Updates in Cooperative AI

Anthropic co-founder Jack Clark puts a >60% probability on AI systems autonomously building better versions of themselves by 2028. If he is right, many important decisions will soon happen inside AI agent workflows that humans cannot review one by one.

This talk asks how democratic accountability works when institutions are run by agents. Today, tools like Polis, Remesh, and the Habermas Machine help citizens turn shared views into policies or commitments. But we lack scalable systems to check whether institutions actually follow them.

That gap is clearest in frontier AI labs, which are likely to automate much of their own work first. Today, some accountability exists because important decisions pass through people: employees can object, leak, testify, be subpoenaed, or leave records in emails and documents. But as work moves into AI agent workflows, the relevant evidence will instead sit in model calls and agent logs.

Michiel argues that we need automated oversight: systems that can read those traces, test specific commitments, produce signed verdicts, and escalate serious violations to humans, regulators, or the public, while preserving privacy and security. The talk will try to make this concrete for frontier labs, then show how the same architecture could apply to governments and companies.

We’re delighted to host this ninth seminar in our 'Updates in Cooperative AI' series. You're welcome to subscribe to our Google Calendar to stay up-to-date on all upcoming events.