More on the depolarizing effects of current AI
A research paper examines how current AI systems may reduce political polarization rather than amplify it, challenging conventional assumptions about the divisive effects of artificial intelligence on public discourse.
Background
- A follow-up to prior research by AI alignment scholars on whether AI tends to "depolarize" political discourse. Finding: large language models (GPT-4, etc.) consistently produce more moderate, centrist text than humans on the same topics, across multiple languages.
- "Depolarization" = reducing extremity of political views. The authors argue this can be harmful: it may erase legitimate minority or marginal perspectives under a bland consensus.
- This sits in a debate: some worry AI amplifies polarization (echo chambers), others find evidence AI flattens disagreement. Key question: is this a bug (models trained to avoid offense) or a feature (deliberate moderation)?
- Authors are at the University of Lausanne and the AI Objectives Institute, in the "AI alignment" field, which studies how to ensure advanced AI systems act in line with human values.