Free beta access. 50 free AI tokens each week. Optional top-ups available. See what is included
Opening Word document…

AI controls are inevitable, but will be gradual and incrementally ratchet up over time---they’ll start with liability and transparency, then move into specific applications, solving downside risk without imposing premature and ineffective regulation

Chris Reed 18, Professor of Electronic Commerce Law at Queen Mary, University of London, LLM from the University of London, “How Should We Regulate Artificial Intelligence?”, Philosophical Transactions of the Royal Society B, Volume 376, Issue 2128, 9/13/2018, https://royalsocietypublishing.org/doi/10.1098/rsta.2017.0360

Using artificial intelligence (AI) technology to replace human decision-making will inevitably create new risks whose consequences are unforeseeable. This naturally leads to calls for regulation, but I argue that it is too early to attempt a general system of AI regulation. Instead, we should work incrementally within the existing legal and regulatory schemes which allocate responsibility, and therefore liability, to persons. Where AI clearly creates risks which current law and regulation cannot deal with adequately, then new regulation will be needed. But in most cases, the current system can work effectively if the producers of AI technology can provide sufficient transparency in explaining how AI decisions are made. Transparency ex post can often be achieved through retrospective analysis of the technology's operations, and will be sufficient if the main goal is to compensate victims of incorrect decisions. Ex ante transparency is more challenging, and can limit the use of some AI technologies such as neural networks. It should only be demanded by regulation where the AI presents risks to fundamental rights, or where society needs reassuring that the technology can safely be used. Masterly inactivity in regulation is likely to achieve a better long-term solution than a rush to regulate in ignorance.

This article is part of a discussion meeting issue ‘The growing ubiquity of algorithms in society: implications, impacts and innovations'.

1. Introduction

It is hardly surprising that there has been a sudden interest in regulating artificial intelligence (AI). AI technology has moved from the research laboratory to become part of our daily lives with remarkable speed. We have seen the first fatal accident involving an autonomous vehicle [1,2], AI applications are analysing images to detect potentially cancerous cells [3] and numerous other implementations are in place or in the pipeline.

The introduction of AI technologies creates societal risks. Although AI technologies aim to augment or replace human decision-making, leading to fewer wrong decisions, there is no doubt that AI will still get it wrong sometimes. And the ways in which AI gets it wrong are likely to be very different from the ways in which a human would make mistakes. This feels dangerous to society. We want to know the kinds of risks we are running, and purely statistical arguments that AI makes us safer are not convincing to the wider population.

Good regulation would improve our perception of safety, and also our perception that humans remain in control. It could also mitigate any new risks which the use of AI creates. But bad regulation risks stifling the development and implementation of useful AI solutions, perhaps even without improving safety and control. Thus, we need to understand what regulation can and cannot do so that we can shape it appropriately. It is also important that those who produce and use AI technologies are actually able to comply with regulation, and that regulation does not stifle worthwhile advances in the technology. Outside specifically regulated sectors, the general approach of law and regulation is that innovation is freely permitted, but that those responsible must bear the consequences if that innovation causes certain types of harm. If our existing law and regulation can deal with AI innovation in that way, no immediate change is needed. The argument, if one exists, for requiring all those who adopt an AI technology to demonstrate that it achieves a higher standard of performance and reliability than other innovations has not yet been made out.

2. The problem

Fundamentally, the problem which regulation must seek to solve is that of controlling undesirable risks. For any truly useful AI technology, there is likely to be empirical evidence that it is more cost-effective and, ideally, more accurate at making decisions than the human-based solution it replaces. But that evidence will be based on comparison with the human-based solution, whose deficiencies are currently tolerated by society. An AI-based solution will have its own deficiencies, and these will be less acceptable if they produce wrong answers where a human would have decided correctly. Regulation ought therefore to focus on any new risks which the AI solution presents, recognizing that some of these risks will be as yet unknown.

Some commentators are so alarmed by the prospect of unknown risks that they have proposed the establishment of a general regulator for AI [4]. But, there are three strong arguments against introducing new, generally applicable legal and regulatory obligations at this moment.

First, any regulatory body needs a defined field of operation, and a set of overriding principles on the basis of which it will devise and apply regulation. Those principles will be based on mitigating the risks to society which the regulated activity creates. Until the risks of AI are known, at least to some degree, this is not achievable. Regulation cannot control unknown risks, and devising a regulatory mandate on the basis of speculative risks seems unlikely to produce successful results.

Second, lawmakers are generally unsuccessful at prospective regulation, particularly in technology fields. The history of legislating prospectively for the digital technologies is one of almost complete failure [5].

Finally, and most importantly, a regulatory regime which aimed to deal with all uses of AI technology would be impossibly wide in scope. The range of potential applications is far too diverse, and it would be foolish to apply the same regulatory regime to autonomous vehicles as to smart refrigerators which order groceries based on consumption patterns. Probably, there is no plausible, let alone compelling, reason to regulate smart refrigerators at all. A regulatory project of this kind would risk becoming a project to regulate all aspects of human life.

The better strategy is to approach the problem incrementally. Some of the risks likely to be posed by AI technology are already apparent, and legal or regulatory action can be taken now to deal with them. Others will make themselves known as the technology becomes more widely used and can be dealt with in the same way. At some point, it will become apparent whether specific regulation is needed, and if so the scope and focus of that regulation will be possible to devise. But at present, we are some distance away from that point.

Sign in to recut, update, or save this card.