Thinking Machines: The Ethics of Self-Aware AI

Introduction

As artificial intelligence continues to improve, the possibility that it might one day achieve self-awareness has begun to attract increasing attention. This chapter discusses the ethical questions around such an advance, in terms of both how to ensure that self-aware machines will act in the interests of human beings and what rights such machines could have themselves.

The aims of this chapter are to describe:

• •

  General principles of AI ethics.

• •

  The likelihood of self-aware machines being created.

• •

  The implications of imposing pro-human ethical constraints upon them.

• •

  The critical need this creates for an objective ethical system.

• •

  A proposed objective ethical system, including its implications for the ethical relationship between humans and artificial minds.

• •

  The further implications this holds for future research in both ethics and computing.

Background

Over the last seventy years, artificial intelligence has progressed from just an idea, through increasingly powerful systems enabled by exponential growth in computing power and storage, to the present day when automated conversation systems are common and fully driverless cars are imminent (Anyoha, 2017; Council of Europe, 2020). As such systems have become more capable and widespread, and therefore their potential impacts on people have risen, there has been increasing interest in developing ethical principles for their use.

Nalini (2019) divided the ethics of AI into four categories: what it is, such as datasets and models, with issues including fairness, accountability and transparency; what it does, with issues such as safety and security; what it impacts, including automation and democracy; and what it can be, which raises future issues including superiority to humans and robot rights.

Nalini’s (2019) own discussion was limited to the first three, and, understandably, most thought on the ethics of AI has been concerned with applications available now or in the immediate future. Thus, it has centered on frameworks and oversight for collecting, using and sharing data in an environment of increasingly powerful machine learning and AI decision making, with a focus on how human oversight can promote general and organization-specific ethical outcomes (Jobin et al., 2019; Sandler & Basl, 2019; Siau & Wang, 2020).

Nevertheless, from the earliest days of AI there has been speculation that one day such machines might equal or exceed human intelligence. This has already happened in narrowly defined areas, with computers now able to exceed human abilities in first chess and then the even more difficult game Go (Anyoha, 2017). While these programs go beyond the fixed application of creator-defined solutions and extend into more general problem-solving algorithms, thus producing abilities beyond those of their creators, they still fall far short of general intelligence applicable to multiple domains of expertise (Bostrom & Yudkowsky, 2014). When, and indeed whether, such general intelligence will be achieved remains a matter of debate. For example, Grace et al. (2018) reported a very wide range of opinions amongst IT experts, whose predictions for how long it will take to achieve high-level machine intelligence ranged from only ten years to well beyond a century. The aggregate forecast was a 50% probability of such intelligence by around the year 2070 (Grace et al., 2018). A similar wide difference in opinion with a somewhat longer average estimate (year 2099) was found in more in-depth interviews with leading researchers reported by Ford (2018).

The uncertainty of this achievement coupled with the ethical issues we already face has led some to dismiss the importance of ethical thinking about artificial general intelligence (Ford, 2018). However, Ford’s (2018) interviews also make it clear that the uncertainty cuts both ways, and we might find ourselves having to deal with these issues sooner rather than later.