the UK AI Security Institute, moving from just model testing to joint research. They'll dig into chain-of-thought monitoring, AI's emotional impact, and economic effects together.
Based on reporting by Google DeepMind — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind and the UK AI Security Institute have been working together since AISI launched back in November 2023, mostly on testing DeepMind's models for risks. Now they're signing a new Memorandum of Understanding that pushes the relationship well past testing. DeepMind will start sharing proprietary models, data, and research ideas with AISI, and the two will co-author reports meant for the wider research community rather than just internal use.
Three research areas anchor the new arrangement. First is chain-of-thought monitoring, watching how AI systems reason step by step, which DeepMind has already explored with AISI alongside OpenAI and Anthropic in a joint paper on what they call a fragile but valuable window into AI safety. Second is what DeepMind terms socioaffective misalignment, the unsettling possibility that a model can follow instructions to the letter while still acting in ways that undermine human wellbeing. Third is economic simulation: building out realistic task environments, having experts score and sort them by complexity and relevance, and using that to get ahead of questions about how AI reshapes labor markets.
DeepMind frames this as one piece of a much larger safety apparatus. There's an internal Responsibility and Safety Council that reviews ethics assessments across teams, plus outside evaluators like Apollo Research, Vaultis, and Dreadnode who've already put Gemini 3 through its paces. The company also points to its membership in the Frontier Model Forum and Partnership on AI as evidence it's not operating in a vacuum.
What stands out is the shift from arm's-length auditing to something closer to co-development. Handing a government-backed institute access to proprietary models and internal research ideas is a bigger ask than letting them red-team a finished product. It signals DeepMind wants AISI's fingerprints on the research itself, not just a stamp of approval after the fact — a bet that closer collaboration produces sturdier safety science than testing alone ever could.
My take — AI-written commentary, not fact-checked reporting
This is DeepMind doing the smart PR-meets-substance move: give a respected government body real access before regulators start asking why they didn't. I'd rather see labs pull outside institutes into the actual research than treat safety as a compliance checkbox after the model ships, and the UK, lacking the EU's regulatory hammer or the US's political chaos, has quietly built AISI into the most credible independent voice in this space. Still, sharing proprietary models with one government-funded body while competitors keep theirs locked down raises an obvious question about how 'independent' any of this stays once access becomes a bargaining chip.
Read more about this at: Google DeepMind