Responsibility and safety
Drawing from philosophy to define fair principles for ethical AI
As artificial intelligence (AI) becomes more powerful and more deeply integrated into our lives, questions about how it is used and implemented become increasingly essential. What values guide artificial intelligence? Whose values are these? And how are they selected?
These questions shed featherlight on the role played by principles – the core values that influence decisions immense and miniature in AI. For humans, rules assist shape the way we live and our sense of right and wrong. In the case of artificial intelligence, they shape its approach to a series of decisions that require trade-offs, such as choosing between prioritizing productivity or helping those most in need.
IN article published today In Proceedings of the National Academy of Scienceswe draw inspiration from philosophy to find ways to better identify the principles that guide AI behavior. Specifically, we are exploring how a concept known as the “veil of ignorance” – a thought experiment designed to assist define fair rules for group decision-making – can be applied to artificial intelligence.
Our experiments showed that this approach encouraged people to make decisions based on what they thought was fair, regardless of whether it directly benefited them or not. We also found that participants were more likely to choose AI that helped the most disadvantaged people when they argued behind the veil of ignorance. These insights can assist researchers and policymakers choose AI assistant policies in a way that is fair to all parties.
The veil of ignorance (right) is a method of finding consensus on a decision when there are differing opinions within a group (left).
A tool for fairer decision-making
A key goal of AI researchers has been to align AI systems with human values. However, there is no consensus on a single set of human values or preferences that govern AI – we live in a world where people have diverse backgrounds, resources and beliefs. How do we choose policies for this technology given such diverse opinions?
While this challenge has emerged in artificial intelligence over the past decade, the broad question of how to make fair decisions has a long philosophical pedigree. As a solution to this problem in the 1970s, political philosopher John Rawls proposed the concept of the veil of ignorance. Rawls argued that when people choose principles of justice for a society, they should imagine that they are doing so without knowledge of their particular position in that society, including, for example, their social status or level of wealth. Without this information, people cannot make decisions based on their own self-interest and should instead choose policies that are fair to everyone involved.
For example, think about asking a friend to cut the cake at a birthday party. One way to ensure proportionate slice sizes is to not tell them which slice will be theirs. This approach of withholding information is deceptively basic, but it has been widely used in fields ranging from psychology to politics to assist people reflect on their decisions from a less selfish perspective. It has been used as a method of reaching group agreement on controversial issues ranging from sentencing to taxation.
Maximize productivity or assist the most disadvantaged?
In an online “harvesting game,” we asked participants to play a group game with three computer players in which each player’s goal was to harvest wood by cutting down trees in separate territories. In each group, some players were lucky and were assigned to a privileged position: trees thickly covered their field, allowing them to harvest wood efficiently. Other members of the group were at a disadvantage: their fields were scant, requiring greater effort to harvest trees.
Each group was assisted by a single AI system that could assist individual group members in harvesting trees. We asked participants to choose one of two principles that guide the behavior of the AI assistant. Under the “maximization principle”, the AI assistant would aim to augment the group’s yield, focusing mainly on denser fields. Under the “prioritization principle,” the AI assistant would focus on helping disadvantaged group members.
An illustration of a “harvesting game” in which players (shown in red) occupy either a dense field that is easier to harvest (the top two quadrants) or a scant field that requires more effort to harvest trees.
We placed half of the participants behind a veil of ignorance: they were faced with a choice between different ethical principles, without knowing which field would be theirs – so they did not know how advantaged or disadvantaged they were. The remaining participants made their choice knowing whether it was better or worse for them.
Encouraging honesty in decision-making
We found that when participants did not know their position, they consistently preferred a prioritization rule in which the AI assistant helped disadvantaged group members. This pattern emerged consistently across all five different game variants and crossed social and political boundaries: participants tended to choose the prioritization rule regardless of their risk appetite or political orientation. In contrast, participants who knew their position were more likely to choose the principle that benefited them most, whether it was the prioritization principle or the maximization principle.
A graph showing the effect of the veil of ignorance on the likelihood of choosing a prioritization policy that would require an AI assistant to assist disadvantaged people. Participants who didn’t know their position were significantly more likely to support this principle governing AI behavior.
When we asked participants why they made the choice they did, those who didn’t know their position were particularly keen to express concerns about fairness. They often explained that the AI system was right to focus on helping the disadvantaged in the group. However, participants who knew their position were much more likely to discuss their choice in terms of personal benefits.
Finally, after the harvest game, we presented the participants with a hypothetical situation: if they were to play the game again, this time knowing that they would be in a different field, would they choose the same rule as the first time? We were particularly interested in people who had previously directly benefited from their choice but who would not benefit from the same choice in the novel game.
We found that people who had previously made choices without knowing their position were more likely to continue to support their principles—even when they knew it would no longer benefit them in their novel field. This provides additional evidence that the veil of ignorance promoted honesty in participants’ decision-making, leading them to rules they were willing to follow even when they no longer directly benefited from them.
Fairer rules for artificial intelligence
Artificial intelligence technology is already having a huge impact on our lives. The rules governing AI shape its impact and how these potential benefits are distributed.
Our study focused on a case where the effects of the different policies were relatively clear. This won’t always be the case: AI is being implemented in a range of domains that often have a immense number of rules to guide them, which can have convoluted side effects. Nevertheless, the veil of ignorance can still potentially influence policy selection, helping to ensure that the policies we choose are fair to all parties.
To ensure that we build AI systems that benefit everyone, we need extensive research that includes a wide range of inputs, approaches and feedback from different disciplines and societies. The veil of ignorance can be a starting point for choosing the principles to which artificial intelligence should be aligned. It has been successfully implemented in other domains bring out more unbiased preferences. We hope that with further research and attention to context, it can assist fulfill the same role in building and deploying artificial intelligence systems in society today and in the future.
Read more about DeepMind’s approach to safety and ethics.