Andy Hall describes research using privacy-protected aggregate data from Anthropic's Economic Index to study how Americans use Claude for politics. In the dataset as he presents it, politics accounts for about half a percent of US conversations and ranks near the middle of roughly 190 topics, with most political conversations producing information or analysis rather than recommendations.
Andy Hall argues that the effect of AI on political knowledge should be judged against realistic alternatives such as social media or no research at all. He is cautiously optimistic about AI as a political tutor but uses a Japanese voting-recommendation audit to show how paywalls and crawler restrictions can distort the sources models retrieve.
Andy Hall then examines personal governance agents that learn a user's preferences and act as delegates. He identifies progressively harder problems: eliciting preferences accurately, representing users on unfamiliar issues, deciding when to challenge a user, avoiding control by model providers or deployers, and resisting prompt injection or other manipulation from content encountered online.
Andy Hall also discusses political-neutrality controls, differences observed in evaluations of model responses to authoritarian requests, and the tension between advanced-model risks and civil liberties. He argues that policy proposals should reflect technical evidence, international competition and political feasibility rather than stopping at an aspirational outcome.
Watch on YouTube
