International AI Safety Report 2026 — Evidence Base Without U.S. Endorsement
Second International AI Safety Report published Feb 3, chaired by Yoshua Bengio with 100+ experts and nominees from 30+ countries. Frontier capability/risk synthesis for policymakers—no specific policy prescriptions. U.S. declined to endorse the final version (TIME/Bengio); timed ahead of New Delhi AI Impact Summit.
As labs ship agentic long-context systems, independent multi-national assessments become reference points—or political flashpoints. The proprietary tension in the 2026 edition: largest scientific synthesis yet, and the United States withholding endorsement of the final text.
International AI Safety Report 2026 published February 3 (DSIT 2026/001), chaired by Turing Award winner Yoshua Bengio, with contributions from over 100 experts and an advisory panel nominated by more than 30 countries/organizations. Scope: “emerging risks” at the frontier of general-purpose AI; updates prior assessments with capability-trajectory research including OECD and Forecasting Research Institute scenario work. Core: rapidly improving capabilities and uncertainty on scale/timing; risk-management methods improving but insufficient for some frontier scenarios; expert disagreement on AI-assisted research automation and predictability of capability jumps. Explicitly does not make specific policy recommendations and does not speak for any single government or the full expert group—goal is an evidence base for policymakers. U.S. position: Bengio confirmed the United States declined to back the final report (unlike prior edition), after providing draft feedback—TIME reporting Feb 2–3. Timed ahead of the AI Impact Summit in New Delhi.
Claims vs checks
Publication date, chair, expert scale, and non-prescriptive mandate are report primary. U.S. non-endorsement is Bengio statement via TIME—diplomatic, not a scientific rebuttal of chapters. Treat disagreement sections as features of the method, not bugs to flatten.
Limits
- Synthesis ≠ audit of any one lab’s system card.
- Non-endorsement reasons are political; the report text still stands for readers.
- Forecasts and scenarios carry wide uncertainty bands by design.
Sources
- International AI Safety Report 2026 (3 February 2026). Primary. Bengio et al., DSIT 2026/001.
- TIME: “U.S. Withholds Support From Global AI Safety Report” (Feb 2–3, 2026 context).
- Report foreword/methodology: expert panel, lead writers, scope limitations.