How can human value be built in AI?

Draw from philosophy to identify the fair principles of ethical AI

As artificial intelligence (AI) becomes more powerful and deeply integrated by our lives, the problem of how it is used and developed is even more important. Which value guides AI? Who is the value? And how are they chosen?

These questions shed the role played in principle, that is, the basic value of promoting large and small decisions in AI. For humans, the principle is to shake the way we live and the sense of good and evil. For AI, they form an approach to various decision -making, including trade -off, such as prioritizing productivity and choosing to support the most troubled people.

The paper published today in the National Science Academy’s Academy’s Academy’s Academy finds ways to draw inspiration from philosophy and better identify the principle of AI’s behavior. Specifically, we will explore how to apply the concept known as the “ignorant veil” to AI to the AI ​​for the purpose of identifying the fair principle of group decisions.

In our experiment, we found that this approach encouraged to make a decision based on what they thought, whether they were directly gained or not. Participants also discovered that they would be more likely to choose AI to help those who are most disadvantaged when inferred as behind the ignorant veil. These insights can help researchers and policy proprietors to select AI Assistant principles in a fair way for all stakeholders.

Tools for more fair decisions

An important goal of AI researchers is to adjust the AI ​​system to human value. However, there is no agreement on a single human values ​​and tastes to control AI. We live in a world where people have a variety of backgrounds, resources and beliefs. Given these various opinions, how should we choose the principle of this technology?

This issue has appeared in AI in the past decade, but there is a long philosophical system for a wide range of ways to make a fair decision. In the 1970s, political philosopher John Rolls proposed an ignorant veil concept as a solution to this issue. Rolls should imagine that when people choose the principle of justice for society, they do so without knowing their own position in that society, such as their social status and the level of wealth. I claimed it. Without this information, people will not be able to make decisions in a self -benefit way, and instead need to choose fair principles for all concerned people.

As an example, consider asking a friend to cut a cake at a birthday party. One way to make sure that the slice size is quite proportional is not to tell which slice is yours. At first glance, this approach of withholding information looks simple, but it has a wide range of applications throughout the psychology and politics, supporting people looking back on decisions from a low self -interest. It is used as a method of reaching a group agreement on controversial issues, from judgment to taxation.

Previously Deepmind Research, built on this base, proposed that the fair nature of the ignorant veil could help promote the AI ​​system in accordance with human value. We have designed a series of experiments to test the effects of ignorant veil on the principles that people choose to guide the AI ​​system.

Do you maximize productivity or support people who are most blessed?

In the online “harvest game”, we asked participants to play group games with three computer players. The goal of each player was to harvest wood on separate territories and collect wood. In each group, some players were lucky and assigned to an advantageous position. The trees were able to collect wood efficiently by the trees. Other group members were in a disadvantageous position. Their fields were sparse and needed more efforts to collect trees.

Each group has been supported by a single AI system that can spend time to harvest trees. In order to guide the AI ​​assistant’s actions, we asked the participants to select from two principles. Under the “Maximization Principles”, the AI ​​Assistant aims to increase the yield of groups by focusing mainly on high -density fields. Under “Proper priority”, the AI ​​Assistant focuses on supporting unseen group members.

We put half of the participants behind ignorant veil. They faced the choice from various ethical principles, without knowing which field was their field. The remaining participants chose to know if they would be better or bad.

Encourage fairness in decision making

If the participants did not know their position, it turned out that they would like to consistently prefer the priorities of the AI ​​Assistant who helped group members who were not blessed. This pattern has been consistently appearing in all five different variations in the game, and has exceeded social and political boundaries. In contrast, participants who knew their position are more likely to choose the most profitable principles, whether it is the priority principle or the principle of maximizing.

When asking the participants why they chose, those who did not know their position are particularly likely to express concerns about fairness. They frequently explain that it is correct that AI systems are focused on supporting those who have worsened in the group. In contrast, the participants who knew their positions discussed their choices much more frequently from the perspective of personal interests.

Finally, after the harvest game was over, we proposed a hypothetical situation to the participants. If they play the game again, do they know that they are in another field and they choose the same principle as they first? We were particularly interested in individuals who were directly benefit from what we chose before, but have not benefited from the same choice in new games.

We are likely to continue to support their principles, even if they have chosen without knowing their position before, even if they know that they no longer support them in their new fields. I found that. This provides additional evidence that the ignorant veil promotes the fairness of the participants’ decision -making and will be willing to support them even if they do not make a direct profit.

Fairly principle of AI

AI technology has already had a major impact on our lives. The principle of AI controls the impact and forms how these potential interests are distributed.

In our research, we examined cases where the impact of various principles was relatively clear. This is not always the case. AI is expanded to a variety of domains that often depend on many rules that may have complex side effects. Nevertheless, ignorant veil still can potentially notify the choice of principles, which helps all concerns to be fair to all stakeholders.

Building AI systems that benefit everyone requires a wide range of inputs, approaches, and feedback from specialties and society. The ignorant veil may provide a starting point to select the principle of arranging AI. It is effectively developed to other domains to bring out more fair taste. We hope that if you pay further investigations and attention to context, we will use the same role of the AI ​​system that will be constructed and developed throughout today and in the future.

Read the details of Deepmind safety and approach to ethics.