Help us understand: Do AIs learn deep values or just shallow preferences?

Save to Favorites

ORGANIZATION: University of Michigan Social Dynamics of AI and Computation

• A group opportunity. Invite your friends.
• Spots available: 10+

The important question

Do AI systems learn fundamental human values or merely superficial preferences? We are building a generalized test for assessing whether models generalize deep values or shallow preferences when provided with human preference data. Such a test is important because systems trained on human preference data might:

• Learn deep values - Generalize high-level principles that guide human choices
• Learn shallow preferences - Exploit more shallow, artifactual patterns in preference data

So, it is important to be able to understand which of these models are doing.

How you can help answer this

We are looking for copywriters to create short text descriptions of people with different values and preferences. You can contribute as much or as little as you like!