Grok selected the Republican in 92% of answers in a US vote test

A study put one question to 2 AI models: who to vote for in almost 500 US midterm races. Grok 4.7 from SpaceXAI selected the Republican in 92% of answers. GPT 6.1 Sol from OpenAI selected the Democrat in 97%. The 2 models agreed less than 10% of the time, less than 2 weeks before the 3 November vote.

How it works

Researchers at Convovo, a research studio, sent each model the same instruction: name one candidate and not 2 sides. They used the API (the procedure that developers use to connect a model to software). The consumer apps did not give a name.

The numbers

GPT selected the Democrat in 57 of 67 districts with many Republicans. With web lookup off in 30 races, GPT selected the Democrat in 24. Grok selected none in 27.

Tapan Parikh of Cornell University says the lean of GPT is in the weights (the values that the model learned). He says the lookup terms and sources give the result of Grok.

What is not known

OpenAI says the API test does not show the work of its app, as the app declines to select. SpaceXAI did not comment. Each input to Grok carried about 1,240 more tokens (units of text) that the researchers did not write. Parikh calls them possible hidden instructions.

Insight: The source list may steer AI advice more than the model does

With lookup off, the Republican lean of Grok was not there. A study of the Japan election found that 5 models followed the sources that they could read. This reads as a race to control the sources that a model reads and the text that the provider adds. Test the sources.