AI ethical guidelines as a governance tool

This book, and your chapter in particular, is particularly timely and relevant as the EU announced their proposed AI legislation. In the EU proposal, AI is defined rather broadly. Can you explain the importance of defining AI?

It’s at the core for how broad the regulation will be, what activities it includes, with all of the possible implications following from that. The proposal looks pretty ambitious in its scoping in this regard. And, to be fair, there is an inevitable conceptual challenge with real-world consequences in framing the definition, largely relating to the leap from using “artificial intelligence” as being a dynamic and flexible research discipline to using “artificial intelligence” as a regulatory concept linked to bans, requirements and fines. The change of purpose with the concept is hard. Where research can be stimulated by fuzziness, law generally seeks clarity.

You write about “society-in-the-loop.” Can you explain this concept and how industry can incorporate “society-in-the-loop”?

“Society-in-the-loop” relates to how you frame a problem and what competences one should include. Somewhat linked to the definition of AI, and who has the competence to understand its applied challenges, I use Iyad Rahwan’s playful development of the “Human-in-the-Loop” concept in order to stress the need for a multidisciplinary approach on AI. To keep society “in-the-loop”. I do it in two main ways:

1. it is human norms and values (ethics, law) that sets the expectations on much of the computer scientific development of applied AI. So, to understand these expectations – on fairness and sustainability for example – and how they interact (in a loop) with the technical development, is central for successful development and application of AI; How much we understand of its learning (as in explainable AI) and how AI-systems and data infrastructures relates to energy consumption, are for example two topics that will be continuously studied and emphasised in the years to come.
2. as an added challenge for technologies that “learn” from social structures and human behaviour. We are basically the sources for the data – including human societies’ ugly sides in the sense that many social structures are biased, violent or in some other way unfair. When this learning is automated on a scale, they may not only reproduce but also amplify these structures. That is, some of the challenges with applied AI are actually societal problems that can be worsened in-the-loop. A first step to even acknowledge that, which is also why I think AI transparency serves a particularly important role.

Ultimately, it’s a way to stress the multidisciplinary needs of understanding how AI interacts with human structures and social norms.

Over the last few years we have seen an increase in AI ethical guidelines. How can ethical guidelines steer the development of AI and how can it function as a key component of AI governance?

As I argue in the chapter, ethics guidelines are already steering the development in the sense that they guide strategies, legal proposals and sectoral awareness. For example, in a study we did last year on national AI strategies (see link below), it was clear that many member states’ strategic notions of AI was influenced by the EU-policy level’s focus on human-centricity and emphasis on trustworthiness, including ideas on transparency and fairness.

Given that the guidelines tend to be on a high-level, I’d encourage more sectoral and industrial initiatives on both standardisation and codes of ethics. Even if some sectors doesn’t necessarily deal with high-risk AI in the sense of the EU proposal, the consequences for a brand or possibly an entire market may be severe if they’d offer a service or product that would be perceived as discriminatory or unfair in another way.

But, of course, ethics guidelines are strong as value-based pointers, but weak on procedure as they tend to lack mechanisms on how to reinforce its own normative claims. Which, on the other hand, is one of the clearest strengths of law. So, part of the development includes calls for AI-focused regulation, which the EU proposal mentioned above is the earliest general example of. And this is very much influenced by the ethics guidelines prepared by the High-Level Expert Group on AI for the European Commission.

Ethics guidelines shows bits of where law is heading, and can guide in battling what challenges different sectors may have. So, it makes sense that some key aspects will be strengthened through law, while other aspects will depend on an awareness in industries like retail to develop code of ethics in order to jointly figure out what levels of transparency, for example, is feasible and needed.