International AI Safety Report to inform global policy discussions
Philip Fox, 29 January 2025
Today, a group of ~100 AI experts published the International AI Safety Report. Mandated by 30 governments, the UN, OECD and EU, this report synthesizes the current scientific understanding of AI capabilities, risks and mitigations.
The result of an unprecedented scientific collaboration – with Nobel laureates, Turing Award winners and experts from around the world among the contributors –, the report is a basis for informed policymaking at a time of increasingly complex discourse around AI.
KIRA members have been involved in this project as Lead Writer (Daniel Privitera) and Writer (Philip Fox).
The report focuses on general-purpose AI, whose capabilities have increased particularly rapidly in recent years and months. It is organized around three core questions:
What can general-purpose AI do?
What risks does it pose?
How can we mitigate those risks?
Key findings:
AI is improving at a fast pace: Over the past 18 months, the best models have gone from „Random Guessing“ to „PhD-Level expert“ performance on tests of programming and scientific reasoning.
Companies are increasingly betting on ‘inference scaling’ – allowing an AI to use ever more computational power for solving problems – as a potential new way to reliably get stronger capabilities going forward.
Several harms from general-purpose AI are already well-established. These include scams, copyright violations and non-consensual intimate imagery (NCII).
As general-purpose AI becomes more capable, evidence of additional risks is gradually emerging. These include large-scale labour market disruptions, AI-enabled hacking or biological attacks, and society losing control over general-purpose AI.
Various risk management techniques exist – such as model evaluations, privacy-enhancing measures or interpretability techniques –, but they all have limitations.
Going forward, policymakers will increasingly find themselves in an ‘evidence dilemma’: pre-emptive measures without sufficient evidence may prove unnecessary or ineffective, while waiting for further evidence could leave society unprepared for major risks.
The report will inform discussions at the upcoming AI Action Summit in Paris on February 10-11.