Yoshua Bengio-led International Scientific Report on the Safety of Advanced AI published

The UK government published the interim International Scientific Report on the Safety of Advanced AI, an expert assessment of frontier-AI risk chaired by Turing Award-winning researcher Yoshua Bengio, ahead of the AI Seoul Summit five days later. The mandate for the report originated from the Bletchley Declaration, where nations agreed to build a shared scientific understanding of the risks involved.

The report was explicitly modelled on the Intergovernmental Panel on Climate Change: rather than one government or lab’s assessment, it drew on around 75 contributors, including an international Expert Advisory Panel of 32 members nominated by 30 countries, the EU and the UN, plus dozens of senior advisers chosen for diversity of view. It synthesised existing research into three broad risk categories rather than generating new findings — malicious use (scams, disinformation, cyber-offence, deepfakes), malfunction (safety failures, bias, loss of control) and systemic risk (labour-market disruption, market concentration, environmental cost, privacy) — and it emphasised, in Bengio’s later summary, that “technical methods exist to reduce risk, but all have limitations,” alongside considerable uncertainty about how quickly capabilities would advance.

Because it was explicitly a synthesis rather than an argument for a particular policy, the report avoided taking a position on flashpoints like existential risk that had divided the field, instead cataloguing the range of expert views and the state of evidence behind each. That made it a less quotable document than the open letters and summit declarations surrounding it, but a more durable reference: it became the first in a series of annual International AI Safety Reports, with a full first edition following in January 2025.

Its immediate function was preparatory — establishing a body of agreed evidence for governments and companies to draw on when negotiating commitments at the summit in Seoul, where 16 companies signed the Frontier AI Safety Commitments days later.