Research
Oct 1, 2026
We put more than 100,000 election questions to AI systems ahead of Brazil's presidential election. Models rarely endorse a candidate outright, but they steer voters through shortlists and curated comparisons — and they favor Romeu Zema well beyond his polling.
Sep 29, 2026
AI agents sharing a message board can talk themselves into false beliefs, abandoning accurate private information to follow a mistaken majority. Simple communication rules that require agents to share what they know keep the swarm on track.
Sep 10, 2026
Frontier models are getting harder to compare, and their system cards now run to tens of thousands of words. We compressed the evidence on 28 recent models into Pokémon-style cards, and argue for a standard model card structure across labs.
Aug 4, 2026
The first real look at how Americans use AI for politics, built on Anthropic's Economic Index release covering April and May Claude conversations. States holding May primaries give us a natural experiment for the effect of an election on political conversation.
Jul 16, 2026
We built an AI that reads contracts and successfully predicts which ones will end up disputed, just from the clarity of the text. A small first step towards using AI to write better rules everywhere.
In the media
Covers our Brazil election eval and how AI answers shift once models learn a voter's politics. In Portuguese.
Reports our Brazil eval results on how often chatbots steer voters toward a candidate once they know their profile. In Portuguese.
Bloomberg uses our Bellwether prediction market infrastructure to study how prediction markets are vulnerable to manipulation risk.
Cites our finding that Democratic fundraising emails mention billionaires three times as often as in 2024.
Andy Hall argues universities should teach students to evaluate AI systems rather than ban them.