Governance without shared scientific foundation is negotiation without a common map. The climate negotiations before 1990 illustrate this problem: governments that disputed whether human-caused warming was occurring at all could not agree on what level of response was appropriate, because they were arguing from incompatible factual premises. The IPCC's first assessment report, published in 1990, established that warming was occurring and was attributable to human activity. It did not end political disputes about what to do, but it changed the character of those disputes from arguments about whether a problem existed to arguments about how to respond to one.
The politics of AI risk are in a pre-IPCC state. There is no authoritative international body that synthesizes the evidence on frontier AI risks, assesses the probability and character of catastrophic outcomes, and produces findings that governments treat as the shared factual basis for governance decisions. This gap is not a minor inconvenience. Governance negotiations that begin without shared factual premises tend to produce frameworks shaped by strategic interests more than by the character of the risk being governed.
How the IPCC actually built consensus
The IPCC does not conduct original research. It synthesizes existing peer-reviewed research, produced by thousands of scientists in universities and research institutes around the world, into comprehensive assessment reports. The synthesis process involves hundreds of expert authors who review and summarize the published literature in their areas of expertise. The summary for policymakers, the section that governments and journalists actually read, is produced through a line-by-line negotiation between scientist authors and government representatives from all 195 IPCC member countries.
This negotiation process is the source of the IPCC's political authority, but also of its conservative character. Any government can object to specific language in the summary, and the process requires that objections be addressed before the summary is adopted. This means the summary reflects the lowest common denominator of what all participating governments will accept, which is almost always more cautious than what the underlying science would support if read without political filters. The IPCC's summaries consistently understate urgency relative to the full assessment reports they summarize, because the summary language must survive government review.
Despite this conservative bias, the IPCC's assessments have been consequential. The progression from the first assessment report in 1990, which established the basic scientific case for human-caused warming, to the sixth assessment cycle completed in 2022, which characterized the evidence as unequivocal and the risks as severe and widespread, has produced a cumulative normative shift that has made it increasingly difficult for governments to publicly deny the basic facts of climate change. The political cost of denial has risen with each assessment cycle because the IPCC's authority is recognized across virtually all political systems.
The Pugwash alternative
Before the IPCC, the most influential science-policy interface for existential risk was the Pugwash Conferences on Science and World Affairs. Pugwash began in 1957 in response to the Russell-Einstein Manifesto, a 1955 statement signed by Albert Einstein, Bertrand Russell, and nine other prominent scientists warning of the dangers of nuclear weapons and urging governments to resolve disputes through peaceful means. The first Pugwash Conference, held in the Nova Scotia village of Pugwash in July 1957, brought together twenty-two scientists from ten countries, including the United States, Soviet Union, and United Kingdom, to discuss nuclear weapons risks.
What made Pugwash different from formal intergovernmental processes was its informal character. Participants attended as individuals, not as government representatives. The discussions were off the record. Scientists from adversarial countries could explore positions and ideas that their governments had not authorized them to advance officially. This informality made it possible to have the kind of candid exchange across Cold War lines that was impossible in formal diplomatic settings.
Pugwash did not produce treaties. It produced the intellectual groundwork that formal negotiations later formalized. The concept of mutual assured destruction as a basis for deterrence stability was refined in Pugwash discussions before it became US official policy. The Partial Test Ban Treaty (1963) and the Non-Proliferation Treaty (1968) both drew on ideas developed in Pugwash discussions. The organization was awarded the Nobel Peace Prize in 1995, jointly with Joseph Rotblat, its long-serving secretary-general, in recognition of this contribution to arms control.
The IPCC and Pugwash represent two different functions that a science-policy interface can perform. The IPCC synthesizes existing research into authoritative consensus assessments with political legitimacy derived from intergovernmental review. Pugwash created informal dialogue across political lines that allowed ideas to develop before they were politically viable. AI risk governance needs both: systematic evidence synthesis to establish shared factual premises for policy, and informal cross-geopolitical dialogue to develop governance concepts before formal negotiations begin. These functions could be performed by a single body or by two distinct ones.
What an AI risk panel would need to do differently
An international panel on AI risk modeled on the IPCC would face several challenges that have no equivalent in climate science.
The proprietary knowledge problem. Most of what is known about the capabilities and risks of frontier AI systems is held by a small number of private companies. The research literature on frontier AI safety is partially published, but the most relevant information, including the results of internal capability evaluations, safety test results, and incident reports, is proprietary. The IPCC synthesizes public peer-reviewed research. An AI risk panel would need access to non-public information to synthesize the most relevant evidence, which would require that frontier AI developers provide that information to the panel under defined confidentiality arrangements. Designing those arrangements is a substantial governance challenge with no IPCC precedent.
The pace problem. The IPCC produces major assessment reports every five to seven years, with updates in between. This pace is suited to a physical phenomenon that changes on decadal timescales. AI capabilities are advancing on timescales of months. An AI risk panel would need to produce much more frequent assessments, or to establish a standing monitoring function that tracks capability developments in near-real time, to remain relevant to governance decisions. More frequent reports would be harder to produce through the IPCC's line-by-line intergovernmental review process, because that process is slow by design.
The dual-use knowledge problem. Climate science is publicly available knowledge that carries no inherent security risk. AI safety research sometimes involves understanding exactly how dangerous capabilities could be misused, which creates a tension between the transparency that makes assessment authoritative and the security concerns that argue for restricting some findings. An AI risk panel would need procedures for handling sensitive capability findings that the IPCC has never needed to develop.
Why the absence of such a body is already costly
The absence of an authoritative international body synthesizing AI risk evidence has concrete governance costs that are visible right now.
Governance negotiations proceed from incompatible factual premises. Different governments, and different factions within governments, operate with fundamentally different assessments of when artificial superintelligence might pose catastrophic risks, what the character of those risks is, and what governance measures would address them. Without a shared factual baseline, these negotiations tend to produce frameworks designed around the most politically convenient view of the risk rather than around the most accurate assessment of it.
Risk communication is fragmented and inconsistent. Warnings about advanced AI risks come from a heterogeneous collection of sources including academic researchers, AI company internal safety teams, think tanks, and advocacy organizations, with no shared methodology, no common vocabulary, and no process for adjudicating disagreements. The result is a public discourse that oscillates between dismissiveness and alarm without a reliable authoritative reference point. The IPCC's value for climate communication was precisely that it gave journalists, policymakers, and the public a single authoritative source that could be cited against both dismissal and exaggeration.
"The IPCC did not create political will to address climate change. It created the shared factual foundation that made ignoring the problem politically costly. AI governance needs the same thing: a body with the authority and methodology to establish what the evidence actually shows about frontier AI risks, so that governance debates proceed from shared facts rather than from strategically convenient uncertainty."
The most important precedent the IPCC provides for AI is not methodological but political. The IPCC was established in 1988, two years before its first assessment report and thirty-four years before the sixth assessment cycle's findings were described as producing impacts that are already widespread, rapid, and intensifying. The institutional investment in systematic evidence synthesis began long before the political will to act on that evidence arrived. For AI risk, the institutional equivalent needs to begin before the window for effective governance closes, not after it.
A panel on AI risk with intergovernmental membership, a mandate to synthesize published and, under appropriate confidentiality arrangements, non-published safety research, a faster reporting cadence than the IPCC, and access to independent technical expertise drawn from AI safety research communities would be buildable within existing international institutional frameworks. It would not require a treaty. It would require political commitment from a sufficient coalition of governments to establish and fund it. It is the kind of institution that can be created before, and in support of, binding governance obligations. Whether it will be created before the technology it would assess renders early assessments obsolete is a question with a closing answer window.
Common questions.
How does the IPCC build scientific consensus?
The IPCC synthesizes existing peer-reviewed climate research rather than conducting original research. Every five to seven years, it produces assessment reports in which hundreds of scientists review the published literature and summarize the state of knowledge. The summary for policymakers is then negotiated line by line between scientist authors and government representatives from all 195 IPCC member countries. This intergovernmental review process gives the summary political legitimacy, at the cost of conservative language that typically understates urgency relative to the underlying science. The process is slow by design, because consensus requires that every participating government accept the final language.
What is the Pugwash model and how does it differ from the IPCC?
The Pugwash Conferences, which began in 1957, brought together scientists from Western and Soviet bloc countries in an informal, off-the-record setting to discuss nuclear risks and arms control ideas. Unlike the IPCC, Pugwash was not intergovernmental: participants attended as individuals, not as government representatives, which allowed candid discussion across Cold War lines that official channels could not accommodate. Pugwash developed the intellectual groundwork for the Partial Test Ban Treaty and NPT before those agreements were politically viable. An AI equivalent might combine the IPCC's systematic evidence synthesis with the Pugwash model's informal cross-geopolitical scientific dialogue.
Is there already an international body that synthesizes AI safety research?
No body currently performs the IPCC function for AI safety. The OECD's AI Policy Observatory tracks national AI policies. The Global Partnership on AI funds research and publishes reports. The UN Secretary-General's AI Advisory Body issued governance recommendations in 2023. The AI safety summits at Bletchley and Seoul produced declarations. But none of these bodies conducts systematic synthesis of AI safety research with the combination of scientific authority and political legitimacy that the IPCC derives from its intergovernmental review process. This gap means governance negotiations proceed from fragmented and contested evidence.
Could an IPCC-style body actually change AI governance outcomes?
Yes, with limits the IPCC's own history makes clear. The IPCC did not create the political will to address climate change. It created the shared factual foundation that made denying the problem politically costly. An AI risk panel that could establish, with similar authority, the nature and likelihood of catastrophic AI risks would change the character of governance debates: from arguments about whether a problem exists to arguments about how to respond. It would not determine the outcome of those arguments, but it would discipline them in ways that shared facts always discipline political debates, raising the cost of positions that cannot survive contact with evidence.