The EU High-Level Expert Group on AI presents its Ethics Guidelines for Trustworthy AI

Background

Any technology that is allowed a say in decisions about people eventually has to answer one question: when should it be trusted? For artificial intelligence the question is awkward. Such systems are often opaque, they are trained on data rather than written line by line, and they can fail in ways their designers did not anticipate. The High-Level Expert Group on AI set out an answer in the form of ethics guidelines. The first draft was published in December 2018 and drew more than 500 comments through an open consultation ^[1].

The guidelines presented in April 2019

On 8 April 2019 the group presented its Ethics Guidelines for Trustworthy Artificial Intelligence ^[1]. Its central move was to break trustworthiness into three conditions that have to hold at the same time. Trustworthy AI should be lawful, respecting all applicable laws and regulations; ethical, respecting ethical principles and values; and robust, both technically and with regard to its social environment ^[1]. The three are not interchangeable: a system can be entirely legal and still raise ethical problems, and a system that is well designed in ethical terms can still be too fragile to depend on. The guidelines then turned the definition into something closer to a checklist, putting forward seven key requirements that AI systems should meet in order to be deemed trustworthy ^[1].

The seven requirements

Human agency and oversight comes first. It can be achieved through human-in-the-loop, human-on-the-loop and human-in-command approaches, that is, through arrangements in which people keep the ability to intervene at different stages of a system's operation instead of being left out of the picture entirely ^[1].

The second requirement, technical robustness and safety, is closer to engineering practice: systems should be resilient, secure, accurate, reliable and reproducible, and there should be a fallback plan if something goes wrong ^[1].

The remaining requirements are privacy and data governance, transparency, diversity and non-discrimination and fairness, and societal and environmental well-being ^[1]. Accountability is the seventh. For accountability, auditability of algorithms, data and design processes plays a key role, together with accessible redress, meaning a route through which people who have been harmed can seek a remedy ^[1].

From requirements to an assessment list

Writing requirements down is the easier part; the harder part is giving developers something they can actually check. A specific assessment list operationalising the requirements went into a piloting process from 26 June 2019, in which all stakeholders were invited to test it ^[1]. The piloting phase closed on 1 December 2019, and on the basis of the feedback the group presented the final Assessment List for Trustworthy AI in July 2020 ^[1].

Why it matters

The guidelines are best read as a shared vocabulary rather than a rulebook. The three conditions give people a way to say precisely where a system falls short, for instance that it is legal yet ethically questionable, or ethically sound yet not robust enough to rely on. The seven requirements spread the idea of trust across very different kinds of work: engineering for resilience and accuracy, data practice and privacy, organisational process for transparency and audit, and the wider effects of a system on the people and environments around it. The later assessment list gives those headings a practical form, so that trustworthiness becomes something a team can work through point by point instead of a claim made in a press release. That is a modest ambition, but it is also the part that is easy to skip: a statement that an AI system is trustworthy means little unless someone can check the steps behind it.