The International AI Safety Report has been published, setting out the “first comprehensive, shared scientific understanding of advanced AI systems and their risks”.

The Report was first launched at the inaugural AI Safety Summit held at Bletchley Park in November 2023, and it represents the work of 96 international AI experts who were tasked with assessing the state of the science relating to general-purpose AI systems.  In particular, they focussed on three core questions: (1) what can general-purpose AI do; (2) What risks are associated with general-purpose AI; and (3) What mitigation techniques exist against these risks?

Importantly, the Report makes clear that it does not “aim to comprehensively assess all possible societal impacts of general-purpose AI”. Instead, its focus is on identifying the risks associated with general-purpose AI and evaluating technical methods for assessing and mitigating them, including by using general-purpose AI itself. However, it points out that “holistic policymaking requires considering both the potential benefits of general-purpose AI and the risks covered in this report”.

It is a dense document, comprising over 200 pages, but its main findings are summarised as follows:

• The capabilities of general-purpose AI continue to increase rapidly and further capability advancements in the coming months and years could be anything from slow to extremely rapid. As if to prove the point, one only needs to consider that between the completion of the writing period for the Report in early December last year, we have seen the release of OpenAI’s o3 model and DeepSeek.
• Several harms from general-purpose AI are already well established. These include deepfakes, scams, child sexual abuse material, biased and discriminatory models, hallucinations, and unreliability. The Report states that “researchers have developed mitigation techniques for these problems, but so far no combination of techniques can fully resolve them”.
• As general-purpose AI becomes more capable, evidence of additional risks is gradually emerging. The Report mentions risks of large-scale labour market impacts, AI-enabled hacking or biological attacks, and loss of control of AI systems. It also points out that there is some disagreement among experts as to how far we are away from such risks eventuating. Some suggest decades; others a matter of a few years.
• Risk management techniques are nascent, but progress is possible. The Report states that various technical methods exist to assess and reduce risks from general-purpose AI, but that they all have limitations. More encouragingly, it points to the fact that progress is being made and that researchers and policymakers are coordinating internationally and trying to standardise risk management approaches.
• The pace and unpredictability of advancements in general-purpose AI pose an ‘evidence dilemma’ for policymakers. As the Report explains, policymakers are faced with either prematurely introducing risk-mitigation measures based on limited evidence that might turn out to be ineffective or unnecessary, or to wait for more evidence only to go past the point at which mitigation measures will be able to address serious risks. It notes that companies and governments are developing early warning systems which either trigger scientific mitigation measures when there is new evidence of risks, or require developers to provide evidence of safety before releasing a new model.
• There is broad consensus among researchers that advances regarding the following questions would be helpful:

• • How rapidly will general-purpose AI capabilities advance in the coming years, and how can researchers reliably measure that progress?
  • What are sensible risk thresholds to trigger mitigations?
  • How can policymakers best gain access to information about general-purpose AI that is relevant to public safety?
  • How can researchers, technology companies, and governments reliably assess the risks of general-purpose AI development and deployment?
  • How do general-purpose AI models work internally?
  • How can general-purpose AI be designed to behave reliably?

• AI does not happen to us: choices made by people determine its future. Concluding on a more optimistic message, the Report states that “the future of general-purpose AI technology is uncertain, with a wide range of trajectories appearing to be possible even in the near future, including both very positive and very negative outcomes. This uncertainty can evoke fatalism and make AI appear as something that happens to us. But it will be the decisions of societies and governments on how to navigate this uncertainty that determine which path we will take. This report aims to facilitate constructive and evidence-based discussion about these decisions”.

The publication of the Report coincided with the commencement of the AI Action Summit in Paris and was intended to inform discussions there. However, some appeared not to get the memo: addressing the audience, the Vice President of the United States, JD Vance, began his speech by declaring “I’m not here this morning to talk about AI safety…I’m here to talk about AI opportunity” before going on to say that “the AI future is not going to be won by hand-wringing about safety”.

Later, it was confirmed that the USA and UK both refused to sign the AI Action Summit’s declaration which called for policies that ensured AI is “open, inclusive, transparent, ethical, safe, secure and trustworthy”. Commenting on the UK’s stance, the Chair of the Science, Innovation and Technology Committee, Chi Onwurah, expressed disappointment not only that the UK had not signed the declaration, but also that there is still no sign of a much-anticipated AI Bill, adding that “these developments together represent an apparent watering down of the UK’s previous stance. The motivations for this remain unclear, as are the national security and global governance concerns cited by the secretary of state in the House of Commons as having prevented the UK from signing the communique. The government needs to urgently clarify its position and offer reassurance that a safety- and citizen-first approach to AI has not been pushed aside”.

To read more, click here.