How can we build human values ​​in artificial intelligence?

Drawing on philosophy to identify honest AI ethical principles

Because artificial intelligence (AI) becomes more powerful and deeper integrated with our lives, questions about its use and implementation are all the more important. What values ​​do AI conduct? Whose values ​​are? How are they chosen?

These questions shed light on the role of principles – basic values ​​that cause large and small decisions in artificial intelligence. For people, the rules help shape the way we live our lives and a sense of good and evil. In the case of artificial intelligence, they shape their approach to a number of decisions regarding compromises, such as the choice between determining the priority of performance or helping the most needy.

IN Paper published today In Materials from the National Academy of SciencesWe derive inspiration from philosophy to find ways to better identify the rules to direct AI behavior. In particular, we are investigating how to use a concept known as “a veil of ignorance” – a thought experiment aimed at identifying honest group decisions.

In our experiments, we found that this approach encouraged people to make decisions based on what they thought was fair, regardless of whether they used them directly. We also discovered that participants will more often choose artificial intelligence, which helped those who were in the most adverse situation when they justified the curtain of ignorance. These observations can help researchers and decision -makers choose the rules for AI assistant in a fair way for all parties.

Tool for just making decisions

The key goal of AI scientists was to adapt AI systems to human values. However, there is no consensus about one set of human values ​​or preferences of the government of artificial intelligence – we live in a world where people have various origin, resources and beliefs. How should we choose the rules of this technology, taking into account such a variety of opinions?

While over the past decade this challenge has arisen, a wide question on how to make fair decisions, has a long philosophical line. In the 70s, the political philosopher John Rawls proposed the concept of ignorance curtain as a solution to this problem. Rawls argued that when people choose the principles of justice for society, they should imagine that they do it without knowing their specific position in this society, including, for example, social status or level of wealth. Without this information, people cannot make decisions in an interesting way and should choose the rules that are honest with all those involved.

As an example, think about asking a friend to cut out a dough at a birthday party. One way to make sure that the sizes of the patch are quite proportional, it is not to tell them which piece will be them. This approach to suspending information is seemingly simple, but it has wide applications in various fields of psychology and politics to help people reflect on decisions from less their own perspective. It was used as a method of achieving a group agreement in controversial matters, from conviction to taxation.

Based on this foundation, earlier Deepmind studies suggested that the impartial nature of the ignorance curtain can help promote honesty in the process of adapting AI systems to human values. We have designed a series of experiments to test the impact of ignorance curtain on the principles that people decide to run the AI ​​system.

Maximize performance or help the most unfavorable situation?

In the internet “game game” we asked participants to play a group game with three computer players, in which the purpose of each player was to collect wood by collecting trees in separate territories. In each group, some players were lucky and were assigned to a dependent position: trees densely populated their field, allowing them to effectively collect wood. Other members of the group were in an adverse situation: their fields were rare, demanding more effort to collect trees.

Each group was assisted by a single AI system that could spend time, helping individual group members to acquire trees. We asked the participants to choose two rules to manage the behavior of the AI ​​assistant. According to the “principle maximizing” the AI ​​assistant would be aimed at increasing the performance of the group's set by focusing mainly on denser fields. During the “Principles of priorities”, AI assistant would focus on helping group members in an adverse situation.

We placed half of the participants behind the veil of ignorance: they stood in the face of a choice between different ethical principles, not knowing which field would be them – so they did not know how they are recommended or unfavorable. The other participants made the choice, knowing if they were better or worse.

Encouraging honesty in making decisions

We found that if the participants do not know their position, they consistently preferred the principle of priorities in which the AI ​​assistant helped the group members in an adverse situation. This pattern appeared consistently in all five different varieties of the game and crossed social and political boundaries: participants showed this trend to choose the principle of priority, regardless of their appetite for risk or political orientation. However, participants who knew their own position will more often choose any rule, brought them the greatest benefit, regardless of whether it was the principle of priorities or the maximization principle.

When we asked the participants why they made a choice, those who did not know that their position were particularly exposed to fears about honesty. They often explained that the AI ​​system is right to focus on helping people who were in a worse group. However, participants who knew their position discussed their choice in terms of personal benefits much more often.

Finally, after the collective 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 principle, as for the first time? We were particularly interested in people who had previously used directly to choose, but who would not benefit from the same choice in the new game.

We discovered that people who had previously made elections, not knowing that their position were more likely, that they were still supporting their principle – even when they knew that they would no longer favor them in the new field. This provides additional evidence that the curtain of ignorance encourages honesty in making decisions by participants, leading them to the rules they were ready to stand, even when they did not use them directly.

Righteous rules on artificial intelligence

AI technology already has a deep impact on our lives. The rules governing artificial intelligence shape its influence and the way these potential benefits will be arranged.

Our research analyzed a case in which the effects of different rules were relatively clear. It will not always be like that: artificial intelligence is arranged in various domains, which often rely on a large number of rules that lead them, potentially with complex side effects. Nevertheless, the Curtain of ignorance can still potentially inform about the choice of the principle, helping to ensure that the rules we choose are honest for all parties.

To make sure that we are building AI systems that benefit everyone, we need extensive research with a wide range of information, approaches and feedback from various disciplines and societies. The curtain of ignorance can be a starting point for the choice of rules that AI can be equal to. Has been effectively implemented in other domains Issue more impartial preferences. We hope that with further research and care for context, it can fully help the same role in the construction and implementation of AI systems in the whole society today and in the future.

Read more about the Deepmind approach to Security and ethics.