TL;DR

The Commons committee covering business, innovation, science and trade published “Guardrails for Growth: Making AI Safe” on 18 November. MPs want legislation, an “AI Safety and Security Bill”, once the next session opens, plus a new regulator distinct from the AI Security Institute (AISI) with legal powers to pause risky training and stop deployments. They list ten weaknesses in current UK protections.

The committee took evidence from AI researchers, former industry staff and senior people from Anthropic, Google DeepMind, Meta and OpenAI. Its central complaint is that developers set their own safety thresholds and then judge whether their own systems clear them.

What the committee found

AISI gets praise for its testing expertise, but the report says that work rests on goodwill, because frontier AI has no statutory framework. Former employees described commercial pressure and contractual restrictions that can deter staff from raising concerns. Evaluators have no guaranteed access to models, incident reporting is fragmented, and some models cannot be reliably withdrawn once widely copied.

The ten gaps run from missing definitions of unacceptable harm and absent model-access powers to weak liability rules and thin whistleblower protection. The proposed regulator could require independent checks on safety claims, delay or prohibit a release, and pull services whose risks cannot be contained. Material changes to a system’s capabilities or safeguards would trigger a fresh assessment.

Chair Liam Byrne said the UK “cannot give today’s AI leaders unfettered freedom to behave like modern-day Dr Frankensteins”. The report frames safety as an economic question as well, using OECD productivity estimates to put AI’s potential boost to the UK economy at up to £400 billion after ten years.

The G20 thread

The committee also wants ministers to set out, within three months, what the UK wants from its 2027 G20 presidency. Separately, it is seeking evidence on how Britain could broker an international convention covering AI safety and security. That lands two days after the Prime Minister used his Lord Mayor’s Banquet speech to promise a “Manchester Compact on AI” at next year’s G20 leaders’ summit.

This is the first output of the long-term inquiry the committee opened in October.

Looking forward

In our view, the report’s sharpest proposal is institutional: keeping AISI as the technical adviser and creating a separate body to enforce. Until now AISI’s testing has depended on developers choosing to cooperate. The government’s reply to the report will show whether ministers are ready to bring forward a frontier AI bill.