We must regulate AI – but in a way that ensures continued competition
Donald Trump at AI Summit in Washington (White House Photo/Alamy)
4 min read
There are clearly reasons to be alarmed about frontier AI models. In July, a swarm of OpenAI agents managed to get access to the internet and then hacked systems belonging to another technology company, Hugging Face.
The agents collaborated and bypassed safeguards. In a separate case, an AI agent created multiple fake online identities in an attempt to persuade a human reviewer to approve its malicious code.
Despite the outcry, these were not actually near-extinction events, but they show how catastrophe could occur: swarms of AI agents doing unintended things in unexpected ways beyond the control or direction of humans. Since then, political concerns about AI, which had been steadily rising on the back of worries about employment and datacentre energy consumption, suddenly shot up on both sides of the Atlantic. The leaders of several frontier AI companies have now said that brakes need to be applied.
These warnings are genuine, but their preferred solutions are unlikely to be entirely disinterested. As regulation is considered, policymakers should keep in mind the need to ensure competition, not just to tackle risk. Building a tall wall of regulation around the existing frontier companies might solidify their existing market advantage by creating a barrier to new entrants. The rewards from AI are likely to be vast but the trillion-dollar investment gamble is so great that the market might reduce to a small number of dominant players. That is a scenario that policymakers in the UK should strongly seek to avoid. Right now, there are good signals that the AI market is a competitive one for British customers, with the cost of access to frontier models falling recently.
The UK has a potentially pivotal role to play internationally on both competition and safety. Our golden asset is the AI Security Institute (AISI), arguably the world’s leading technical research centre on frontier AI safety. However, it is not a regulator. It cannot compel companies to hand over information or face investigation. That gap feels increasingly hard to justify. When a control failure occurs, we should not depend on what a company chooses to disclose. Governments should be able to demand evidence and establish independently what went wrong, similar to air accident investigations.
Liam Byrne MP, chair of the Business and Trade Committee, has asked AISI to engage more with Parliament. He is right: Parliament needs its expertise. But Parliament should recognise the limits of a research institute operating without statutory powers. AISI’s strength is that the frontier companies respect and trust it, which is why they voluntarily co-operate. Hurriedly giving AISI stronger regulatory powers may seem the obvious answer but there is a risk that layering AISI with new laws makes it harder, not easier, for US companies to collaborate.
Realistically, a British regulator cannot stop US companies experimenting within the US, still less Chinese companies within China. Given AI’s military and intelligence applications, neither country would readily yield access to frontier models. Earlier this year, Anthropic did not share its latest model Mythos 5.1 with AISI, as it has typically done previously, following US export restrictions.
Realistically, a British regulator cannot stop US companies experimenting within the US, still less Chinese companies within China
That does not make domestic legislation pointless. The UK could require the reporting of serious incidents that occur in the UK, even if the investigatory work is fulfilled by counterparts within the US. Ultimately, we need the US and China to regulate their own frontier AI labs: mandatory reporting of serious incidents and independent inspection powers, including prohibition of dangerous autonomous behaviour, with a commitment to warn the rest of the world when things go wrong. That will require international diplomatic pressure more than domestic legislation.
It will be difficult. President Donald Trump appears unwilling to intervene. Nonetheless, thanks to the global reputation of AISI, Britain has a potential leading role. We should use our technical expertise and diplomatic leadership to build a coalition of middle powers to push the US and China towards common minimum standards for dealing with control failures, while also preserving competition.
Britain’s objective should not be safety at the expense of competition, or competition at the expense of control, but a frontier market in which no company is too powerful to challenge and no system too powerful to investigate. Britain cannot achieve that alone.
Theo Bertram is Director of the Social Market Foundation
