The concept comes from the political scientist Peter Haas, who studied how a transnational network of atmospheric scientists shaped the response to ozone depletion. An epistemic community is a network of professionals with recognised expertise and a shared set of beliefs about a problem — its causes, its stakes, and the policies that would address it. When such a community achieves consensus and gains the ear of decision-makers, it can supply the shared factual basis that makes international agreement possible among states that otherwise trust nothing the others say.
How expert communities move treaties
The mechanism is not that experts have formal power; they do not. Their influence comes from providing something governments need under uncertainty: a credible, shared account of a complex problem. When facing a novel technical threat, policymakers lack the in-house knowledge to define the problem or evaluate solutions, and an epistemic community fills that gap. Crucially, because the community is transnational and bound by professional standards rather than national interest, the understanding it provides can be accepted across borders — giving rival governments a common starting point that national experts, suspected of serving their governments, cannot.
- They frame the problem. By establishing what the threat is and how serious it is, experts shape the agenda governments respond to.
- They build cross-border consensus. A shared professional understanding that transcends national lines gives distrustful states a common factual basis to negotiate from.
- They design the options. Experts translate a diagnosis into concrete, workable policy proposals that diplomats can then bargain over.
- They legitimise action. When a credible expert consensus says a threat is real and manageable, it becomes harder for governments to justify inaction.
The precedents
The ozone regime is the classic case: atmospheric scientists reached consensus on the mechanism and danger of ozone depletion, and that consensus, carried into the policy process, made the Montreal Protocol possible. Nuclear arms control had the Pugwash movement — scientists from both Cold War blocs who maintained dialogue and built shared technical understanding that fed directly into arms-control agreements, work recognised with a Nobel Peace Prize. The climate regime has the IPCC, an institutionalised epistemic community whose assessments are the agreed factual basis for negotiation. In each case, the expert network did essential work that governments could not have done for themselves.
Where AI's epistemic community stands
AI's situation is mixed. On one hand, a serious safety research community exists, and it has produced landmark expressions of consensus — the 2023 statement on AI extinction risk signed by leading researchers, the 2025 call to prohibit unsafe superintelligence, and the first International AI Safety Report, an IPCC-style synthesis chaired by a senior figure in the field. These are exactly the artifacts an epistemic community produces. On the other hand, the field is younger, more internally divided, and far less institutionally connected to policymakers than the atmospheric-science or nuclear-physics communities were at the equivalent moment.
The gaps are consequential. The AI safety community remains split on key questions of timelines and severity, which weakens the unified signal that moves governments. It is heavily concentrated in a few countries and institutions, limiting the transnational breadth that gave ozone and nuclear expertise their cross-border credibility — and, notably, its links to Chinese counterparts are thin, precisely where shared understanding is most needed. And it has only begun to build the durable institutional bridges to policymakers that turn expert consensus into political action.
Every treaty that governed a technical danger stood on a foundation of shared expert understanding that came first. For AI, building that community — consensus enough to speak clearly, and connected enough to be heard, across borders — is not a side project. It is part of building the treaty.
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
Building the community as governance work
The lesson is that strengthening AI's epistemic community is itself a governance strategy, not a precondition to be assumed. The consensus statements and the International AI Safety Report are the right kind of work, and they should be deepened: broadening the community across borders, especially to China; building the institutional channels that carry its findings into the policy process; and forging enough shared understanding of the core risks to present governments with the clear, credible signal that made past agreements possible. History suggests that the treaty and the expert community that makes it possible are built together. For AI, investing in the community — its consensus, its breadth, and its connection to power — is among the most durable contributions to the eventual agreement. The experts do not sign the treaty. But without them, there may be no treaty to sign.
Common questions.
A concept from political scientist Peter Haas for a network of professionals with recognised expertise and shared beliefs about a problem — its causes, stakes, and the policies that would address it. When such a community reaches consensus and gains the ear of decision-makers, it can supply the shared factual basis that makes international agreement possible among states that otherwise distrust each other's claims.
Not through formal power but by providing what governments need under uncertainty: a credible, shared account of a complex problem. They frame what the threat is and how serious it is, build cross-border consensus that gives distrustful states a common basis to negotiate from, translate diagnosis into workable policy options, and legitimise action. Because they are transnational and bound by professional standards, their understanding can be accepted across national lines.
The ozone regime, where atmospheric scientists' consensus made the Montreal Protocol possible; nuclear arms control, shaped by the Pugwash movement of scientists from both Cold War blocs; and the climate regime, whose IPCC is an institutionalised epistemic community providing the agreed factual basis for negotiation. In each case the expert network did essential agenda-setting and consensus-building work that governments could not have done themselves.
Partially. A serious safety research community exists and has produced consensus artifacts — the 2023 extinction-risk statement, the 2025 call to prohibit unsafe superintelligence, and the first International AI Safety Report. But the field is younger, more internally divided on timelines and severity, heavily concentrated in a few countries with thin links to China, and less institutionally connected to policymakers than the ozone or nuclear communities were. Strengthening it is itself a governance task.