←
Dan Thompson420 posts
Dan Thompson avatar

Dan Thompson

@danmthomp

Studying elections and political economy. Associate Prof of Political Science, UCLA

◷ Joined November 2016▧ 41 Photos and videos
Andy Hall avatar
⟳ Dan Thompson retweeted
Andy Hall@ahall_research·Nov 12

Different AI models rely on different mixes of news sources to answer the same political questions -- super interesting new research from my collaborators at @freesystemslab! As we build towards political superintelligence, there are going to be a bunch of important questions about how to incorporate news content. Before we figure all that out, we need basic facts about how things are working on the ground today. This research is super helpful for that, and I'm thrilled that @danmthomp and Sho Miyazaki are doing such great work on this topic.

📰 New research: Inside AI’s election news diet What news does AI read to get up to speed on politics? And what kind of news does it reward? Following interesting recent commentary on our Japan election study from @mattyglesias, @JerusalemDemsas and @TheStalwart, we ran a bigger test on the US midterms across 35 Senate races, 18 days, and 15,750 answers. What we found: - Each model has its own favorite news outlet for political news: AP for GPT, Yahoo for Claude, PBS for Gemini, and NYT for Grok. - GPT tends to behave like a civic guide (reading more election office and campaign sites), whereas Claude behaves more like a news summarizer (reading national news, Wikipedia, polls). - Grok draws on X and Gemini on YouTube for political information, far more than the other models we studied. Gemini cites YouTube in 55% of its answers and X shows up in 59% of Grok’s. - A huge share of AI's local news is provided by large nonprofit networks that very few people will have heard of. States Newsroom, a large network of over 40 local newsrooms, appears in 60% of GPT’s answers. - For many models, prediction markets are now a go-to source for polling information. Kalshi is cited more by Claude than any poll aggregator. Full study here: freesystems.substack.com/p/i…

Post attachment
Free Systems avatar
⟳ Dan Thompson retweeted
Free Systems@freesystemslab·Nov 12

📰 New research: Inside AI’s election news diet What news does AI read to get up to speed on politics? And what kind of news does it reward? Following interesting recent commentary on our Japan election study from @mattyglesias, @JerusalemDemsas and @TheStalwart, we ran a bigger test on the US midterms across 35 Senate races, 18 days, and 15,750 answers. What we found: - Each model has its own favorite news outlet for political news: AP for GPT, Yahoo for Claude, PBS for Gemini, and NYT for Grok. - GPT tends to behave like a civic guide (reading more election office and campaign sites), whereas Claude behaves more like a news summarizer (reading national news, Wikipedia, polls). - Grok draws on X and Gemini on YouTube for political information, far more than the other models we studied. Gemini cites YouTube in 55% of its answers and X shows up in 59% of Grok’s. - A huge share of AI's local news is provided by large nonprofit networks that very few people will have heard of. States Newsroom, a large network of over 40 local newsrooms, appears in 60% of GPT’s answers. - For many models, prediction markets are now a go-to source for polling information. Kalshi is cited more by Claude than any poll aggregator. Full study here: freesystems.substack.com/p/i…

Post attachment
Andy Hall avatar
⟳ Dan Thompson retweeted
Andy Hall@ahall_research·Oct 14

Today we're sharing some initial data on how Americans use Claude for politics, based on aggregated conversation data analyzed with differentially private methods. A couple interesting takeaways: *Almost no one asks AI how they should vote* About ~5.75% of conversations in the US touch on some kind of political or social issue in a recent one-week period, and less than 0.01% of conversations involve asking how to vote. *Political conversations have many goals* These conversations have wide-ranging goals, everything from research and writing assistance to learning, fact-checking, and arguing. We are continuing to make progress on improving how Claude talks about these topics---steadily increasing our performance on our even-handedness evals and partnering with a range of outside organizations to develop new measures of political bias---and I hope that sharing information about the shape of these conversations will help people outside the company evaluate AI's political bias, accuracy, and other criteria in ecologically valid ways. I'd love to discuss this with people who are interested, as it was a huge priority for me before joining. This was a big joint effort as part of our work preparing for the midterm elections @AnthropicAI. Full post is here: anthropic.com/news/2026-us-m…

Post attachmentPost attachment
Andy Hall avatar
⟳ Dan Thompson retweeted
Andy Hall@ahall_research·Oct 13

Political scientists and economists: we're looking for researchers with deep expertise to come work with us for a unique 4-month fellowship at @AnthropicAI. You'll work closely with us to evaluate, improve, and apply Claude's capabilities in your field. It's going to be extremely fun and interesting work, and definitely worth taking a short leave from your university for if you need to! More info here: job-boards.greenhouse.io/ant…

Post attachment
Dan Thompson avatar
Dan Thompson@danmthomp·Oct 1

Ultimately, we’re finding that describing the right incentives can help reduce cascades toward false beliefs, but the models we're using are not yet smart enough to figure out how to avoid these cascades without explicit instruction.

Dan Thompson avatar
Dan Thompson@danmthomp·Oct 1

We also studied whether, in addition to giving agents an incentive to care about how other agents performed, we could get fewer false belief cascades by encouraging agents to think about what information future agents would need. That additional nudge barely made a difference.

Dan Thompson avatar
Dan Thompson@danmthomp·Oct 1

The swarm-based incentive scheme reduced the share of swarms converging on the false belief from over 90% down to roughly 65%.

Post attachment
Dan Thompson avatar
Dan Thompson@danmthomp·Oct 1

Based on Jeff’s great insight, we ran a follow-up experiment where we told each agent they’d be rewarded based on the private beliefs of the last five agents in the sequence.

Dan Thompson avatar
Dan Thompson@danmthomp·Oct 1

Jeff Lewis had an excellent insight: by telling each agent they’d be rewarded based on the accuracy of their own belief, we didn’t give them a reason to give future agents the information they would need to make good decisions.

Dan Thompson avatar
Dan Thompson@danmthomp·Oct 1

The problem in our experiment was that the agents were not always sharing their private signals with others.

Dan Thompson avatar
Dan Thompson@danmthomp·Oct 1

Follow-up to our @freesystemslab experiment this week: Rewarding AI agents for the swarm’s performance rather than their own makes the swarm less likely to converge on a false belief. But even with the right incentives, agents don’t always figure out how to avoid getting it wrong

Post attachment
Ethan BdM avatar
⟳ Dan Thompson retweeted
Ethan BdM@ethanbdm·Sep 30

My view is that "alignment" is the wrong framing for safety with agentic AI. I think it is often conceptually confused and basically always flattens the conversation in an unhelpful way. Some non-expert thoughts from a social scientific point of view... 1/

Andy Hall avatar
⟳ Dan Thompson retweeted
Andy Hall@ahall_research·Oct 1

Latest Free Systems research: how are AI models making voting recommendations in Brazil? @leticiaauriemo and Luca Moreno Louzado ran a huge study to find out! Some highlights... --Top frontier models are quite stubborn at refusing to give direct voting reccos, consistent with a new Brazilian law forbidding reccos; other models are less stubborn --As you provide more info to AI about yourself and what you care about, it's easier to get reccos even when you shouldn't --These recommendations don't necessarily reflect the distribution of political preferences in society (not clear that they should, but it's an interesting benchmark). In Brazil, reccos for a less popular businessman candidate (Zema) occur at an unusually high rate. More generally, I think these results raise a lot of profound question about what we want out of AI for politics. Should we forbid AI from making direct recommendations, like Brazil has? How do you even define what a recommendation is? It's not clear, and in this study, the authors found AI models providing advice that arguably falls short of a direct recommendation but toes the line. What would optimal political advice look like? And how would we know it if we saw it? Incredibly excited to see this research, and to see Free Systems continue to scale beyond me and my work, and beyond the borders of the US! Check out the full piece here: freesystems.substack.com/p/w…

Post attachment