⚡ Quick Summary
Published by the Department for Science, Innovation and Technology (DSIT) under Crown copyright, this report synthesises the scientific evidence on the safety of general-purpose AI, defined as "AI that can perform a wide variety of tasks". It is the first full International AI Safety Report, written by 96 experts under Chair Prof. Yoshua Bengio, with an Expert Advisory Panel nominated by 30 countries, the EU, the OECD and the UN, and published ahead of the AI Action Summit in Paris in February 2025.
The report answers three questions: what general-purpose AI can do, what risks it poses, and what techniques exist to assess and mitigate those risks. It groups risks into malicious use — fake content, manipulation of public opinion, cyber offence, biological and chemical attacks — malfunctions (reliability, bias, loss of control) and systemic risks (labour market, global AI R&D divide, market concentration, environment, privacy, copyright). A Chair's note adds OpenAI's o3 results, which emerged after the 5 December 2024 writing cut-off.
Its conclusion is that capabilities have advanced rapidly; that risk management methods, while progressing, have "highly significant limitations" and offer no quantitative risk guarantees; and that policymakers face an "evidence dilemma" when weighing pre-emptive mitigation against waiting for stronger evidence. The report states that it does not recommend specific policies.
🧩 What’s Covered
After an introduction, the report has three main sections and a conclusion.
- About this report and the Chair's note: sets out the purpose, the experts' independence, the focus on general-purpose AI and the three core questions; the Chair's note records OpenAI's o3 results as evidence that the pace of capability advances may remain high or accelerate.
- Capabilities (Section 1): defines models, systems, compute, deep learning, weights and AI agents; describes the lifecycle from data collection and pre-processing through pre-training, fine-tuning, system integration, deployment and post-deployment monitoring; reviews current abilities across text, code, audio, images, video, robotic actions and proteins; and projects future trends, citing 4x per year growth in pre-training compute, 2.5x per year in dataset size, and roughly 100x more training compute than 2023's most compute-intensive models by 2026, reaching 10,000x by 2030.
- Risks from malicious use (2.1): fake content that harms individuals (deepfake pornography, CSAM, fraud, blackmail, sabotage), manipulation of public opinion, cyber offence, and biological and chemical attacks.
- Risks from malfunctions (2.2): reliability issues, bias, and hypothetical loss-of-control scenarios.
- Systemic risks (2.3): labour market impacts, the global AI R&D divide, market concentration and single points of failure, environmental effects, privacy, and copyright infringement.
- Open-weight models (2.4): how release decisions shape risk, and the shift to assessing openness in terms of "marginal" risk relative to existing alternatives.
- Technical approaches to risk management (Section 3): risk management overview; technical and societal challenges such as the "evidence dilemma" and the information gap between companies and governments; risk identification and assessment; and mitigation and monitoring, covering training more trustworthy models, monitoring and intervention, and technical privacy methods.
- Conclusion: future trajectories depend on the choices societies and governments make.
💡 Why it matters?
For those who govern, assess or audit AI, the report offers a map of where scientific agreement exists and where it does not, rather than prescriptions. It separates harms that are already well established — scams, non-consensual intimate imagery, bias, reliability failures and privacy violations — from those that remain contested or evidence-poor, and explains why general-purpose AI is unusually hard to manage. Its account of the "evidence dilemma", of evaluations that miss hazards, and of frameworks that trigger mitigations when risk rises is directly usable when designing monitoring, evaluation and escalation duties, and it ties explicitly to the Bletchley Park, Seoul and Paris summit process.
❓ What’s Missing
The report deliberately excludes the benefits of general-purpose AI and makes no policy recommendations, so it cannot serve as a compliance checklist. It states that risk management methods have "highly significant limitations" and that there are no quantitative risk estimates or guarantees of the kind available in other safety-critical domains. Several areas rest on thin evidence: statistics on deepfake harms, the effectiveness of manipulation campaigns, and much biological and chemical risk research, which is classified. Narrow AI, military applications and lethal autonomous weapons are out of scope. The Chair's note concedes that capabilities advanced after the 5 December 2024 cut-off and that no evidence yet exists on o3's real-world impacts. Only the first 80 of 298 pages were available for this review.
👥 Best For
Best for policymakers and their advisers who need a defensible overview of the scientific evidence on AI risk; for risk, compliance and audit teams building evaluation, monitoring and escalation frameworks; and for technical teams designing capability and safety evaluations. Researchers will find the report's explicit evidence gaps a useful agenda.
📄 Source Details
International AI Safety Report: The International Scientific Report on the Safety of Advanced AI, January 2025, published by the Department for Science, Innovation and Technology (DSIT) under Crown copyright; research series number DSIT 2025/001. Chair: Prof. Yoshua Bengio. 298 pages, in English. No URL for the report itself is printed in the extracted pages. The input was a partial text extraction covering pages 1–80 — the front matter, Section 1, Section 2.1 and the start of 2.1.4; sections 2.2 onwards were not available.
