Towards a Code of Ethics for Artificial Intelligence

This brief introduction sets a context for the subject of the book: the challenges that arise in developing codes of ethics for artificial intelligence (AI). To start with, an overview of some of the concerns about AI and current developments in AI. AI encompasses a wide range of applications, of very varying natures, which means that there will be complex debates about its benefits and risks. Some of the many and varied current initiatives concerned more specifically with AI and ethics are introduced briefly. There are diverse approaches to tackling ethical issues in AI which may complement the development of codes of ethics, and these too are briefly outlined.

Consideration of ethical questions in AI requires an understanding of some central questions and ideas in ethics. This chapter provides an introduction to ethics which will be used as a basis for further explanation of the particular questions about ethics in AI. Ethics is sometimes seen entirely negatively as restricting developments, but can also be used more positively as assisting in the promotion of beneficial activities. Standard normative ethical theories are outlined, but the focus here is on spelling out underlying questions in ethics. We need to understand that there are diverse accounts of the root need for ethics, questions about the nature of ethical concerns, and questions about who, or what, is the proper object of our moral concern, all of which need to be addressed in thinking about AI. There are also contentious questions about the nature of argument and justification in ethics, including questions about moral relativism , which are especially pertinent to the issue of developing codes of ethics, and which we will need to consider carefully. The issue of transparency in ethics parallels concerns with transparency in AI. Questions about the nature of moral agency and moral motivation are also of prime relevance to discussions of AI.

In developing codes of ethics for AI, it’s important to consider how the ethical issues concerning AI relate to other ethical issues. Some of the questions overlap with those relevant to other rapidly developing technologies, especially where those technologies involve questions about how we humans understand ourselves and our place in the world. Issues that concern both AI and other areas of science and technology include the difficulty of considering ethical issues where predictions about the future may be difficult to make; where technology is embedded in complex ways in society, the economy and culture; where the analysis and use of large amounts of data are at issue; and where complex, large systems give rise to difficult questions about the distribution and attribution of responsibility. Characteristic ethical questions regarding AI concern its typical enhancement or replacement of human agency; crucially, questions of agency are at the heart of how we understand ethics. AI, like other technologies, is often hyped up; it’s important to see through the hype to as realistic a picture as possible. Hype can distort our ethical thinking in dangerous ways, and some of the key hazards of hype regarding ethics are outlined.

This chapter outlines the features of the professional practice which lead to the necessity for codes of professional ethics, and which underpin the nature and typical content of such codes. There are a variety of codes and regulations regarding professional practices, which may serve different purposes. Members of a profession possess certain skills, knowledge and capacities that their clients and the general public typically lack. This creates a gradient of power and of relative vulnerability between the professional and others. Codes of ethics aim to mitigate the potentially deleterious effects, or the misuse, of such professional power. Codes of professional ethics may be backed up by hard or soft power. Since each profession deals with a certain area of endeavour, codes of professional ethics typically concern themselves with values, benefit and harms in relation to their own area of expertise. Nonetheless, there are general values underlying such codes, even if these are implicit. These may be hard to articulate and may indeed be controversial. The value of autonomy is examined as an example especially relevant to AI. Codes of ethics can only function effectively with both adequate institutional and societal backing. Understanding the history and context of development of codes of ethics is important to understand their underlying values, and especially where social and technological change is occurring. Codes of ethics may develop in response to catastrophe, in anticipation of problems, and with reference to codes of ethics in key areas, and all of these may give rise to problems. Codes of ethics may have certain failings, and in some cases even make a situation worse.

Having considered those aspects of professional practice which underpin the need for professional codes of conduct, and the nature of such codes, we turn now to consider how AI presents particular challenges for developing professional codes of ethics. Although many working in AI may be members of a professional body, work in AI may be carried out by those outside of any formal organisational setting. In addition, in AI, resources of money, technological capacity, and capacity over the dissemination of information, may be concentrated into certain hands. At the same time, the control problem in AI undermines the expertise gradient that in most professions gives them power and authority. This means that in AI, there is a particular problem with professional vulnerability in relation to their own products. Codes of professional ethics generally deal with two matters: the behaviour of professionals, and the impact of their products or services on clients and on the wider public. In AI, we have a third, additional, layer of complexity that must be addressed: the behaviour of machines.

As with other technologies, one of the tasks facing those concerned with AI is how to develop codes of ethics in the face of rapid and perhaps unpredictable technological change. We need to consider carefully what methodology we use in considering ethical questions: abstract principles and accounts of virtue based on the past may or may not be appropriate, and a consequentialism which assumes the assessment of harms and benefit may likewise be of scant use where social and technological change makes the identification of benefit difficult. Close attention to context and complexity will be needed. Technological change and social change are closely intertwined, and may impact on many of our key value concepts, including those which may not at first sight be carriers of social and cultural value; in considering AI, we must watch out for such effects. The complexity of the question of technological changes brought by AI is illustrated by briefly considering the impact of AI on employment and by asking questions about the value and meaning of work. Technological changes are likely to have global impact in at least some cases, and the implications of this for how we think about universal values such as human rights, for relativism, and for cultural diversity are briefly considered. Lastly, as one solution to the complex task of developing ethical guidance for AI, a proposal is to ensure diverse participation in discussions; what is meant by ‘diversity’ in this context, and why it is essential especially for AI, is discussed with reference to some recent research findings.

Those developing codes of ethics for AI must of necessity consider the ethical issues that AI presents. There are some common pitfalls and gaps in argument to watch out for here. A full treatment of this topic would take much longer, but this chapter simply aims to alert readers to some of the main traps to avoid. There is always a balance between abstract and concrete thinking in ethics. Work in AI and ethics may concentrate too much on the idea that what distinguishes humans is their intelligence, and subsequently, idealisation or oversimplification of what is involved in both human and machine agency may occur. There may be different expectations for human and machine agency which are present but not fully articulated. This can have concrete and deleterious impacts upon any ethical conclusions which are drawn. AI is used to enhance or replace human agency. This means we must pay attention to questions about the boundaries of human agency and ‘normal’ human functioning. There needs to be careful consideration of different cases, given the varying nature of AI. The impacts of AI may not be just on its immediate use, but further afield within complex social systems, and careful attention should be paid to this. Lastly, clarity of language and of definitions is frequently an issue in AI; common language may mask deep disagreement.

This final chapter makes some provisional suggestions for the development of codes of ethics based upon the discussion so far. This will be of necessity incomplete, but there is a need to contribute to ongoing debate. Any code of ethics needs to be embedded well into an organisation and its culture, and specific ways in which codes of ethics for AI might face problems are indicated. Procedures for drawing up and implementing codes need to take note of diversity of thinking style and of experience in participants. The problems of transparency inherent in the operation of some AI, together with the important public concerns about the impact of AI, means that maximising transparency and openness in codes of ethics, appropriate to a particular organisation, is highly desirable. Codes of ethics need to balance attention to abstract principles with specificity, especially in AI where application of ethical ideals must be translatable into concrete practice. Procedures for revision and critique of codes are essential. Ethical discussion leading up to codes of ethics, as well as the codes of ethics themselves, must include consideration of issues concerning boundaries of human functioning, which is a key issue in AI and which may be left out of some ethical debates. Particular attention to the implications of replacing or extending human agency, and impacts upon complex social systems, would be useful. Lastly, the Asilomar AI Principles are briefly discussed, as an example of a recent attempt to produce principles intended to stimulate debate and discussion about beneficial and ethical AI.