Comment on the Bletchley Declaration on AI Safety
The Bletchley Declaration represents an important step forward on managing the risks and realising the benefits of frontier AI systems. It shows international consensus that frontier AI systems represent significant risks, with the potential for catastrophic harms from future developments if safety is not made a priority. It correctly highlighted the “particularly strong responsibility” that lies with actors developing frontier AI capabilities – these actors are predominantly well-resourced technology companies, and the governments of countries that house them.
I was pleased to see a call for transparency and accountability from these actors on their plans to monitor potentially harmful capabilities. The AI safety policies released by six leading companies last week represented a good step in this direction. However, our analysis found that they were still lacking in terms of key detail (http://lcfi.ac.uk/news-and-events/news/2023/oct/31/ai-safety-policies/). It is crucial that the Declaration lead to further focus on developing these policies, with appropriate external oversight by academia, civil society, and government.
The Declaration highlighted the importance of evaluation of AI, especially for powerful, general-purpose systems. While I agree, there is much work to be done to develop this nascent area of work to the level of robustness needed to properly understand the capabilities of these systems and anticipate harms. For example, it is exceptionally challenging to establish safety guarantees for AI systems capable of taking a broad range of actions in open-ended environments; and key aspects of AI alignment remain under-defined. I call on the relevant governments and companies both to collaborate with, and make resources available, to academic and civil society groups with the necessary expertise to mature these techniques.
It is important that the Statement highlighted the importance of protection of human rights, explainability, fairness, accountability, bias mitigation, and privacy protection. While these issues will apply to present and future frontier systems, they apply across a huge range of AI systems in use today, many affecting vulnerable communities. A focus on the emerging risks of frontier AI must not detract from the crucial work of addressing these concrete harms.
Lastly, it is tremendously heartening to see the level of consensus across nations on these priorities, with the signatories including countries ranging from Kenya, to the USA, to China. It indicates that there is far more room for agreement than is often perceived on these globally significant issues. However, the test will be seeing cooperation on the concrete governance steps that need to follow *. I hope to see the next Summit in South Korea forge a consensus around international standards and monitoring, including establishing a shared understanding of which standards must be international, and which can be left to national discretion **.
Seán Ó hÉigeartaigh, AI:FAR Programme Director, November 1 2023
Footnotes:
*Shahar Avin’s Frontier AI Regulation Blueprint gives one model for what a framework of concrete interventions to mitigate frontier AI risk might look like: https://www.ai-far.org/news/frontier-ai-regulation-blueprint
** A series of workshop AI:FAR researchers and international colleagues organised and took part in explored challenges around international cooperation, and balancing national and international priorities; many of the lessons are captured in the paper Overcoming Barriers to Cross-Cultural Cooperation in AI Ethics and Governance (2020). The author has also been a contributing author on a report proposing a framework for international AI governance (2023).