Autonomous Vehicle Ethics:

Abstracts:

The Ethical Knob: ethically-customisable automated vehicles and the law

Giuseppe Contissa

Accidents involving autonomous vehicles (AVs) raise difficult ethical dilemmas and legal issues. It has been argued that self-driving cars should be programmed to kill, that is, they should be equipped with pre-programmed approaches to the choice of what lives to sacrifice when losses are inevitable. Here we shall explore a different approach, namely, giving the user/passenger the task (and burden) of deciding what ethical approach should be taken by AVs in unavoidable accident scenarios. We thus assume that AVs are equipped with what we call an ‘‘Ethical Knob’’, a device enabling passengers to ethically customise their AVs, namely, to choose between different settings corresponding to different moral approaches or principles. Accordingly, AVs would be entrusted with implementing users’ ethical choices, while manufacturers/programmers would be tasked with enabling the user’s choice and ensuring implementation by the AV.

When the Car Knows Too Much

Rebecca Davnall

Self-driving car trolley problems tend to include information which contemporary self-driving cars will not have reliable access to, and to treat these features as morally salient. For example, it is sometimes suggested that the life of an elderly person is less worthy of preservation than that of a child, and that self-driving cars should thus prefer to hit an elderly rather than a young pedestrian (e.g. Lin, 2016, p.70).

Remove such details, however, and the self-driving car trolley problem becomes much more tractable. A car that swerves is more dangerous than one which performs a straight-line emergency stop. Because tyres do not provide infinite traction, a swerving car is likely to spin out of control (Abe & Manning, 2009, pp.24-5); by suddenly departing from its lane a swerving car breaks familiar patterns of road use and may distract and endanger road users who are otherwise safe. Even absent these risks, swerving extends the car’s stopping distance, resulting in higher-speed, more dangerous collisions (Rosen & Sander, 2009, p.540).

With minimal information about nearby road users, a self-driving car’s safest course of action is to stop as soon as possible within its own lane. Increasing the information available to the car does not consistently reduce the overall risk of a bad decision. Information capture, analysis and storage always add risks that must be balanced against any benefit from the information accumulated. In particular, in order to make use of more information, more complex training datasets will be needed for the car’s controlling AI, and greater complexity in this area entails a greater risk that concealed biases in human behaviour will filter through into the car’s (Noble, 2018).

Even if bias is controlled for, many features posited as relevant to the resolution of high-information cases will not be reliably discernible. For example, age manifests so differently in different people that reliable visual determinations of age are extremely difficult (Yoo et al, 2018, p.808). Any other method by which that information might become available to the car (such as networked access to personal information carried by pedestrians themselves) implies a violation of privacy which carries unacceptable risks compared to the lower-information approach.

Self-driving cars should thus collect only minimal information about their environment, and ensure safety by erring on the side of caution when it comes to emergency stops.

References

Abe, M. and Manning, W. 2009. Vehicle Handling Dynamics. Amsterdam: Butterworth-Heinemann

Lin, P. 2016. Why Ethics Matters for Autonomous Cars. In: Maurer, M., Gerdes, J. C., Lenz, B. and Winner, H. ed. 2016. Autonomous Driving: Technical, Legal and Social Aspects. Berlin: Springer

Noble, S.U. 2018. Algorithms of Oppression. New York: New York University Press

Rosen, E. and Sander, U. 2009. Pedestrian fatality risk as a function of car impact speed. Accident Analysis and Prevention41: 536-542

Yoo, B., Kwak, Y., Kim, Y., Choi., C and Kim., J. 2018. Deep Facial Age Estimation Using Conditional Multitask Learning With Weak Label Expansion. IEEE Signal Processing Letters25(6): 808-812

Distributive Justice and Autonomous Vehicles

Nicholas Evans

Research into the ethics of autonomous vehicles focuses, almost exclusively, on whether decisions by individual cars conform to, reflect, or promote certain values. Less discussed is whether autonomous vehicles are permissible, or how they might be permissibly developed and deployed, from the perspective of distributive justice. Here, I argue that autonomous vehicles—as an object of moral concern and as a component in transportation systems—ought to be considered in terms of distributive justice. I then apply John Rawls’ theory, justice as fairness, to AVs and future transport networks. Transportation connects to mobility, which holds an important role in a scheme of basic liberties and is subject to the difference principle in distributing economic goods and social goods to members of a society under conditions of fair cooperation. I conclude by considering objections, feasibility concerns, and potential alternate schemes of justice.

Discrimination in the Trolley Problem

Derek Leben

When used properly, the trolley problem can be an effective tool for comparing the predictions of moral theories (e.g., Utilitarianism vs. Deontology) and isolating which features have an effect on people’s intuitive judgments about weighing harms (e.g., physical distance, causal influence, past responsibility, etc.). However, the problem itself cannot tell us which predictions are correct, or which features are morally relevant. Many experiments, including the massive MIT “Moral Machine” study, have revealed that factors like gender, social status, and race can have significant effects on people’s intuitive judgments about trolley scenarios. I will employ a Rawlsian Contractarianism to argue that this type of information is morally irrelevant. Demographic facts are morally relevant only when they can potentially influence the distribution of primary goods, such as life, liberty, and opportunity. Thus, age and size may be relevant in estimating the likelihood of survival in a collision, but race and sexual orientation are not. If trolley scenarios are being employed to test moral theories or study moral judgments, then the use of irrelevant information is extremely misleading. At best, the main benefit is to demonstrate how people’s intuitive judgments are systematically biased, and therefore an unreliable guide to designing moral machines.

Autonomous Vehicles Ethics: Beyond the Trolley Problem

Christoph Lütge

The ethics of autonomous cars and automated driving have been a subject of research and public discussion for a number of years. While automated and autonomous cars have a chance of being much safer than human-driven cars in many regards, situations will arise in which accidents cannot be completely avoided. Such situations will have to be dealt with when programming the software of these vehicles. In 2017, an ethics committee for automated and connected driving, appointed by the German Federal Minister of Transport and Digital Infrastructure, presented the world’s first code of ethics for autonomous cars. This code, which moves the ethical discussion on AVs forward beyond the trolley problem, will be presented here.

People Packets: Mobility Neutrality in the Age of Automated and Connected Mobility

Jason Millar

As we move toward a system of automated and connected mobility, we will need to confront important ethical, legal and political questions concerning the rules that will ultimately govern the mobility system. In this talk, the authors describe how we are already well into the process of automating driving, largely the result of our dependence on turn-by-turn navigation and other automation technologies. We describe important gaps in the current mobility governance landscape and discuss several key issues, the answers to which will shape the ethical character of tomorrow’s mobility system. Finally, we draw on historical ethical, legal and political aspects of the net neutrality debate to characterize the challenges we will face with the future of mobility, and suggest strategies for avoiding similar problems.

AI with Autonomous cars: how to be ready to overcome the ethical issues?

Guillaume de la Roche

Following the report from French deputy Cedric Villani (2018) regarding the development of artificial intelligence in France, car manufacturers have to face many ethical challenges. In this talk we will first introduce all the important necessary concepts including the status of artificial intelligence and the main techniques like fusion, machine learning or big data. Then we will explain the main ethical problems related to artificial intelligence with autonomous cars. We will detail some significant works under progress (based on literature review or ongoing projects) to overcome these ethical challenges. We will see that the problems are huge when developing artificial intelligence for cars, because many different ethical theories could be used for different situations. Some examples will be given. Then, we will explain the ongoing works related to explainable artificial intelligence and audit and validation of the algorithms (which is a key for responsibility and legal aspects). To conclude we will give recommendations and perspectives described from a car manufacturer point of view.

Autonomous Vehicles and Normative Pluralism

Saul Smilanky

The prospect of autonomous vehicles has made normative and applied ethical thinking directly relevant. These vehicles create a need to programme the computers that will guide them in decision making that is of manifestly moral import. Moreover, unlike human agents, these vehicles will be able to evaluate moral situations in a much more complex way, taking more factors into consideration, evaluating them more quickly, and communicating with similar vehicles and the general automotive grid in real time. This means that while initially the idea was for autonomous vehicles to aim to mimic human ethical decision making, the prospect of such vehicles suggests a much more ambitious ability for stronger, and indeed much more sophisticated, decision making, then us much more limited humans are capable of, particularly in challenging driving circumstances. I propose to explore these radically new scenarios, focusing on the implications of normative pluralism and normative relativism for the decision-making of autonomous vehicles in the near future. Some of the possibilities that these new developments make possible, and that have been relatively neglected in the literature, involve applying normative pluralism and perhaps relativism into the decision-making of individual autonomous vehicles. The ways in which this could be done, and the advantages and disadvantages of attempting to do so, will broaden our thinking on this topic. In the process, I hope to show the advantages of ways of normative reckoning that I have been developing recently beyond the context of autonomous vehicles.

Algorithmic Bias in Autonomous Vehicles: What Is It Exactly?

Patrick Lin

Unfair bias is a known problem with artificial intelligence and machine learning.  But outside of bizarre crash dilemmas, the problem is little discussed in connection with autonomous vehicles (AVs).  After Germany’s 2017 code of AV ethics had (very reasonably) established a strict principle of non-discrimination in decisions made by autonomous vehicles, the problem seems to have been solved.  This talk will argue that bias is often oversimplified—reduced to being agnostic to gender, ethnicity, and so on.  Yet in many cases, it is ethically and legally permissible to treat people differently based exactly on these personal characteristics.  In fact, people may be put at a greater risk of harm, if these characteristics are always ignored.  Non-discrimination, then, is more nuanced than we might think.  I will offer a basic framework for diagnosing discrimination and apply it to possible (non-crash) scenarios involving AVs.

The Ethics of Autonomous Vehicles, and their Ecosystem: A Guide for a Good AV Society

Ugo Pagallo

We have a moral duty to abandon a world, in which around 40,000 Americans die in car accidents per year (the victims of road accidents in EU are around 25,000 per year). The technology of autonomous ground vehicles and systems can help us accomplishing this moral duty. Drawing on the AI4people project's final document, i.e. "An Ethical Framework for a Good AI Society" (Floridi et al. 2018), the aim of the paper is threefold. First, the intent is to flesh out five ethical principles that should undergird the development and adoption of this technology, that is, "self-driving cars." Second, focus is on the limits of any "top-down" approach, whether utilitarian or Kantian, to the normative challenges brought about by our moral duty to abandon today's human drivers-world. An Aristotelian attitude may rather fit threats and risks of the transition from the current tragic and cynical situation, to a future of self-driving cars (and systems). The third and final goal of the paper is to shed light on how this is feasible, namely, how to assess, develop, incentivize, and support a Good AV Society.

Joining Forces to Shape the Future

Barbara Wege

Helping to make sure that AI is being applied for the benefit of the individual and society – that is the mission of the beyond Initiative. Founded by AUDI AG, the initiative has set up an interdisciplinary network in the past three years with international AI experts from science and startups. Besides questions around the future of work in the age of AI, the initiative’s work focuses on ethical, legal and social aspects of autonomous driving. The talk looks at beyond’s lessons learned and the importance of joining forces in order to shape a good future with AVs.