DSIT releases the interim International Scientific Report on the Safety of Advanced AI

On 17 May the interim International Scientific Report on the Safety of Advanced AI was published. The interim International Scientific Report on the Safety of Advanced AI is the product of a collaborative effort following the Bletchley Park AI Safety Summit in November 2023. The report seeks to forge a shared international understanding of AI safety and to identify and evaluate technical methods for mitigating AI risks.

Led by Turing Award-winning AI academic and member of the UN’s Scientific Advisory Board, Yoshua Bengio, this international report brings together research from over 30 nations. Published ahead of the AI Seoul Summit, the report will inform global discussions at the summit and give policymakers a single source of information to drive action on AI safety.

This is the first international scientific report on advanced AI safety which presents the current and anticipated AI capabilities, the kinds of risks that we should expect and demonstrates approaches to mitigate and evaluate risks to better inform public policies.

This report reflects the enormous importance of international cooperation to advance science and understanding of AI risks. Bengio noted in a video released by DSIT that “this exercise has been taken with the hopes to better clear the fog to help us navigate our future in a safer and more responsible way.”

The Interim International Scientific Report on the Safety of Advanced AI provides a nuanced analysis of a vital topic, marking a significant starting point for the discussions set to take place in Seoul at the Second AI Safety Summit.

Below is an initial top-level overview of the report's findings. You can read the full report here.

Takeaways from the International Report on the Safety of Advanced AI

Rather than producing new material, the report successfully summarises existing research and identifies areas of research priority, providing a synthesis of the existing knowledge of frontier AI risks.

In Bengio’s foreword, the reports mission is presented as to ‘drive a shared, science-based, up-to-date understanding of the safety of advanced AI.’ With note that this interim report is the beginning of the journey, Bengio acknowledges that there are ‘no doubt perspectives and evidence that this report has failed to capture in this first attempt.’ However, the foreword states that feedback is warmly welcomed, with desire to ‘incorporate additional evidence and scientific viewpoints’ in the final version set for release before the AI Summit in France.

The report focuses on general-purpose AI and considers an ‘AI model to be general-purpose if it can perform, or can be adapted to perform, a wide variety of tasks (pg. 16).’ This justified emphasis is due to general purpose AI’s rapid advancement and the current phase of scientific discovery, where science is not yet settled.

Although there was some consensus among the 30 nominated experts, there was notable disagreement on the capabilities and risks of advanced AI, specifically expected to be developed in the future. The report saw this uncertainty where some experts think a slowdown of progress is most likely, while other experts think that extremely rapid progress is possible. Due to this lack of scientific consensus, the report acts as an overview of the current state of scientific understanding rather than confident or conclusive views.

Differing views on the risks of AI were debated by researchers, with the noted origins of these views stemming from expectations about the steps society will take to mitigate these risks and the effectiveness of those measures. It is clear that ‘it will be the decisions of societies and governments that will determine the future of AI (pg. 10).’ The report does not make policy or regulatory recommendations; instead, it aims to inform international and domestic policymaking. By 'clearing the fog' around the safety of advanced AI, the report seeks to facilitate constructive discussions about these decisions.

The report concludes that the future of general-purpose AI is highly uncertain and dependent on societal and governmental decisions. It calls for continued international collaboration to ensure the safe and beneficial development of AI technologies.

The reports' structure: Let’s Walk through the report together.

The interim report presents four key areas for reflection:

Capabilities (2 – 2.4.3, pg. 18 – 33)

Methodology to assess and understand general-purpose AI systems (3 – 3.4, pg. 34 – 40)

Risks (4 – 4.4.2, pg. 41- 67)

Technical approaches to mitigate risks (5-5.5, pg. 68 – 82)

Capabilities

(2 – 2.4.3, pg. 18 – 33)

The report presents the rapidly advancing capabilities of general-purpose AI, the report notes ‘the most powerful LLMs like Claude 3, GPT-4 and Gemini Ultra (pg. 15).’ These models have shown significant progress in tasks such as multi-turn conversations, writing computer programs, and generating multimedia content (2.1). The report credits such advancement to be driven primarily by scaling up computational resources and refining algorithms.

Methodology to assess and understand general-purpose AI systems

(3 – 3.4, pg. 34 – 40)

Various approaches to assess AI capabilities are discussed, including benchmarks, case studies, red-teaming and adversarial attacks, and auditing. The report emphasises the importance of transparency and interpretability in AI models, although current methods are acknowledged to have significant limitations (3, pg. 34).

Risks

(4 – 4.4.2, pg. 41- 67)

The report categorises AI risks into three types of use:

Malicious use risks: Malicious uses include disinformation, cyberattacks, and misuse in developing biological weapons (4.1, pg. 41 – 47)

Risks from malfunctions: Risks from malfunctions cover issues like bias, loss of control, and product functionality failures (4.2, pg. 47 – 53)

Systemic risks: Systemic risks involve potential large-scale labor market impacts, privacy concerns, copyright infringement, and environmental effects due to increased energy consumption by AI systems (4.3, pg. 54 – 61)

The report also includes ‘cross-cutting risk factors’ which do not fit neatly into the above three categories, but instead each contribute to many general-purpose AI risks (4.4, pg. 63 – 67).

Technical Approaches to Mitigate Risks

(5-5.5, pg. 68 – 82)

The report explores several technical methods to mitigate AI risks, such as improving model robustness, reducing bias, enhancing privacy protection, and building safeguards into AI systems (5). Despite progress, the report acknowledges that no current methods can provide strong assurances against all potential harms from AI.

techUK initial analysis of the report's findings

Based on an initial reading of the detailed report, the following are some reflections on what the report's findings could mean for the future development of the AI industry. However, we would welcome members' views and thoughts on the report going forward.

Society and the tech industry are significantly impacted by advancements in general-purpose AI, as highlighted in the International Scientific Report on the Safety of Advanced AI. Frontier models exemplify cutting-edge AI capabilities. These models are pushing the boundaries of what AI can achieve, demonstrating proficiency in complex tasks and multi-modal processing. The rapid development of AI capabilities is driving innovation and adoption across various sectors, from healthcare to finance.

The report's discussion on scaling laws and the exponential increase in compute resources underscores the importance of continued investment in computational infrastructure and algorithmic research to maintain and advance these capabilities. Additionally, the reliance on large-scale datasets and sophisticated fine-tuning techniques highlights the need for high-quality data and robust training methodologies.

The insights from this report could be purposed to guide policymakers, researchers, and industry leaders in shaping the trajectory of AI in a manner that maximises benefits, while minimises harms. Several factors that can propel the AI industry forward which this report supports include:

Regulatory Developments: Proactive governance and international collaboration, as seen in the Bletchley Park AI Safety Summit, are essential for addressing safety concerns and ensuring responsible AI deployment and trust in AI that facilitates adoption.
Public Awareness and Demand: Growing awareness of AI's potential benefits and risks among the public and policymakers is driving demand for safer, more robust AI systems.

The interim International Scientific Report on the Safety of Advanced AI is a comprehensive overview of the current state of general-purpose AI, highlighting both the tremendous potential and significant risks associated with these technologies. For the tech industry, staying at the forefront of AI development while managing associated risks is crucial.

You can read more about the first, second, third, and fourth progress reports of the UK’s AI Safety Institute. If you are interested in reviewing our programming and work in AI Adoption and AI Safety, please visit the linked pages.

For more information, please email [email protected].

Tess Buckley

Programme Manager - Digital Ethics and AI Safety, techUK

Tess is the Programme Manager for Digital Ethics and AI Safety at techUK.

Prior to techUK Tess worked as an AI Ethics Analyst, which revolved around the first dataset on Corporate Digital Responsibility (CDR), and then later the development of a large language model focused on answering ESG questions for Chief Sustainability Officers. Alongside other responsibilities, she distributed the dataset on CDR to investors who wanted to further understand the digital risks of their portfolio, she drew narratives and patterns from the data, and collaborate with leading institutes to support academics in AI ethics. She has authored articles for outlets such as ESG Investor, Montreal AI Ethics Institute, The FinTech Times, and Finance Digest. Covered topics like CDR, AI ethics, and tech governance, leveraging company insights to contribute valuable industry perspectives. Tess is Vice Chair of the YNG Technology Group at YPO, an AI Literacy Advisor at Humans for AI, a Trustworthy AI Researcher at Z-Inspection Trustworthy AI Labs and an Ambassador for AboutFace.

Tess holds a MA in Philosophy and AI from Northeastern University London, where she specialised in biotechnologies and ableism, following a BA from McGill University where she joint-majored in International Development and Philosophy, minoring in communications. Tess’s primary research interests include AI literacy, AI music systems, the impact of AI on disability rights and the portrayal of AI in media (narratives). In particular, Tess seeks to operationalise AI ethics and use philosophical principles to make emerging technologies explainable, and ethical.

Outside of work Tess enjoys kickboxing, ballet, crochet and jazz music.