International AI Safety Report 2025

Introduction

The first International AI Safety Report, which is the foremost independent report and an international AI safety manual, was published in January 2025. The report is inspired by the Intergovernmental Panel on Climate Change (IPCC) report of the United Nations (UN). It outlines the better and rigorous application of the scientific method in determining AI safety. It includes the insights into AI safety or the risks related to advanced AI, as provided by 100 globally-acclaimed AI experts coming from 33 countries, including the USA, the UK, France, China, etc., and some are even associated with the international organisations, like the UN, the European Union (EU), and the Organisation for Economic Cooperation and Development (OECD). These AI leaders collectively form the Expert Advisory Panel at an international level.

This report highlights the state of several modern AI systems available in today’s world, their capabilities, and the risks associated with them. It encompasses the first detailed and common scientific analysis of these systems. The report is formed under the able leadership of Yoshua Bengio, one of the most popular computer scientists in the world and a prominent AI educator who shared the ACM A.M. Turing Award (often called the Nobel Prize of Computing) with Geoffrey Hinton and Yann LeCun in 2018. The report will serve as a guideline for policymakers while framing policies on AI safety.

Yoshua Bengio is currently working as a full professor at Universite de Montreal and the founder and scientific advisor of Mila, a Quebec AI Institute.

Earlier Developments

The discourse on AI safety has evolved over the years, driven by ethical, regulatory, and geopolitical concerns. In 2017, the Asilomar AI Principles, organised by the Future of Life Institute, set foundational guidelines emphasising transparency, accountability, and human oversight. In 2018, the OECD AI Principles reinforced the need for fairness, trust, and responsible AI deployment. The EU’s AI Act, 2021 introduced a risk-based classification framework for the AI regulation. During the AI Safety Summit held at Bletchley Park (the UK) in November 2023, it was agreed upon by 30 nations to create an international and comprehensive report regarding the capabilities and risks associated with advanced AI. In May 2024, an interim version of this report was released which was discussed at the AI Seoul Summit.

Yoshua Bengio presented the report at the AI Action Summit held in Paris, France, in February 2025. The Department for Science, Innovation, and Technology (DSIT) was responsible for operational support.

Objective of the Report

The main objectives of this report include understanding the capabilities and evolution of general-purpose AI (GPAI) models and assessing the risks associated with it at an international level. It also includes a number of techniques to reduce these risks. It has shown rapid advancement recently, and a large number of technology companies have installed this type of AI for varied commercial purposes. For the purpose of attaining this objective, the GPAI has been considered in the report. GPAI is used to carry out a wide range of tasks.

In this report, risks are highlighted, interpreted, and categorised into technical, societal, economic, and geopolitical dimensions. The report also proposes actionable policy recommendations which include the method in which the GPAI itself reduces the risks.

Some Highlights of the Report

1. Capabilities of the GPAI Recently, there has been a drastic improvement in the capabilities of the GPAI. It includes composing computer programs, getting involved in long open-ended dialogues, producing photorealistic images, scientific reasoning, etc. However, its earlier large-language models failed to give even logical text as their output.
2. Development of the GPAI Agents The AI systems that can plan, act, and even assign tasks without human assistance to fulfil their objectives are referred to as the GPAI agents. A large number of companies are adopting GPAI agents for their day-to-day tasks, which will help in their advancement. With time, the advanced GPAI agents will become capable of completing lengthy projects using computers unlike their current counterparts. These will provide additional benefits along with imposing additional risks.
3. Evolution of the Capabilities of GPAI This evolution can be determined by two things: (i) if new models can be trained by companies by swiftly providing them with more data and computational power, and (ii) if models would be able to deal with their constraints after getting trained in such a manner. According to a recent study, it may be physically possible to improve models quickly for some years. However, drastic improvement in their capabilities would depend on other factors as well, such as unpredictable innovation in research and how far the new approach used for training models by the companies succeeds.
4. Certain Disadvantages of GPAI Non-consensual intimate imagery, scams, reliability concerns, violation of privacy, child sexual abuse material, and outputs of model that show favouritism towards particular groups of people or certain viewpoints, while oppose other groups or viewpoints are some of the major disadvantages of the GPAI. Many researchers have tried to lessen their adverse effects through some techniques but have failed to eradicate these issues completely. In the past few months, new subtle bias carried out by the GPAI systems has been witnessed.
5. Risks with Increased Capability of GPAI Some of the major risks are biological attacks or hacking facilitated by AI, wide-ranging impacts of labour market, and the GPAI becoming out of control of the society. Several researchers and programmers do not see eye to eye on these risks. Some opine that such risks will appear only after several decades, while others think that risks imposed by the GPAI at a societal level will be evident within a few years. With increasing capabilities of the GPAI, such as programming and scientific reasoning tests, potential risks like hacking and biological attacks by AI, have become apparent. As a result, a notable AI company has rigorously assessed the biological risk imposed by its top-class model.
6. Newly Formed Risk-Management Techniques Risks imposed by the GPAI can be assessed and reduced through a number of techniques adoptable by both the developer and the regulator. However, there are certain limitations of these techniques as well. Let us consider one of these techniques, known as current interpretability technique. This technique gives a reason behind any given output generated by a GPAI model. So, the researchers have also found ways to deal with these limitations.
   
   Besides, the standardisation of risk-management techniques by policymakers and researchers is underway. They also intend to coordinate on an international level.

7. Lack of Evidence due to Unpredictability of GPAI The improvements in the GPAI are often swift and unexpected. So, policymakers are required to analyse the benefits as well as the risks imposed by these AI improvements in the absence of adequate scientific evidence. At this point, a dilemma appears in front of them. The risk-mitigation measures may prove to be unsuccessful if they are formed on the basis of limited evidence of potential risk. Contrary to this, obtaining compelling evidence may take too long to prepare the society to address the risks. Thus, it may lead to an unprepared society, wherein mitigation may not be an option. This may happen if AI capabilities suddenly improve, which will bring related risks.
   
   This dilemma can be tackled by introducing early warning systems and creating risk-management frameworks. This task has been taken up by the governments and various companies.
   
   Upon encountering new evidence of risks, certain warning systems may initiate a particular mitigation measure. While in other warning systems, a new model is launched only when evidence of safety is given by developers.  

8. Mutual Consent by Researchers A large number of researchers have mutually consented that they could be benefitted with the improvements in the following areas: the rapid advancement in the capabilities of the GPAI in the near future; the reliable measurement of this advancement by the researchers; the sensible upper limit of risk that leads to mitigations; the access to public safety-related information about the GPAI by policymakers; the reliable assessment of risks associated with the creation and installation of the GPAI by researchers, governments, and technology companies; the internal functioning of the GPAI models; and the designing of the GPAI to facilitate its reliable behaviour.

9. Dependence of the Future AI on People’s Choice There is much uncertainty with respect to the future of the GPAI technology, which can produce both positive and negative outcomes. It is up to the government, industries, and societies to make necessary decisions to deal with this uncertainty. These decisions can be taken after effective discussions based on evidence.

Conclusion

To conclude, this report is an attempt to bridge the gap that has been created due to swift and unpredictable advancements in AI. The report does so by creating a scientific basis of evidence, which will enable policymakers around the world to improve the safety in the applications enabled by AI. It encompasses critical challenges and threats imposed by swift advancements in the AI technologies. However, the societal impacts of the GPAI are not measured by the report thoroughly. Nor does the report reveal the benefits of the GPAI, whether current or of potential future. Experts and researchers would be required to take into account the potential benefits and risks associated with the GPAI so that policymaking can be done holistically. The report states that the decisions made by policymakers will determine the outcomes. For the people across the world to safely experience the full benefits of the GPAI, it is essential to ensure that its risks are properly managed.

© Spectrum Books Pvt. Ltd.