Brief IA
›

British MPs Propose to Suspend Dangerous AIs

Regulation & Ethics·Tom Levy·

British MPs Propose to Suspend Dangerous AIs

British MPs Propose to Suspend Dangerous AIs
Key Takeaways
1A British committee is calling for intervention powers as soon as AI models are developed.
2The AI Safety Institute is already evaluating models, without the power to prohibit their dissemination.
3The government has not committed to the requested legislation.
Why it matters : the proposed measure would remove the sole decision-making power from companies regarding whether to continue training or deploying models deemed dangerous.

In the United Kingdom, a parliamentary committee is calling for a law that would empower a regulator to halt the development or dissemination of dangerous AI, starting in the next parliamentary session. This non-binding proposal has yet to receive the government's commitment.

The AISI Tests Without the Power to Ban Dissemination

The UK's AI Safety Institute (AISI) is already evaluating some of the most powerful systems in the world. However, it relies on the voluntary cooperation of companies and has no veto power over their dissemination, even in the presence of safety concerns. MPs propose to maintain its evaluation mission while a separate authority would enforce the rules.

For systems classified as high-risk, developers would need to provide evidence supporting their safety assurances and allow for examination by independent researchers. Any significant changes to the model's capabilities or protections would require additional evaluation.

Control would also involve employees. Former staff have testified before MPs about the potential costs of whistleblowing: contractual constraints, commercial pressures, or financial repercussions can, according to them, discourage speaking out. The committee is calling for enhanced legal protection for these individuals, as well as a requirement to report incidents such as near-misses.

Companies Should Accept External Oversight

OpenAI, Google DeepMind, Anthropic, and Meta have all assured parliamentarians that they would withhold models they deem dangerous. However, their safety criteria differ, and UK law generally does not impose any prior external expertise before dissemination. These companies have been heard, along with researchers and former employees, by the UK Parliament's Committee on Business, Innovation, Science, and Trade.

In its recommendations titled "Guardrails for Growth: Making AI Safe," the committee identifies ten shortcomings in the national framework. It specifically mentions the lack of mandatory incident reporting, restricted access for external evaluators, and insufficient protection for whistleblowers. Its chairman, MP Liam Byrne, warns of the risk of allowing companies to create unchecked dangers that could become uncontrollable.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

The desired law would enable action before a model is made available: halting training or testing that presents a high risk, prohibiting deployment, or even suspending service if the danger cannot be managed. Byrne demands that companies justify their assurances before independent experts and that an authority be able to intervene. He emphasizes that some models, once disseminated, cannot be reliably recalled and calls for protection from the design and testing stages.

Language models contain billions of learned parameters; their designers do not always know how they will behave in every situation. Therefore, MPs refuse to allow OpenAI or Anthropic to set acceptable limits on their own. The government also seeks regulation that does not rely solely on industrial assurances, similar to the aviation, pharmaceutical, or nuclear safety sectors.

The Committee Estimates UK Economic Potential

The committee estimates that AI could add up to £400 billion to the UK economy over the next decade, based on OECD productivity estimates. It believes that more stringent guarantees could encourage companies, particularly smaller ones, to adopt these tools. It also envisions economic opportunities in the independent evaluation of models, leveraging the expertise of the AISI.

For Liam Byrne, the decision on acceptable risks for the public rests with ministers, who are accountable to Parliament. While the committee is calling for a vote in the next parliamentary session, its recommendations are not binding on the government. The government has not committed to the proposed text or to the power to halt the development of systems.

Parliamentarians are also advocating for an international approach. They wish for the UK's G20 presidency, scheduled for 2027, to be used to persuade other countries to tighten their requirements regarding evaluations, incident reporting, and the dissemination of high-risk models.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.