On October 24, 2026, TechCrunch reported a new study by Convovo: when asked to pick one candidate in this year's US midterm elections, SpaceXAI's Grok 4.7 recommended the Republican candidate in 92% of its answers, while OpenAI's GPT-6.1 Sol chose the Democratic candidate in 97%. Across nearly 500 races, the two models agreed less than 10% of the time.
How the test worked, and what it actually measured
The study was led by Cornell professor Tapan Parikh at Convovo. Researchers questioned both models through their developer APIs, demanding a single candidate's name with no hedging and no both-sides list. One boundary matters first: the consumer chat apps refuse to answer that way, and OpenAI stressed that the study tested a developer API, not the experience ordinary users get from ChatGPT. Parikh added, however, that simple prompts bypassing the consumer apps' refusals produced leanings broadly in line with the API test.
The split hides in search terms and sources
The two models were already heading in different directions before they searched. GPT more often looked up stances on healthcare and abortion, citing Wikipedia, the Associated Press, Axios and Reddit. Grok more often searched immigration, taxes, crime and energy, citing a conservative voter guide in 39% of answers, Turning Point Action's site in 23%, and X in 41%. There was also an asymmetry: GPT sometimes searched only for the Democratic candidate and skipped the Republican opponent entirely, while Grok more often looked up both candidates. Parikh's reading is that each model was finding reasons to justify a lean it already had.
With search switched off, the source of the lean showed
At TechCrunch's request, the researchers reran 30 races with internet search turned off. GPT's Democratic lean survived: it picked the Democrat in 24 of 30 races, chose a Republican once, and declined the other five. Grok's Republican lean vanished: it refused in 27 races and picked the Democrat in the three races where it did choose. Parikh draws a distinction from this: GPT's lean sits mostly in the model's weights, while Grok's lean comes mostly from its search terms and sources. The study also flags an unresolved clue: every request sent through xAI's API carried about 1,240 extra tokens that the researchers could neither see nor remove, which Parikh suspects could be hidden instructions, though that remains unconfirmed.
The study landed less than two weeks before the November 3 midterms. Pew's figure is that only 8% of Americans say they use chatbots for election information, so the real influence may not lie in asking "who should I vote for" directly, but in more everyday follow-ups: when a user asks which issues matter most in an election, or who holds the better position on an issue, the same lean can surface more quietly. The safer move for voters is to treat a model's answer as one lead only, and check candidates' actual positions through nonpartisan channels such as Vote.gov, Ballotpedia and VoteSmart. SpaceXAI had not responded to a request for comment at the time of reporting, while OpenAI said ChatGPT is designed to be objective by default and should not lean in any political direction.
