The International AI Safety Report 2026, published on 3 February 2026 and led by Turing Award winner Yoshua Bengio with more than 100 independent experts, concludes that general-purpose AI systems now solve graduate-level problems and write functional code, while measurable evidence of malicious use, malfunction and systemic harm is accumulating faster than institutional safeguards can absorb it. For research-integrity and governance teams, its findings translate directly into fabrication, authorship and peer-review risk that existing standards bodies are only beginning to address.
The International AI Safety Report 2026 is an independent, government-backed scientific synthesis — not a regulation — that reviews peer-reviewed evidence on general-purpose AI capabilities and risks for policymakers worldwide.
- What is the International AI Safety Report 2026?
- What does it say about general-purpose AI capabilities?
- What risks does it flag for research and scientific integrity?
- How does it relate to the OECD, UNESCO and the AI Safety Institute Network?
- What should research institutions and publishers do next?
- Answer-first Q&A
What is the International AI Safety Report 2026?
The report is the second edition of a scientific review process first commissioned after the 2023 UK AI Safety Summit at Bletchley Park. An Expert Advisory Panel nominated by more than 30 countries and organisations, including the OECD and the EU, oversees the work; Yoshua Bengio chairs the writing team of over 100 independent experts. The 2026 edition follows the inaugural full report of 29 January 2025 and is structured around three questions: what general-purpose AI can do today, what emerging risks it poses, and how those risks are being managed.
Crucially, the report is descriptive rather than prescriptive. It does not set binding rules; it gives governments, standards bodies and institutions a shared, evidence-based picture to legislate and self-govern against — the same function CASRAI’s frameworks perform for research administration.
What does it say about general-purpose AI capabilities?
The report documents rapid gains in reasoning, multilingual fluency, code generation and scientific problem-solving. Models can now generate and evaluate multiple candidate solutions to improve performance on specialised tasks, including protein design and graduate-level mathematics.
- General-purpose AI is increasingly used by researchers for literature review, data analysis and experimental design.
- Reasoning models can iterate over intermediate steps, improving accuracy on multi-stage scientific tasks.
- Since the 2025 report, the number of companies publishing Frontier AI Safety Frameworks has more than doubled, according to the report’s technical safeguards update.
These gains are precisely why the report’s authors describe an “evidence dilemma”: acting on a risk before conclusive proof exists may produce weak or misdirected policy, but waiting for definitive evidence risks leaving institutions unprepared when harm materialises at scale.
What risks does it flag for research and scientific integrity?
The report groups risk into three categories — malicious use, malfunctions and systemic risks — and each maps onto a live research-integrity problem that CASRAI, NISO, COPE and ICMJE are already tracking.
| Report risk category | Research-integrity manifestation |
|---|---|
| Malicious use | Paper-mill-generated manuscripts, fabricated datasets, and AI-written peer reviews submitted without disclosure |
| Malfunctions (hallucination, flawed output) | Fabricated citations, invented statistics and confidently wrong methodology sections that pass initial editorial screening |
| Systemic risk | Erosion of trust in the scholarly record as AI-assisted authorship becomes undisclosed and unattributed at scale |
The report notes that current detection and safety training methods have not eliminated hallucination or flawed code generation, and that growing AI agent autonomy makes human review harder to insert before harm occurs. For journals and institutions, that is a direct argument for mandatory AI-use disclosure and for contributor-role transparency — the function the CRediT contributor role taxonomy already performs for human authorship.
How does it relate to the OECD, UNESCO and the AI Safety Institute Network?
The International AI Safety Report sits inside a wider, pre-existing governance ecosystem rather than replacing it. Understanding how the pieces relate — which none of the report’s own summary pages set out — is the practical question research administrators actually need answered.
| Framework | Type | Adopted | Function |
|---|---|---|---|
| OECD AI Principles | Intergovernmental standard | 2019, updated 2024 | First intergovernmental AI policy standard; underpins G20 AI principles and national strategies |
| UNESCO Recommendation on the Ethics of Artificial Intelligence | Global normative instrument | November 2021 (193 member states) | First global standard-setting instrument on AI ethics, including research and education use |
| International Network of AI Safety Institutes (“AI Safety Institute Network”) | Technical evaluation network | Formed following the May 2024 Seoul AI Summit | Coordinates frontier-model testing and incident reporting across national AI safety institutes |
| International AI Safety Report 2026 | Independent scientific synthesis | 3 February 2026 (2nd edition) | Synthesises evidence for all of the above to draw on when setting policy |
The report explicitly feeds the evidence base that the AI Safety Institute Network’s technical evaluations, the OECD’s principle updates and UNESCO’s ethics guidance all draw on. None of these bodies regulates research publishing directly, which leaves that gap to sector standards bodies and journals themselves.
What should research institutions and publishers do next?
Research administrators do not need to wait for binding AI legislation to act. The report’s own “evidence dilemma” argument supports proportionate, revisable policy now rather than delay.
- Require disclosure of generative-AI use in manuscript preparation, data analysis and peer review, consistent with existing ICMJE and COPE guidance.
- Extend authorship and contribution attribution policies to explicitly exclude AI systems as authors while requiring disclosure of their use.
- Treat the report’s malfunction findings — persistent hallucination and flawed code generation — as grounds for mandatory human verification of AI-assisted literature reviews and statistical outputs.
- Track institute-level AI evaluation outputs from the AI Safety Institute Network alongside research administration risk registers, since frontier-model incidents can affect research tools before they affect the public.
Answer-first Q&A
What is the International AI Safety Report 2026?
It is the second edition of an independent scientific review of general-purpose AI capabilities and risks, published 3 February 2026 and chaired by Yoshua Bengio with over 100 experts backed by more than 30 countries. It informs policymakers rather than setting binding rules.
Who wrote the International AI Safety Report 2026?
An Expert Advisory Panel nominated by over 30 governments and organisations, including the OECD and EU, oversaw the report. The writing team of more than 100 independent experts was chaired by Yoshua Bengio, a Turing Award-winning AI researcher.
What are the main risks the report identifies?
The report groups risk into three categories: malicious use (fraud, deepfakes, disinformation), malfunctions (hallucination, flawed code, dangerous advice) and systemic risks from AI’s widespread integration into critical sectors, including research.
What is the AI Safety Institute Network?
The International Network of AI Safety Institutes is a technical coordination body formed after the May 2024 Seoul AI Summit. It links national AI safety institutes to share frontier-model evaluation methods and incident data that feed reports like this one.
Looking ahead
The International AI Safety Report 2026 does not resolve the “evidence dilemma” it describes — it sharpens it. For research-integrity and governance functions, the report’s malfunction and malicious-use findings are not abstract policy risk; they are already surfacing in manuscript submissions, peer review and grant applications. Institutions that align AI-disclosure policy with existing contributor-attribution standards now will be better placed than those waiting for the OECD, UNESCO or national regulators to catch up with a fast-moving evidence base.