UK and US AI Safety Institutes sign testing partnership

The UK and US governments signed a Memorandum of Understanding committing their respective AI Safety Institutes to work together on testing frontier AI models, the first bilateral agreement of its kind. UK science secretary Michelle Donelan and US commerce secretary Gina Raimondo signed the pact, which took effect immediately.

The agreement committed the two institutes — the UK AI Safety Institute and its newly formed US counterpart — to develop a shared, “interoperable” approach to evaluating AI models, to carry out at least one joint testing exercise on a publicly accessible model, to collaborate on technical safety research aimed at advancing the science of frontier-model evaluation, and to explore exchanging personnel between the two organisations. It followed directly from commitments both governments had made at the Bletchley Park AI Safety Summit in November 2023, where the UK had launched its institute and the US had announced plans for one.

The partnership addressed a practical problem facing government AI-safety testing: no single national body had the staff or compute to evaluate every frontier model release on its own, and divergent testing methodologies between countries risked producing inconsistent judgments about the same model. By agreeing to build compatible evaluation approaches and to run joint exercises, the two governments aimed to make it possible for either country’s testing to inform the other’s, rather than duplicating the work independently — a template other governments’ safety institutes referenced in the discussions that followed on international coordination of AI evaluation standards.

As with the summit commitments it stemmed from, the memorandum created no binding testing requirement on AI developers themselves; both institutes continued to rely on voluntary access arrangements with labs including OpenAI, Anthropic and Google DeepMind to obtain pre-deployment access to models for evaluation.