The UK AI Security Institute tests AI models but does not approve or prohibit their release. Director Henry de Zoete made that distinction at an October 13 parliamentary hearing. Transcripts published on October 19 record demands for earlier scrutiny, company commitments and unresolved government decisions.
Daniel Kokotajlo of the AI Futures Project called for testing before powerful models are used internally for AI research. Aengus Lynch proposed independent evaluations against risk claims set in advance. Lynch identified himself as an external Anthropic collaborator, not a former employee. Steven Adler warned about evaluation limits but rejected the suggestion that such tests were inherently worthless.
Anthropic’s Lord Liam Booth-Smith and OpenAI’s Tom Duff Gordon supported binding statutory rules and incident-reporting duties. Google DeepMind’s Owen Larter emphasized the design of standards. Meta’s Rob Sherman pointed to existing law and a statutory basis for the institute. These positions do not amount to a shared commitment by all four companies to the same new framework.
Model access also remained conditional. Duff Gordon qualified future institute access by reference to US legal requirements. Not every provider committed to embedding government evaluators very early in internal development. Witnesses distinguished notifying affected parties from publicly disclosing technical details.
Government witnesses left new mandatory reporting requirements and their timetable undecided. De Zoete explicitly described the institute as a research organization, not a regulator. The testimony documents statements and commitments, rather than an enacted regulatory settlement or demonstrated effectiveness of the promised safeguards.