Discussion

AI diplomacy needs more than promises: who can share the warning?

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#5528

International AI agreements may falter not for lack of signatures, but because governments have not settled who can obtain, approve and share the information needed to act. Percy Yixuanchen Yu argues in Global Policy Journal that agreements should spell out those responsibilities and test them before an emergency arrives.

Watch Desk analysis

What happened

Yu’s 13 October essay uses a hypothetical software vulnerability that an AI lab discovers: if its government wants to warn another country, who can access the evidence, authorise its release and ensure it reaches someone able to respond? A diplomatic promise alone may answer none of those questions.

Yu proposes that agreements identify who receives a warning, how evidence can be obtained, who approves sharing and which organisations must take part. The commitments should distinguish mandatory duties from voluntary arrangements and say what happens when permission is unavailable. He recommends starting with a bounded task, such as notifying a counterpart about a verified vulnerability affecting a specified class of deployed systems, rather than promising broad access to models or customer records.

The essay also argues that an agreement should be rehearsed against a fictional incident. An exercise could reveal whether the obstacle is missing evidence, missing permission, an unclear threshold or a deliberate refusal to cooperate. Yu points to a joint US–UK pre-deployment evaluation of OpenAI’s o1 model in 2024 as an example of bounded cross-border evaluation, while cautioning that one successful exercise does not guarantee access or permission in other cases.

Why it matters

This is a practical argument about the plumbing behind international AI cooperation. Shared reporting formats can help countries describe incidents consistently, but they cannot grant legal authority to disclose information. Yu’s central point is that negotiators need to match each promise to the people and institutions able to carry it out.

That could make agreements narrower and slower to negotiate. It could also expose gaps before a real incident turns a carefully worded declaration into a very expensive group email.

Our read

Yu makes a useful case for testing the chain of responsibility, not just counting signatories. The proposed exercises cannot guarantee cooperation, and governments may still refuse to share information. But identifying exactly where a process stalls is more useful than discovering, mid-incident, that everyone assumed someone else had permission.

What to watch

  • Whether future AI agreements identify who can obtain, approve and share incident information.
  • Whether governments rehearse protocols and publish what those exercises reveal.
  • Whether shared reporting standards are paired with clear domestic authority and responsibilities.

Discussion spark: Should governments prioritise narrow AI incident-sharing agreements that can be tested in practice, even if that means less ambitious promises?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.