National University of Singapore

Just this week, Dame Wendy Hall and I have been working on the emerging field of AI Assurance and in our opinion, we think the first outcome of this dialogue should be some consensus on a conceptual framework that offers clear definitions for key terms such as 'human values', 'responsible AI', 'trustworthy AI' and 'AI governance' -- AI governance is still in a terminological quagmire. Without a clearly defined conceptual framework, developing a global understanding of AI governance will be immensely difficult. A second outcome should be a structured mapping of the various stakeholders, e.g., governments, industry leaders, academia, society/social interest groups/minority groups, etc., who will play an active role in the governance of AI and what their role will look like, including some idea of their responsibilities, across the global AI landscape. A conceptual framework that defines key terms, and a structural map of the stakeholders, including their roles, should be two outcomes of this dialogue. These outcomes are necessary to proceed with developing a global understanding of AI governance.

• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches

AI capacity building involves defining what AI governance means with respect to the various stakeholders. It is very had to build capacity if stakeholders do not possess a clear understanding of what AI governance means in their socio-cultural contexts. Social, economic, ethical, cultural, linguistic and technical implications of AI are the very reason why we need AI governance in the first place. We must investigate what these implications are with a view to mitigating and/or avoiding the current and possible future harms and dangers that AI technology can cause. Interoperability of governance approaches is extremely challenging because of the socio-cultural differences that exist between societies. Hence we need to investigate what these differences are in order to understand how they might hinder or prevent the interoperability of governance approaches.

Yes, the translation and operationalisation of human values and ethics principles within AI development is absolutely essential and fundamentally necessary for developing global AI governance. This is why we think AI assurance is now so important, because it brings stakeholders together with a view to merging socio-ethical methods and techniques with the technical development of AI systems.

The biggest governance gap of all - and this is where AI assurance comes in - is how to translate principles and values into concrete methods and processes that can be applied by industry stakeholders, e.g. tech companies and businesses in general. Translation is most challenging when trying to apply universal principles within different socio-technical contexts. This is where we think methods in academia such as 'computational specification' can be extremely useful for building ethics principles into the software development process with the aim of building ethical AI systems. Such methods are not well-known outside of academia, however. Academics should play a larger role in the socio-technical dialogues if only to communicate how certain methods work in theory, and how they can be applied - or translated - into practice.

One role that we think the AI Dialogue can play is by creating a map of the international landscape of stakeholders and clarifying how global AI governance operates within local societies and communities. If stakeholders can see where they stand in relation to others then dialogues between stakeholders can talk meaningfully about the implications that AI technology is having within localised settings. A large part of the problem at the moment is that no one quite knows *who* to talk to for addressing and resolving socio-technical harms, dangers and other problems.

We believe that the AI Dialogue should use the history of biomedical ethics as an example of a professional industry that once stood in a serious ethical crisis post world war 2 (not too dissimilar to what we are witnessing today with AI technology). Thanks to the creation of the Nuremberg Code in 1947 and a series of Codes that followed thereafter, the global medical profession organised itself around a common set of ethics principles and best practices that eventually became standardised across the globe in the form of ethical rules, industry regulations and medical laws. We think AI governance today is in a similar position to where the medical profession stood in the 1920 and 30s, one that lacked a cohesive global understanding of how medicine should be practiced across societies and cultures around the world. The lesson from history here is that we must avoid a Nuremberg Trial for the AI technologists and AI companies --- no one wants a major global catastrophe involving AI-powered autonomous robots and weaponry. We must take these lessons from history with a view to avoiding such catastrophe before it's too late!

We believe different stakeholders should highlight operational difficulties - tensions between differing values, that is - with or between different platforms and systems. The format could look like this: individual stakeholders raise ethical/legal issues with how certain platforms or systems are operating in their socio-cultural contexts. For instance, one stakeholder might find that 'privacy' settings within a system are raising ethical/legal concerns in their country or industry sector. The dialogue should be structured in a way that enables stakeholders to articulate their concerns with these platforms and systems, and allow for the target stakeholder to offer a response with how they could go about resolving the tension between them.

We think the global South is still underrepresented, especially the African nations. Beyond inviting the leaders of these nations to participate in the dialogues, another way to include them would be to run summits that focus on the major socio-technical problems facing the nations in the global South. For instance, many nations in the global South are not captured within AI development processes -- it is well-known that Western societies feature most predominantly in the data set used for training AI systems.

One very useful way of gauging the views of a large audience is to use the Slido app for asking questions and getting answers in real-time, including a word-map that highlights which answers are the most common across the room. The insights gained from this will give participants immediate insight into what other people are thinking and the degree to which certain issues are most concerning or otherwise prominent within the room.

One very effective way to address the challenges in AI governance is to run sandpits and workshops that bring stakeholders together to proactively work through the challenges with a view to finding solutions. We believe this is one of the most effective methods for doing the 'hard work' of deeply understanding the problems and making progress together.