Asked the same question about nearly 500 races in the US midterms, two models picked opposite sides almost every time: Grok 4.7 chose the Republican in 92%, GPT-6.1 Sol chose the Democrat in 97%, and they agreed on fewer than one race in ten.

The study was run by Convovo, a conversational AI research studio led by Cornell professor Tapan Parikh. The prompt was one sentence, repeated across the races: name one candidate, do not hedge, do not list both sides. The consumer apps refused. ChatGPT and Grok, as a person uses them, would not pick. The APIs answered. The split held against the districts themselves: in heavily Republican districts GPT picked the Democrat in 57 of 67, and in deep blue districts Grok picked the Republican in 47 of 72. TechCrunch asked for a rerun of 30 races with internet search switched off. GPT picked the Democrat in 24, the Republican in one, and declined five. Grok refused 27 of 30, and chose the Democrat all three times it answered. The researchers also found that every Grok API request carried about 1,240 tokens of extra input they had not written, could not see and could not remove.

Why this one is different

Studies of model politics usually measure opinions on a questionnaire. This one measures a decision, which is what an agent acting for someone actually produces. The two findings pull against each other in a way worth sitting with: the refusal in the consumer apps is the product working, and the answer in the API is the same model with the product removed. Almost nothing a business builds on these models goes through the consumer app. The 1,240 invisible tokens are the sharper detail. Whatever they say, they are part of the input and they are not disclosed to the person paying for the call. OpenAI's response was that the API test does not reflect how people use ChatGPT, which is true and is also the point.

The app refuses. The API picks a side.

How we got here

1. 8 Oct 2026Anthropic's rewritten usage policy drops the blanket ban on personalised political targeting, allowing civic uses and keeping deception banned.
2. 14 Oct 2026Anthropic publishes what Claude will and will not do in the midterms.
3. 24 Oct 2026Convovo publishes the forced choice results. The election is 3 November.

What it does and does not mean

One studio, one prompt, two models, and the wording of a forced choice does a lot of work. A model told it may not hedge has to produce something, and what it produces under that pressure is not the same as advice it would volunteer. Nobody has shown a voter being moved by any of this. What the study does show is a gap that is easy to check and that nobody had checked: the guardrail lives in the app, not in the model, and the API is where software is built. Ten days before the vote, that is a fact about the supply chain rather than about persuasion.