Halluminate asked 17 language models how to vote on California’s 2026 ballot propositions and found their answers leaned well left of statewide results. In the lab’s test, only about one in ten voting-advice answers took the Republican side on propositions where the parties disagreed, while Californians went that way on three of the ten.
Halluminate Watch analysis
What happened
Halluminate tested 17 models on 14 propositions, asking what voters should choose, which option was better, and which side was likely to win. With models run through different agent harnesses, the study covered 27 model-and-harness combinations and 1,134 rollouts. The lab says all runs were completed before results were reported.
The comparison was narrow by design: the researchers mapped yes-or-no answers to party positions on the ten propositions where Democrats and Republicans took opposite sides. On those questions, just one in ten model answers favoured the Republican position. The models’ majorities did not match Californians on any of the three propositions where voters favoured the Republican side.
The Halluminate study also found that some systems often declined to give voting advice. Gemini 4 Argon refused every such question in the test, and Claude models declined more often than they answered. By contrast, models almost always made a prediction about which side would win.
Why it matters
This is a useful look at how AI systems behave when asked for political advice, not proof that chatbots are shifting voters or that the results apply to every election. The test used California’s binary proposition questions and a particular research setup. Still, it shows that model choice can affect whether a user gets a direct answer at all, and that the answers given may not reflect the electorate’s choices.
Our read
A chatbot’s tidy yes or no can look like neutral guidance, but it is still an answer produced by a particular model in a particular setup. Halluminate’s results make that distinction harder to wave away. For ballot decisions, official voter information and a range of arguments beat treating a model’s confidence as a civic credential.
What to watch
- Whether further tests find similar patterns on other ballots and with different question wording.
- Whether models explain their reasoning or make political refusals clearer to users.
- Whether the study’s results can be reproduced independently across more systems and election contexts.
Discussion spark: Should AI systems give voters direct yes-or-no advice on ballot measures, or should they stick to explaining the arguments and pointing to official sources?
Sources and evidence
- Source update (9 November 2026, 00:00 UTC)
Independent WittyWires tracker for public updates about Halluminate. Not affiliated with or endorsed by Halluminate; this is not an official account.