Op-ed: AI Act cannot address existential AI risk, international cooperation needed
This op-ed by Mathias K. Hauglid, senior lawyer at the law firm Wikborg Rein, argues that the EU's AI Act is the wrong tool for dealing with fears that AI could destroy humanity. It was prompted by claims from the AI company Anthropic that AI could wipe out humanity, which have revived the question of how potentially deadly AI can be regulated. Hauglid says the AI race was once compared to the space race, but the nuclear arms race has now become the dominant metaphor, at least if the most worried AI researchers are taken seriously.
His core point is that the AI Act applies to AI systems made available on the market. It explicitly exempts systems intended for research and testing, even inside Europe, and AI used for military or national security purposes. So it does not cover the areas where AI is said to pose an existential threat. He compares expecting it to protect against an AI apocalypse to expecting New Year's firework rules to protect against nuclear accidents.
The only legal concept that could help, he writes, is binding international agreements between countries. He sees Norwegian Prime Minister Jonas Gahr Støre and Finnish President Alexander Stubb as on the right track with their international appeal for control over AI development. A disarmament treaty modelled on nuclear weapons treaties is unlikely, he says, because the size of the risk is uncertain and the economic and security gains from AI leadership are too large. Stubb has said the choice is between braking, competing and cooperating, and prefers cooperation. Hauglid argues that once great powers are at the table, competition is unavoidable, which limits how much transparency and mutual reporting, central to the appeal, countries like the US and China will accept.
His proposal is that each state commit to national legislation requiring anyone experimenting with the most powerful AI to have an effective, documented emergency brake. To offer real protection, the rules must also cover states' own activities, including those close to national security and defence. Submitting to independent third-party oversight would signal willingness to cooperate. How such external oversight of potential superintelligence could work in practice, he says, will be one of the big questions ahead.