Rethinking the ethics of AI - and ours

Edmond Awad, Sydney Levine and Theos Evgeniou

AS investments and innovations in artificial intelligence (AI) grow, impacting all aspects of our lives from how we consume to improved medical treatments or to how we redesign our future transportation or energy systems, alarm regarding the potential risks these technologies can create is also becoming increasingly stronger. 

Already organisations such as the OECD or Partnership in AI are setting up AI incidents databases to track potential undesirable behaviours of these technologies as they get deployed and used. Risks range from straightforward safety ones - such as an autonomous vehicle creating an accident - to more “nuanced” ones, for example related to potential discrimination that AI can lead to, in domains ranging from credit scoring, hiring, facial recognition-based security systems, to personalised pricing, etc.

As a result, there has been growing discussions in recent years to better understand the behaviour of AI systems and ensure they are “ethical”. Never before has technology - not how people use it - been able to be unethical by itself. What sets AI systems apart from all other technologies is their ability to autonomously (or semi-autonomously) make decisions. With decision-making power comes ethical responsibility. Ensuring AI is ethical is therefore critical if we wish to build a future where AI not only creates enormous value but also does not lead to enormous value destruction. 

Achieving this requires “all hands on deck”: regulators, businesses (particularly tech companies but also all users of AI), civil society, and - given the complexity of the issues - academics. The latter play a critical role in this journey, as we are still lacking fundamental answers to critical questions. Never did we have to answer questions such as “how do we make sure we build technology that by itself behaves ethically?”. 

While “human ethics” has been a subject of debate for thousands of years, “machine ethics” is only becoming critical today. Answers to “machine ethics” questions will have great impact on our society going forward - and significant implications for everyone involved, from businesses that face new AI risks, to politicians and citizens who may see democracies at risk due to, well, algorithms. 

This is a key reason why a group of about 20 academics from global institutions have come together to study what we call “Computational Ethics” (see endnote). In a recent article that is published in the top journal Trends in Cognitive Sciences, we propose several key questions that we collectively need to study if we wish to ensure a future digital world that is safe and beneficial for all. 

How can we build a machine that can fairly balance the values of different stakeholders? How can we evaluate whether a commercialised algorithm reproduces human biases or whether it leads to a concerning change in how humans treat each other? How can we build machines that coexist with humans in shared environments, ensuring that neither of them converges to unethical behaviour?

We drew lessons from history - there are plenty of them - to propose an approach to these important questions. Going back a few decades ago, at the early stages of the AI revolution, academics again planted the seeds for what became today an “AI world”. 

SEE ALSO

The dark side of using AI to design drugs

Back then, people like the late David Marr and Massachusetts Institute of Technology’s Tomaso Poggio posed a similar question about another part of “human nature” and the human brain: how we can build computational systems, namely AI, that have similar visual capabilities as humans. It was the beginning of what led to all the marvels of today’s AI-based computer vision technologies: intelligent security access solutions, smart photo editing, medical image-based AI diagnosis, and the risks of potentially a surveillance economy. At the time, in the 1970s, academics set the foundations of what is called “computational neuroscience”. The impact has been profound: this led not only to better computer vision technologies, but also helped scientists - neuroscientists in this case - better understand how our brains, particularly our visual system that itself occupies a significant part of our brain, works. A perfect marriage of science and engineering with profound consequences over the following half century - and beyond. 

In this new article we aim for something as ambitious. If computational neuroscience helped us both better understand how human vision works and to build impressive AI-based computer vision technologies, can computational ethics help us both better understand human moral behaviour and build ethical AI systems going forward? 

Much like how computational neuroscience, that started almost 50 years ago, led to major practical AI innovations with profound socioeconomic impact today, so may computational ethics going forward. It can help us design solutions to important problems using the power of computation and AI. Consider the case of matching kidney donors to compatible patients. Practical approaches to this problem often assume a utilitarian solution: that is, matching should be done to maximize the number of recipients (ie matches) given a certain set of practical constraints (ie, biological compatibility, logistical feasibility, or geographic proximity). However, there can be multiple solutions that have the maximum number of matches. 

How should we decide among them? Such a complex multi-faceted question can be approached from different angles. Some work may take inspiration from arguments and frameworks laid down by ethicists about how such decisions should be made. Other work, by cognitive and social scientists, may instead focus on understanding how ordinary people evaluate the value of lives of others. These two research lines have different purposes and will likely result in different conclusions, but they can also inform each other. Yet, in real situations of kidney matching, decisions made on the spot may result in different outcomes. How do we reconcile these different angles together?

Ethically using AI to answer questions such as the one above requires that we develop new tools, processes and practices going forward. We propose that the answer is in formalising them in algorithmic terms. Formalising the problem and the ethical decisions would push us to better understand the principles and rules behind the decisions; it would require the decisions by clinicians to be studied, understood, specified; and it would force us to articulate and commit to specific consistent resolutions in such situations. 

These tools will not only allow us to develop a better AI future for the world but may also help us generate insights on human nature itself. Computational ethics can help us better understand millenia-old moral questions, much like computational neuroscience helped us better understand how our vision system works. For example, how do theories of ethics and of our moral judgment fare when applied to concrete specific cases? What should we do when these theories lead us to decisions that are inconsistent with our intuitions? 

One of the most profound and influential pieces of work, arguably largely defining our civilization, has been Aristotle’s treatises on ethics. This was written almost 2400 years ago - way before the digital revolution. Perhaps we are now in a position to better understand some of the deepest questions about human nature that have been discussed over the centuries - ironically because of today’s silicon-built AI systems. This is not only an intellectual discussion, but one that can have profound implications both in terms of technological and business innovations and regarding some of the most important ideas humanity has ever discovered: ethics, democracy, and many others that we need in order to build a safe and beneficial world.  

Edmond Awad is lecturer (assistant professor) at University of Exeter Business School, associate research scientist at Max-Planck Institute for Human Development, and Turing fellow at Alan Turing Institute.  

Sydney Levine is research scientist in Massachusetts Institute of Technology's brain and cognitive sciences department and in Harvard University's psychology department.

Theos Evgeniou is a professor at INSEAD; co-founder of Tremau; member of the OECD Network of Experts on AI; a World Economic Forum academic partner on AI; and advisor for the BCG Henderson Institute.

"Computational Ethics", published in the Trends in Cognitive Sciences journal, is co-authored by 20 authors across several disciplines and top universities. https://www.sciencedirect.com/science/article/pii/S1364661322000456