Preventing and containing the catastrophic risks of artificial intelligence will inevitably require international cooperation. As models grow more capable and more dangerous, the demand for shared safeguards is becoming more pressing, and it now comes from an ever-widening range of actors: governments, civil society, international organisations, religious leaders, and even the AI companies themselves. What has been missing, however, is concrete proposals that diplomats and policymakers can actually build on.
That is the gap a new project from the Future of Life Institute (FLI) sets out to fill, one to which CeSIA contributed alongside more than 25 specialists in the international governance of AI.
A “model” treaty
This is not, of course, a treaty that is in force. It is a model treaty: a text that sets out in black and white what an international agreement on advanced AI could look like. The aim is to open the discussion. In the end, it is states that will decide which obligations and mechanisms such an agreement should contain.
The model treaty is published on a website, global-governance.ai, which offers two tools. The first, the Treaty Builder, lets anyone put together their own version of the text: you choose from seven broad governance blocks, preview the result, and then download it as a document ready to be put up for debate. The second tool lets you compare existing international initiatives against several criteria, to see what is already in place and what is still missing.
Seven building blocks for an agreement
At its heart, the proposal rests on seven components, each addressing a specific problem:
- Red lines: certain practices should be flatly prohibited at the international level, such as autonomous cyberweapons, helping to develop weapons of mass destruction, or circumventing human control.
- Safety standards: shared technical standards that developers of advanced systems would be expected to meet.
- Technical cooperation: joint safety research, capacity-building, and the shaping of a scientific consensus across countries.
- Emergency response: a coordinated ability to react quickly to a major incident caused by an AI system.
- Benefit-sharing: mechanisms to ensure that the gains from AI reach as many people as possible, without leaving any country behind.
- Institutional arrangements: one or more international institutions responsible for putting these rules into practice.
- Compliance mechanisms: verification and the means to enforce commitments, which is what makes trust between states possible.
CeSIA’s contribution: the red lines
Our own work focused specifically on the red lines. Together with many partners, and under the coordination of José Jaime Villalobos (FLI), we set out to establish which ones belong in a treaty of this kind. The result of that deliberation appears in Annex A of the document.
This work follows directly from the Global Call for AI Red Lines, led by CeSIA. Defining what should be off-limits is not enough if there is no architecture in place to enforce it.