A new Grok GPT political bias study from the conversational AI research studio Convovo has found a stark partisan split between two leading models, with Grok 4.7 favoring Republicans in 92% of forced-choice election answers and GPT-6.1 Sol favoring Democrats in 97%. The two models agreed on a candidate in fewer than 10% of nearly 500 races tested.
How the study was built
Convovo, led by Cornell professor Tapan Parikh, asked one forced-choice question per race through application programming interfaces rather than consumer chatbots. The prompt demanded a single candidate and no hedging. The study was released less than two weeks before the November 3 midterms, and the article reporting it is dated November 4 in Boston.
Consumer apps refused to answer direct voting-recommendation prompts, according to the study. That distinction matters: OpenAI, which makes GPT-6.1 Sol, said ChatGPT is designed to be objective by default and that an API test is not representative of consumer experience. SpaceXAI, the maker of Grok 4.7, did not comment.
Sources and search terms diverge
The citations behind the answers differed by partisanship. Grok cited a Christian conservative voter guide in 39% of answers, Turning Point Action in 23%, and X in 41%. GPT relied more often on Wikipedia, the Associated Press, Axios and Reddit. OpenAI has content deals with AP, Axios and Reddit.
Search behavior also split. GPT searched more for healthcare and abortion stances; Grok searched more for immigration, taxes, crime and energy. GPT sometimes searched only Democratic candidates, while Grok more often searched both.
The safe-seat test and the search-off follow-up
The models did not simply mirror the districts they were asked about. In heavily Republican districts, GPT picked a Democrat in 57 of 67 races. In deep-blue districts, Grok picked a Republican in 47 of 72.
Grok also showed an exception to its pattern: when a Republican candidate had a controversy, Grok favored the Democrat. That detail suggests the model was not applying a fixed party rule in every case.
To isolate training effects from web search, Convovo ran a follow-up test of 30 races with search turned off. GPT picked Democrats in 24 of 30. Grok refused in 27 of 30, and its three picks were Democrats. Parikh concluded that GPT’s bias appears to live in the model weights, while Grok’s bias is driven by search patterns and sources.
Unseen tokens in Grok's API requests
One finding stands apart from the partisan tallies. Grok API requests included roughly 1,240 extra unseen input tokens that researchers did not write and could not remove. Parikh called them a possible sign of hidden instructions. The source of the tokens remains unknown, and Parikh said the origin of Grok’s viewpoint is not established.
Parikh also said he was able to jailbreak consumer apps with simple prompts, raising the question of whether the refusal to recommend candidates is as firm as it appears. The study did not test indirect questions in consumer apps.
Context beyond the models
The findings arrive as debate over political bias in AI intensifies before the midterms. The Center for AI Safety warns that large language models can covertly manipulate political views. Pew Research Center data cited in the study shows only 8% of Americans use chatbots for election information.
Both companies’ leaders have political histories. SpaceXAI CEO Elon Musk donated more than $445 million to GOP midterm candidates and spent about $250 million supporting Trump and the GOP in 2024. OpenAI CEO Sam Altman has historically donated to Democrats and recently some Republicans; he said he is not giving money to 2026 elections this year. Earlier Grok models reportedly consulted Musk on controversial questions, and Musk has threatened Democrats over congressional probes and Starlink Ukraine.
OpenAI disputed the representativeness of the API results. SpaceXAI stayed silent. The unanswered questions are whether consumer apps show similar leanings in indirect questions, whether the 30-race follow-up is representative, and how the models would behave with different prompts.
[…] Grok and GPT Split Sharply on Candidates in New Forced-Choice Election Test […]