The AI safety report made evidence limits part of the story

Published on February 3, 2026, the International AI Safety Report assessed general-purpose AI using evidence published before December 2025. Its publication date therefore did not make it a record of every later model or incident. The authors distinguished documented harms involving generated media from future risks investigated through modelling, laboratory studies and theory. Those evidence categories answer different questions; a controlled experiment does not establish how frequently a problem occurs in deployed services. ^[1]

The project’s institutional history explains its purpose. Countries participating in the 2023 Bletchley summit sought a shared scientific baseline. The report series, chaired by Yoshua Bengio, asks what general-purpose systems can do, which risks they pose and which technical mitigations exist. Its About page describes a research synthesis without policy recommendations. An international assessment supplies evidence for decisions; it is not itself a product approval or a prescription for a particular public service. ^[2]

An earlier update, published November 25, 2025, explained why safeguards needed more than a checklist. Controls could reduce particular vulnerabilities, yet performance varied across deployment contexts and implementation was inconsistent. It also identified limited shared metrics for evaluating safeguards over time. The existence of a published safety framework was consequently different from evidence that documented controls produced the intended outcomes. ^[3]

For readers assessing a safety claim, those distinctions suggest concrete questions: which conditions were tested, how does the proposed deployment differ, and what evidence connects the controls to real-world results? This retrospective did not evaluate a model or verify a company’s implementation. ^[3]

About this item: Retrospective report. No first-hand reporting or independent testing is claimed. Corrections policy.