There is no need to slow down the race towards AI in order to control it: a case for oversight modelled on the aviation and arms control sectors

In *Noema*, the auditor Anna August argues that the registration of models, the prompt reporting of incidents and independent audits would be sufficient to make AI transparent.

Rather than betting on a global pause in the development of AI, which she considers out of reach, Anna August, a specialist in AI governance auditing, defends another way in Noema: to make the work of laboratories visible from the outside, without depending on their goodwill. According to her, effective supervision does not require companies to slow down, but that they can no longer operate in opacity.

Two historical precedents

The author relies on the Cold War, where disarmament agreements became credible thanks to reconnaissance satellites, data exchanges and on-site inspections, and not thanks to trust between adversaries. She cites civil aviation in particular: independent investigations, mandatory reporting of incidents, black boxes and information feedback without finding a culprit have made it one of the most secure complex systems in the world.

It draws three proposals: the registration of border models with a regulator before their deployment, the reporting of incidents within a fixed legal period, on the model of the 72 hours already mentioned, and independent audits with access to test environments, not just marketed products. It also reviews a fragmented regulatory landscape, from the ambitious but slow to enforce European AI Act to the American FRONTIER Act.