Title: Using AI to Align Economic Policies with Human Values
In a recent paper published in Nature Human Behaviour, researchers demonstrated that deep reinforcement learning (RL) can be used to find economic policies that align with human values. This is a key challenge in AI research, and the paper provides a proof-of-concept for addressing it.
The researchers created a simple game with four players, each playing over 10 rounds. They were given funds and had to decide whether to keep them or invest in a common pool. After the game, they voted for a referee who would redistribute the funds. One referee was a pre-defined policy, and the other was designed by a deep RL agent trained to maximize the votes of virtual players.
The results showed that the policy designed by deep RL was more popular than well-known baselines. The AI system learned to maximize the stated preferences of a group of people, ensuring that it aligned with human values. It redistributed funds based on relative contribution and rewarded more generous players.
This study harnesses the principle of majoritarian democracy, but further research is needed to understand how to balance the preferences of majority and minority groups in economic policies. This approach has the potential to ensure that AI systems produce human-compatible solutions.