Yoshua Bengio wants frontier AI treated like nuclear power: license it before it ships.
Speaking in New York today, Bengio urged the UN Security Council to require licenses for the most powerful AI systems, comparing the idea to oversight regimes already used for medicine, aviation, and nuclear energy. Bengio co-chairs the UN's scientific panel on AI, which gives the proposal more institutional weight than a typical open letter. His core complaint is simple: the companies building frontier models have no convincing answer to the risks those systems pose. He's taking that case straight to the body responsible for international peace and security, not to a standards group or a national regulator.
Nuclear power and aviation are licensed because failure in those industries is catastrophic and hard to reverse, not because operators asked for the paperwork. Applying that same framework to AI is a pointed statement that voluntary safety commitments and internal evals from labs aren't sufficient. Routing the ask through the Security Council also puts it above any single government, which matters given how inconsistent national AI rules have been so far.
Whether a body built for wars and sanctions is the right one to gatekeep software development is a separate question Bengio's pitch doesn't quite answer.