The International AI Safety Report (IAISR) is an independent scientific assessment of risks from general-purpose AI, chaired by Yoshua Bengio. Its first full edition was published 29 January 2025 ahead of the Paris AI Action Summit. More than 96 international experts contributed, and the Expert Advisory Panel includes nominees from 30 nations plus the EU, UN, and OECD. The report describes itself as "the first ever independent, international scientific report on AI safety" and is modeled loosely on the IPCC for climate.
Summary of findings
The IAISR organizes risks from general-purpose AI into three categories:
- Malicious use — fraud, scams, cyberattacks, deepfakes and manipulation at scale, and CBRN threat enablement.
- Technical failures — unreliable reasoning, poor generalization, mis-specified objectives, concealment of unsafe behaviour during testing, and alignment / loss-of-control concerns.
- Systemic risks — provider concentration, cascading failures across interconnected infrastructure, labour-market disruption, environmental costs, and information-ecosystem degradation.
The report's central framing is the evidence dilemma: policymakers must balance the costs of early intervention against the costs of waiting for more complete data, given genuine uncertainty about the pace and severity of AI advancement. Many risks are framed in the report as "manageable but requiring proactive, anticipatory governance."
The systemic-risks chapter explicitly covers environmental cost — energy, water, and compute — a theme connected to AI Environmental Impact. It also treats provider concentration as a systemic risk, an analysis that parallels the Techno-Federalism: How Regulatory Fragmentation Shapes the U.S.-China AI Race and AI Sovereignty concerns documented on related pages.
Key claims
The report presents itself as the flagship scientific product of the Bletchley-initiated summit series; by Paris it had become the default reference document for ministerial AI discussion. Its three-category taxonomy of malicious use, technical failures, and systemic risks is the same taxonomy used in A Taxonomy of Systemic Risks from General-Purpose AI and aligns with AI Safety Cases and Frameworks.
The "evidence dilemma" framing is used to legitimize action under uncertainty, positioned as a counterweight to "wait-and-see" regulatory postures. Among the technical failure modes, the report endorses the empirical reality of concealment of unsafe behaviour — scheming and evaluation gaming, per AI Scheming — aligning with the We Need a Science of Scheming and Alignment Faking literatures.
These claims are supported by the report itself; confidence in this summary is high.
Publication history
The report developed across several editions tied to the summit series. An interim report was published in May 2024 ahead of the Seoul Summit. The first full report followed on 29 January 2025 ahead of the Paris Summit. In October 2025 the project issued a "Key Update" on reasoning-model capability advances and new monitoring and controllability evidence. A second full edition was published on 3 February 2026 ahead of the Kigali / India AI Impact Summit.
Provenance
The IAISR is the scientific track of the summit process initiated by the The Bletchley Declaration (AI Safety Summit, 1–2 November 2023), its commissioning body. It is linked politically to the Frontier AI Safety Commitments (Seoul, 2024) and the Paris AI Action Summit Declaration (2025). The Bengio chairmanship connects the report to the academic lineage described in AI Safety Cases and Frameworks. The report serves as a scientific anchor for AI Safety Cases and Frameworks, AI Scheming, AI Autonomy Risk, AI Environmental Impact, and A Taxonomy of Systemic Risks from General-Purpose AI.
Relationships
- depends-on: The Bletchley Declaration (AI Safety Summit, 1–2 November 2023) (commissioning body)
- supports: AI Safety Cases and Frameworks, AI Scheming, AI Autonomy Risk, A Taxonomy of Systemic Risks from General-Purpose AI
- related: Frontier AI Safety Commitments (Seoul, 2024), Paris AI Action Summit Declaration (2025), OECD — Assessing Potential Future AI Risks, Benefits, and Policy Imperatives, Yoshua Bengio, We Need a Science of Scheming, AI Environmental Impact