There has been a recent explosion of interest in the field of Responsible Artificial Intelligence (aka AI Ethics). It is well understood that AI is having (and will continue to have) significant social implications as the technology currently exists. Concerns over bias, fairness, privacy, and labour market impacts do not require the arrival of some mythical Artificial General Intelligence (AGI) to be socially destructive. Organisations from around the world are publishing frameworks for the ethical development and deployment of AI, corporations are affirming their commitment to social responsibility by funding responsible AI research and creating review boards, and a political interest in AI has led to state guidelines from numerous countries. Despite this, to many observers, little feels to have changed. Algorithms with suspected bias continue to be used, privacy continues to be a major concern, and accountability is near non-existent.
The failure of progress in responsible AI stems from a mistaken approach which prioritises good outcomes over good behaviour. It is uncontroversial to say that bias is bad and that privacy is good, but what this means in practice is more contentious. By attempting to simplify the work of achieving good outcomes to “frameworks” or “principles” the work being done in the field risks bearing little fruit. Our understanding of how AI systems can lead to problematic social outcomes is inherently reactive, in that we respond to problems that can be documented. The goal of responsible AI, however, is to be proactive by anticipating potential harms and mitigating their impact. Checklists on what ought to be done can never achieve the full range of potential risks that responsible AI seeks to address, and as a result are inherently limited. Proactive concern with socially beneficial outcomes requires not just work on frameworks for ethical use, but the cultivation of virtuous technologists and managers, who are motivated to take the concerns of responsible AI seriously.
The importance of virtue is clear when we consider the failure modes of principles-driven approaches to Responsible AI: non-adoption and recuperation. Either principles will be ignored, or they will be contorted to serve the interests of the status quo.
The risk of non-adoption is currently the more pervasive problem facing responsible AI work. In the rush to develop principles and guidelines competing approaches fail to generate consensus, and as a result have low impetus for adoption. Where a set of guidelines have low implementation costs, they may be championed to little or no effect. Where the burden of principles is too high, nobody cares to use them. If work is too technically focussed, it’s hard to communicate what has been achieved. If work is too value driven, it’s hard to audit whether anything has been done and it has limited practical applicability.
Even if consensus were to exist and the ideal set of principles formulated, this would not in and of itself motivate adoption. A lot of work has gone into consensus building among ethicists and commoditising the various implementations of responsible AI tools such as differential privacy, but this alone does not make technologists or business leaders care. Regulatory approaches seek to create the right incentives for following ethical guidelines by penalising bad behaviour, but these run into huge cost barriers to enforce and are slow to develop and diffuse. Even if ideal enforcement mechanisms were discovered, they would fail to create the proactive concern with social outcomes that responsible AI practitioners desire. Good incentives are no substitute for good citizens.
The degradation of responsibility and civic duty that incentive driven adoption creates leads to the second failure mode of current responsible AI approaches, recuperation. Recuperation is the risk of sincere work being co-opted by those who are in power. The common cries of “ethics washing” and “ethics theatre” that dominate high profile efforts by corporations to respond to the concerns of responsible AI practitioners reveal that without a sincere motivation to care, principles can be reduced to PR gimmicks.
Technical jargon can bury sincere conversations about social consequences, making it seem as if lots is being done. Review boards that hold no actual power over decisions become poster children for corporate efforts. Internal memos may normalise a certain sort of discourse around responsibility, with little to no actual understanding of the content. Regulatory compliance may become a way to capture markets by increasing barriers for new entrants. Even without the malice associated with naked self-interest, recuperation can result from laziness. If ethical concerns are reduced to checklists one simply has to tick off, the letter of the law may be followed, but its spirit completely lost. This would require the field of responsible AI to remain in constant vigilance to identify new risks and formulate new approaches for everyone to follow. In doing so, the law would get larger, but adherence would decline.
None of this is to say that work on principles is inherently useless, but rather that good behaviour relies on good people. Principles are put into place by people, and they embody the character of those who implement them. In order to achieve the goals of the responsible AI community, making people care about being good needs to be the primary goal. This requires a cultivation of virtue.
Dating back to Ancient Greece the concept of the Cardinal Virtues has been a pillar in moral philosophy as prerequisites for living a good life. A virtuous person is one capable of doing good, and they provide a bedrock upon which the internal motivation to do the right thing rests. Traditionally, the Cardinal Virtues are Prudence, Fortitude, Temperance, and Justice. By focussing on cultivating these virtues among the practitioners who develop and deploy AIs, from researchers to ML engineers to data scientists to project managers to executives, the responsible AI community would be more successful in ensuring proactive social concern. Each of these virtues is essential, and none on their own is sufficient to guarantee prosocial outcomes.
The first of these virtues is Prudence, which can be roughly equated to foresight or practical wisdom. It is probably the only virtue that is emphasised in education and work environments today. A prudent person is able to judge what the right thing to do is at the right time. The ability to make rational judgements about what one can best spend their time doing, what is the best use of the resources at their disposal, and whether a risk is worth taking are the fruits of prudence.
In the AI context, prudent practitioners are able to understand the consequences of what they are building, accurately describe the limits of their programs, and evaluate the opportunity costs of solving different problems using AI or solving a problem using AI rather than more conventional methods (such as human labour or rules-based programs). All of these decisions are critical for evaluating whether the outcomes of a system are socially beneficial or harmful, and the process by which this understanding comes about cannot be reduced to a checklist.
The next virtue is Fortitude, which can be equated with moral courage or perseverance. Fortitude enables one to do the right thing no matter the personal consequences or the stigma associated with a course of action. There are fields where fortitude is emphasised, such as in the military and in nuclear power management, where there is a pervasiveness of uncertainty and the consequences of error are disastrous. A lack of fortitude among AI practitioners poses many risks as without fortitude those who notice a problem may be afraid to speak up, the voice of one’s conscience may be drowned out by a desire to go with the flow or to protect one’s career, and doubt may confuse one’s ability to make effective judgements. Cultivating fortitude goes beyond the education of what the right thing to do is that currently grabs the attention of the responsible AI community by providing practitioners with the tools to act on the right thing.
Then there is Temperance, which can be equated with self-restraint or moderation. Plato himself considered this the most important of all the virtues as it enabled one to be humble and avoid acting rashly. Temperance enables people to understand their own failings and critically examine their own whys for acting. This virtue has played a foundational role in most of the world’s great religions and cultures, though its popular emphasis ebbs and flows. Organisations such as the Boy Scouts seek to cultivate this virtue, as do philosophy and meditation. The risks of intemperance for AI are severe as it would allow personal desires, for recognition or power or profit or even pure intellectual curiosity, to cloud judgement. Without temperance practitioners would be less willing to acknowledge their own biases and limitations, and as a result may refuse to acknowledge the harms that could be caused by what they develop.
Finally, there is Justice, which encompasses concepts of fairness and charity. The scales of justice in classical iconography sum up the balance between selfishness and selflessness that defines the just person. The golden rule that one ought to treat others as one ought to be treated cuts to the heart of just behaviour. The concerns of justice are front and centre in the work of most responsible AI researchers, as they seek to ensure that AIs do not discriminate against the disenfranchised and that diversity among practitioners provides representation to all voices. The risks of being unjust are evident and well understood, that the development of AIs will prioritise the people building it or people like them at the expense of the whole. Despite this understanding, the development of the moral sense of justice is typically absent from most schools and workplaces, with a carrot and stick approach often being used to enforce concern with social justice.
The value of these virtues I hope have been made clear, not only in their necessity in creating the motivation to develop and deploy responsible AI, but also in their ability to help foster a good society filled with good people. Formation in the virtues, however, is quite uneven, with prudence and to a lesser extent justice being the primary focus, with fortitude and temperance being more niche concerns. Calls to make ethical reasoning core to AI education should make personal virtue their aim, providing the tools for students to reason about the particulars on their own. At the Rotman School of Management, my alma mater, the business ethics curriculum was recently redesigned, moving away from a more legalistic compliance curriculum to the study of Aristotle’s Nicomachean Ethics and a focus on the personal moral inquiry of students and their character. Should this be a more popular approach, the benefits would be immeasurable. But the development of virtue does not end in school, it is rather a life-long task.
To be trained in virtue is to be given the opportunity to develop the right habits and frames of mind. This often comes though having good routines and being faced with challenges that exercise one’s moral faculties. Some ideas for how workplaces can cultivate this include:
- Regular workplace exercises that have individuals and teams solve complex moral problems from day to day life would allow for critical inquiry of values and enable character development. Michael Sandel’s Harvard Justice course provides a lot of insight on how to organise such problems and engage people with sound ethical reasoning.
- Encouraging time for reflection and philosophical reading would allow workers to develop their moral faculties. Scheduled unstructured time is critical for children’s moral development, but it is often lacking in schools and homes. There is no reason to believe that providing it for adults would fail to bring about the same results in allowing for improved communication, empathy, and social concern.
- Interdisciplinary work environments, which not only have diverse teams, but have individuals go out of their comfort zone to work on tasks on which they are less skilled, would aid in character building. When the Apollo missions had infighting about what the ideal approach to landing on the moon would be, each competing team was asked to complete a research document arguing for the option they disagreed with. This approach enabled greater understanding of others and a concern for full knowledge of the circumstance. Adapting these lessons to more workplaces can help take individuals outside of themselves.
This is a non-exhaustive list as the ways to build virtue are myriad and there is no checklist for how to do it. The particular environment and the circumstances of the people in that environment call for tailored strategies. Helping managers and human resources staff understand the value of virtue development in the workplace and working with them to create easily implementable solutions for their organisations would lead to larger returns to responsible AI work than focussing solely on reactive solutions to identified problems.
There has been significant valuable research in the responsible AI community in developing tools to address important problems posed by AI, clarifying ethical questions, and circulating frameworks that allow for prosocial development and deployment. The major hurdles faced now are in adoption and workplace culture. To address these, attention needs to be paid on the character of people involved in AI, and the moral virtues that they possess.
Op-ed contributed by our researcher Ryan Khurana, who has worked in tech policy and is a graduate of UofT’s Rotman School of Management in the Master’s of Management Analytics program.