It is a curious feature of the history of artificial intelligence (AI) that its successes have often been measured in games. Early programs were taught bounded problems like tic-tac-toe and draughts. These were novelties, but the defeat of chess world champion Gary Kasparov by IBM's Deep Blue in 1997 was presented as a threat to the intellectual dominance of humanity—comparable, perhaps, to the Cold War rivalry that had played out in the match pitting Bobby Fischer of the United States against the Soviet Union's Boris Spassky quarter of a century earlier. Another 25 years on, and the machines have beaten us in even more complex games, such as Go, as well as idiosyncratic ones, such as Jeopardy!.
Ludology offers a relatable measure of machine achievement. Yet it is curious because such games are, by definition, meant to be fun. Deep Blue, AlphaGo, and other AI systems have many qualities, but the ability to have fun is not among them. Another explanation might be that we focus on trivial measures because it makes the advances of our metal and silicon creations seem less threatening.
As Ryan Abbott's The Reasonable Robot makes clear, those advances will affect every aspect of human society and economy. That much we have heard before, from the World Economic Forum's breathless talk of a Fourth Industrial Revolution to prophetic warnings of the coming singularity. Abbott's contribution is to try to offer clarity in how law should respond.
Many attempts tend to follow Isaac Asimov, articulating rules to shape AI behaviour. The past five years has seen hundreds of lists, most failing to understand that Asimov's literary career was built on the fact that his ‘Three Laws of Robotics’ might have been wonderful in theory but did not work in practice. Abbott predicates his own approach not on what the rules should be so much as how regulators should develop them. The new guiding tenet, he argues, should be ‘AI legal neutrality’.
That tenet, appropriately enough, works in both directions: the law should not discriminate against AI nor against humans. At present it does both. AI systems, for example, will (eventually) be safer drivers than humans, but are prohibited from plying the roads. Humans, by contrast, may be better at customer-facing jobs, but are being replaced by machines to save on taxes. Neutral legal treatment, he argues, ‘would ultimately benefit human well-being by helping the law better achieve its underlying policy goals’.
Abbott is not claiming that AI systems should be treated as if they are persons, with rights or legal personality. His more subtle argument is that a presumption of neutrality helps clarify areas of the law in which AI should be treated more like humans, and where humans may sometimes need to be treated more like AI. This, he argues, will promote competition, improve safety, incentivise innovation, and reduce antisocial behaviour.
In addition to being a law professor, Abbot is a physician and a patent attorney. Both qualifications are on display in his research. In medicine, diagnostics requires the analysis of signs and symptoms to identify a disease, condition, or injury. A human doctor relies on his or her training and experience, which might include years of study and decades of seeing hundreds of patients or more. An AI system can be programmed with every textbook ever printed and millions of patient records. Misdiagnoses and accidents will occur in either situation, but how should the law respond?
Abbott warns that imposing strict liability on AI will discourage innovation. In areas where AI may ultimately be safer than humans—medical care, transportation—a negligence standard that treats AI ‘like a person’ would more appropriately weigh the costs and benefits of automation. A decade ago, Ryan Calo went further and argued that robot manufacturers needed immunity to remove the uncertainty of potential lawsuits. Neither approach has been embraced by any major jurisdiction, and yet there does not appear to have been any appreciable slowdown in research and development. (China may offer a counter-example, where its dominance in AI is often attributed to minimal restrictions on data collection and tort law is comparatively underdeveloped. Even there, however, data protection laws and limits on technology companies are being strengthened, rather than weakened.)
It is in the field of intellectual property that Abbott's work spills off the page and into the real world. Even as he wrote the book, his argument that AI already generates intellectual property was being presented not just to Cambridge University Press but in courts around the world. Again, the nuance of his position is important: he is not proposing that AI systems should ‘own’ their creations but highlighting a gap in the law that AI-created patentable inventions are not recognised for want of a natural person who qualifies as the ‘inventor’.
Working with computer scientist Stephen Thaler, Abbott named the AI system DABUS as the inventor of a functional container design and a type of emergency signal and applied for patents in various jurisdictions. Initial responses were not promising. The British Intellectual Property Office, the European Patent Office, and the US Patent and Trademark Office all rejected the application on formal grounds: the relevant legislation presumes a human inventor. Only the first of these cases had been decided when his book went to press. The second and third might have been disheartening.
Then in July 2021, a year after publishing The Reasonable Robot, Abbott's team prevailed in both South Africa and Australia within the space of a week. If he is correct and AI inventions become more common, these cases will be seen as landmark decisions.
While I agree with much of what Abbott argues, his guiding tenet of neutrality—as he himself acknowledges—cannot answer many of the regulatory problems thrown up by technology. The final section of the book includes a case-by-case examination of different areas of practice to determine whether specific sectors warrant change or not. But in areas such as tax, the lens he offers does help clarify policy questions that governments may be loath to confront. One might have a debate about whether AI should replace workers in a particular area of the economy, for example, but the more immediate argument should be whether governments ought to be encouraging that replacement through taxes that are higher on labour than on capital.
AI systems are destined to surpass humans in many areas, perhaps most areas. Abbott is sanguine about this, comparing it to long-distance running. Once of practical importance for transmitting messages or packages, the advent of telecommunications and transportation made distance running redundant. Yet people still compete in marathons. Similarly, people still play games like chess and Go—even if we know that a computer somewhere could thrash us at both.
The shifting reasons why we do such things are questions for philosophy and psychology. Here Abbott stays in his lane, noting in a deadpan aside that ‘it is beyond the scope of this book to establish the meaning of life’. Fair enough, but his emphasis on a purely utilitarian approach to AI regulation is itself a moral position.
Though well-written and admirably concise, the prose sometimes slides into techno-utopianism: ‘Once superintelligent inventive AI is run-of-the-mill, the financial costs of innovating will be trivial, the push to incentivize will be unnecessary, and future innovation will be self-sustaining.’ It is atypical for a lawyer (or a doctor) to be so optimistic, but perhaps he is looking forward to a future in which academics continue to write in the same way we continue to play chess: for the fun of it, rather than because there is anything truly new to say that the machines have not thought of already.
Until that time, The Reasonable Robot is an accessible and illuminating account of the problems AI poses for law—and those that law, for the time being, might pose for AI.
