Alan Winfield, a British national, is the world’s only AI ethicist. As a professor in the Department of Engineering, Design, and Mathematics at the University of the West of England, his primary research focus is on “robot ethics”. He refers to himself as “a professional worrier” due to his daily contemplation of AI issues. However, unlike Stephen Hawking, who constantly worried about AI destroying humanity, Winfield proposes more practical solutions. In his view, while AI presents troubling ethical issues, some anxieties are unnecessary.
Interviewed by｜Li Feiran
Edited by｜Jin Yan
Illustrated by｜Chao Chunbin
“Personage”: Are you really the only AI ethicist in the world? How does an AI ethicist work?
Alan Winfield: While it cannot be completely confirmed, it is probably not wrong to call me that. Of course, there are many outstanding researchers in the field of AI, and I think the biggest difference between them and me is that I have included this in my job title.
AI has developed to a point where it brings many ethical impacts, involving social, economic, political, and environmental domains. AI ethics is a vast topic that not only concerns how AI affects each individual but also how it ethically impacts society and even the entire planet. The task of an AI ethicist is to discover and study the “unethical issues” brought about by AI development in these areas.
However, I want to clarify that there is actually no boundary between AI ethics and human ethics; in other words, the subjects I study are still humans—the creators of AI, which may include designers, developers, or maintainers.
“Personage”: Can you provide some examples of research in AI ethics?
Alan Winfield: In our laboratory, we have a very intuitive experiment—we let some small robots play the roles of humans, and they keep moving forward while training a robot to learn how to prevent them from falling into a hole ahead.
Initially, the protective robot performed well. When a small robot representing a human moved towards the hole, the protective robot would rush over and push it to the side. However, when two small robots representing humans appeared, the protective robot began to panic; it had to make a judgment—who should it save?
Sometimes, it insisted on saving one person while the other fell into the hole. There were also instances where it moved quickly and saved both. However, in a total of 33 tests, there were 14 instances where, faced with a moral dilemma, the protective robot spent a lot of time hesitating and could not make a decision, resulting in both individuals falling into the hole.
We refer to such robots as “moral zombies”; although they physically save others, they do not understand the reasons behind their actions. A very intuitive example for everyone is autonomous vehicles. Although it may take a long time before we see them on the road, they provide a good way to understand AI ethical issues—when an autonomous vehicle carries passengers and encounters pedestrians or other cars in a crisis situation where a collision is unavoidable, how should humans design the rules? Should it prioritize protecting the passengers in its vehicle, or minimize the total number of casualties in the accident?
In contrast, a related area of research is creating “moral AI”, meaning AI can make moral judgments and choose how to act. However, this is a very cutting-edge field, and currently, only a few researchers worldwide are engaged in related work, and it is still uncertain whether humans can create such systems. Therefore, we may have to wait many more years to see the birth of a “moral AI”.
“Personage”: What kind of “unethical behaviors” might AI exhibit?
Alan Winfield: Currently, many cases show that biased AI has been developed. Some AI systems exhibit discrimination when making judgments; for example, traffic monitoring AI tends to favor middle-aged white individuals while perceiving black individuals as more likely to violate traffic rules. The primary reason for this is that developers have introduced significant biases and errors during the design process, and all human biases have been learned and absorbed by the machine, which represents a moral crisis faced by AI development.
Facial recognition is a typical example in this regard. For AI, the quality of the training data you provide determines the quality of the AI it produces. If the training data comes directly from the internet, and that data itself is biased, the resulting AI will likely favor white individuals, especially white males.
Of course, the AI that triggers a series of “unethical issues” is not limited to AI robots; it also includes AI systems.
In the political realm, we have noticed that AI systems like Facebook have been used to influence election outcomes. To a large extent, this is a form of targeted, unethical advertising. However, it is not Facebook itself that is unethical; it is the human behavior of manipulating election results using AI that is unethical.
Similarly, in the economic field, I can think of “alternative trading systems”, which are AI systems that can automatically buy and sell stocks or other securities based on certain rules. I am not a financial expert, but I have noticed that their behavior led to the Flash Crash, specifically the “flash crash” event of 2010.
To me, actions that cause harm are unethical. A very easy-to-understand unethical case is a computer virus—a very small AI program designed to cause harm, which is a clear and intentional form of unethical behavior.
“Personage”: What gives robots their morality? What factors determine whether AI is moral?
Alan Winfield: In fact, this statement should be phrased differently—most robots are not amoral; they are simply “morally irrelevant”. There is no distinction between moral and amoral; it is the way humans use them that places them in ethical dilemmas.
Take cars as an example: when you drive to help others, such as taking your grandmother to the supermarket or taking a patient to the hospital, the car is a morally good machine. But a car can also be used for evil; if you are a malicious person, you can use it to rob a bank or commit murder.
An AI developed by a moral developer is clearly more ethical than one created by a developer with lower moral standards. However, for a robot to autonomously make ethical judgments and know how to choose in ethical dilemmas, much more is required than just a moral developer.
To this end, I have participated in developing a series of robot moral standards. Among them, BS8611 is the world’s first robot moral standard; it is more like a “toolbox for engineers”. This guide provides a series of basic principles for AI design, such as robots should not be designed primarily to kill or harm humans, and every robot should have a responsible person who is accountable for the robot’s actions. You can use it to conduct a “moral risk assessment” of your system, thereby minimizing the ethical risks that AI may bring.
“Personage”: If companies apply these moral risk standards, will it affect their profitability? Are you confident in persuading them to sacrifice profits for greater morality?
Alan Winfield: This may affect direct profits, but in my view, a company that can actively take on moral responsibility will find it easier to achieve commercial success. For example, I refuse to use products from companies with low ethical standards, even if they save me money or provide convenience. At least in the UK, there are more and more consumers like me.
“Personage”: Many scientists hold a negative view of AI; for instance, Professor Stephen Hawking once said, “Developing complete artificial intelligence could lead to the extinction of humanity. Once they are designed, they will self-start and rapidly self-update. Humans, limited by slow biological evolution, cannot compete with them and will ultimately be replaced.” Do you also have anxieties regarding the ethical issues of artificial intelligence?
Alan Winfield: I have mentioned in an interview with the BBC that I believe Professor Hawking’s viewpoint is meaningless. The issues he discusses are based on a small probability that follows a series of chain reactions. Artificial intelligence is far from being that intelligent, and the problems he mentioned are not inevitable. In contrast, we should consider more realistic anxieties, such as the wealth and gender inequalities brought about by AI.
Currently, the wealthiest individuals in the world, this generation of billionaires, are becoming richer due to AI, such as Jeff Bezos of Amazon. I believe a more moral AI should benefit society as a whole, rather than a small number of super-rich individuals.
On the other hand, there is not enough female participation in the AI field, especially at mid to senior levels. The development of AI does not accurately reflect the needs of women.
“Personage”: Your daily work involves contemplating ethical issues related to robots. As a human, have you encountered ethical dilemmas? How did you resolve them?
Alan Winfield: In fact, not just me, every decision we make involves ethical issues to some extent. You can also think about how you handle these situations. You think, care, love, and face them with a sense of justice and responsibility. This is also one of my strategies for finding solutions to ethical dilemmas: when you need to make a decision, remember to have a sense of justice and responsibility.
Click“Read the original article”
to purchase this issue of “Personage”
