ABOUT OUR GUEST
Dr. Rosa Lim
Research Director, Centre for AI Governance · Oxford
Rosa studies multi-stakeholder coordination in emerging technology governance. She previously advised the UN Secretary-General’s Roadmap on Digital Cooperation and has published widely on institutional design for AI oversight.
We’ve been framing AI safety as a technical challenge. But the harder question is whether competing institutions can agree on the same goal at all.
In this month’s guest curation Dr. Lim dives into this question and gets to the center of why AI safety is critical.
You can solve the technical problem perfectly and still end up with a catastrophically misaligned outcome — if the labs, governments, and users all have different objectives.”
— Dr. Rosa Lim
Question 1
What we’re looking at is a classic multi-player coordination game with asymmetric information and misaligned incentives. Each actor does best by defecting (moving faster, setting lower safety bars) as long as they believe others will do the same. The tragedy is that everyone knows this — and it changes nothing.
What we’re looking at is a classic multi-player coordination game with asymmetric information and misaligned incentives. Each actor does best by defecting (moving faster, setting lower safety bars) as long as they believe others will do the same. The tragedy is that everyone knows this — and it changes nothing.
What we’re looking at is a classic multi-player coordination game with asymmetric information and misaligned incentives. Each actor does best by defecting (moving faster, setting lower safety bars) as long as they believe others will do the same. The tragedy is that everyone knows this — and it changes nothing.
Consider what happened during the 2024 wave of frontier model deployments. Each lab had its own red-teaming standards, its own definition of “sufficiently safe,” and its own timeline pressures. None of them were acting in bad faith. But the aggregate result was a race to a threshold nobody had collectively agreed on.
Question 2
What we’re looking at is a classic multi-player coordination game with asymmetric information and misaligned incentives. Each actor does best by defecting (moving faster, setting lower safety bars) as long as they believe others will do the same. The tragedy is that everyone knows this — and it changes nothing.
What we’re looking at is a classic multi-player coordination game with asymmetric information and misaligned incentives. Each actor does best by defecting (moving faster, setting lower safety bars) as long as they believe others will do the same. The tragedy is that everyone knows this — and it changes nothing.
What we’re looking at is a classic multi-player coordination game with asymmetric information and misaligned incentives. Each actor does best by defecting (moving faster, setting lower safety bars) as long as they believe others will do the same. The tragedy is that everyone knows this — and it changes nothing.
Consider what happened during the 2024 wave of frontier model deployments. Each lab had its own red-teaming standards, its own definition of “sufficiently safe,” and its own timeline pressures. None of them were acting in bad faith. But the aggregate result was a race to a threshold nobody had collectively agreed on.
of LMIC ministries report stalled AI curriculum rollouts 62%
governments proposing joint cross-border AI safety frameworks 3
ideas for narrowing the learning divide identified by the Observatory 18
We asked two experts to respond to Rosa’s observations
Tolu Oke
EU Tech Policy Office, Brussels
Rosa is right that the framing matters. The EU AI Act implicitly assumes a technical compliance model — but if the underlying problem is coordination, compliance frameworks won’t get us there. We need binding multilateral agreements, not just audits.
Marco Ricci
Economist, CEPR
I’d push back slightly. The coordination framing underestimates how much technical progress on interpretability changes the incentive landscape. Once you can actually verify alignment properties, the game theory shifts considerably.
Hear From More of Our Guests
3Mpower: Mobile Learning for Empowerments of Marginalised Mathematics Educators
The aim of 3MPower was to address the global evidence gap on the use of EdTech for teacher professional development (TPD), specifically in relation to the needs and experiences of…
A Strategic Toolkit for EdTech Policy and Decision-Making
Developing policy in the real world rarely follows a straight line. It unfolds amid ambiguity, competing agendas, and external disruptions like leadership changes or shifting public moods. This toolkit was…
Funding statement.
