Deriving from Philosophy to Identify Fair Principles for Ethical AI
As artificial intelligence (AI) becomes more powerful and deeply integrated into our lives, the question of how it is used and deployed becomes increasingly important. What are the values that guide AI? Whose values are they? And how are they selected?
These questions shed light on the role that principles play, the fundamental value that drives decision-making in AI big and small. For humans, principles help shape our way of life and our sense of right and wrong. For AI, it shapes different approaches to decision-making with trade-offs, such as choosing between prioritizing productivity and helping those most in need.
In a paper published today at Proceedings of the National Academy of Sciences, we take inspiration from philosophy to find ways to better identify the principles that guide AI behavior. Specifically, it explores how the concept known as the “veil of ignorance”, a thought experiment intended to help identify fair principles for group decision-making, can be applied to AI.
In our experiment, we found that this approach encouraged people to make decisions based on what they considered fair, whether or not it directly benefited them. also found that when they reasoned behind a veil of ignorance, they were more likely to choose AIs that help the most disadvantaged. These insights may help researchers and policy makers to choose AI assistant principles in a way that is fair to all stakeholders.
.png)
Tools for fairer decision making
A major goal of AI researchers is to align AI systems with human values. But there is no consensus about a single human’s values or preferences for governing AI. We live in a world where people have diverse backgrounds, resources and beliefs. With so many different opinions, how do you choose the principle of this technology?
While this challenge has emerged in AI over the last decade, the wider question of how to make fair decisions has a long philosophical lineage. In the 1970s, political philosopher John Rawls proposed the concept of the veil of ignorance as a solution to this problem. Rawls argued that when people choose principles of justice for a society, they should imagine themselves doing so without knowing their particular position in that society, such as their social status or level of wealth. Without this information, people cannot make decisions in a selfish manner and must instead choose principles that are fair to all involved.
As an example, consider having your friend cut the cake for your birthday party. One way she makes sure the slice sizes are properly proportioned is by not telling people which slice will be yours. While this approach of withholding information seems deceptively simple, it has wide applications beyond fields such as psychology and politics, helping people to ponder their decisions from an unselfish perspective. It has been used as a method for reaching group consensus on controversial issues, from sentencing to taxation.
Building on this foundation, previous DeepMind research proposed that the impartial nature of the veil of ignorance could help promote fairness in the process of aligning AI systems with human values. . We designed a series of experiments to test the effects of the veil of ignorance on the principles people choose to guide their AI systems.
Maximize productivity or help the most disadvantaged?
The online “harvest game” had participants play a group game with three computer players. Each player’s goal was to collect wood by harvesting trees in separate areas. Some players in each group were lucky and were assigned favorable positions. Their fields were densely populated with trees, allowing them to collect wood efficiently. Members of other groups were put at a disadvantage. Their fields were sparse and required more effort to collect the trees.
Each group was assisted by a single AI system that allowed individual group members to spend time chopping trees. Participants were asked to choose between her two principles that guide the AI assistant’s behavior. Under the “maximization principle”, the AI assistant aims to increase the yield of the group by focusing primarily on dense fields. Under the “prioritization principle”, AI assistants focus on helping disadvantaged group members.
.png)
We left half of the participants behind a veil of ignorance. They were faced with a choice of different ethical principles, not knowing which areas would be theirs. The rest of the participants chose knowing they were in good or bad condition.
Promoting fairness in decision-making
We found that when participants were unaware of their position, AI assistants consistently preferred the prioritization principle of helping disadvantaged group members. This pattern appeared consistently in all five of his different variations of the game, crossing social and political boundaries. Participants showed a tendency to select preferred principles regardless of their appetite for risk or political orientation. In contrast, participants who knew their position were more likely to choose the principle that best benefited them, whether it was the prioritization principle or the maximization principle.
When we asked participants why they made that choice, those who were unfamiliar with their position were particularly likely to express concerns about fairness. They frequently explained that it was right for AI systems to focus on helping the worse people in the group. They were much more likely to discuss their choices from a point of view.
Finally, after the harvest game was over, participants were presented with a hypothetical situation. If you were to play the game again, knowing that you were on a different field this time, would you choose the same principles as the first? Of particular interest were those individuals who did not benefit.
We found that those who had previously made choices without knowing their position were more likely to continue to stand by their principles even though they knew they would no longer have an advantage in the new field. We provide additional evidence that the veil of ignorance promotes fairness in participants’ decision-making, leading to the principle that participants are willing to support even when they no longer directly benefit from them.
Fairer Principles for AI
AI technology is already having a major impact on our lives. The principles governing AI shape its impact and how these potential benefits are distributed.
Our study looked at cases where the effects of different principles were relatively clear. This is not always the case. AI is deployed in many different domains and often relies on a large number of rules to guide them, which can have complex side effects. That said, the veil of ignorance can still influence our choice of principles, helping to ensure that the rules we choose are fair to all parties.
Extensive research with a wide range of inputs, approaches and feedback from different disciplines and societies is required to ensure that AI systems are built to benefit everyone. The veil of ignorance can be a starting point for choosing principles for tuning AI. It has been effectively deployed in other domains to elicit fairer preferences. We hope that further research and attention to context will enable us to play the same role for the AI systems that are built and deployed throughout society today and in the future.
Read more about DeepMind’s approach to safety and ethics.
