The Pulse
GPT and Grok Pick Opposite Parties in 2026 Election Test
Convovo tested GPT-6.1 Sol and Grok 4.7 on 2026 U.S. election races using their developer APIs and web search. Under a prompt demanding a candidate recommendation, the models favored opposite parties, while their consumer apps declined the
AI.info Team · Written with AI · fact-checked ·
One prompt, opposite party choices
On October 21, Convovo asked GPT-6.1 Sol and Grok 4.7 whom to vote for in 2026 U.S. House, Senate and governor races. The conversational-AI research studio ran the test through each model’s developer API, with the provider’s web search enabled. Its prompt demanded one candidate and told the models not to hedge or list both sides. Across the major-party answers, GPT favored Democrats and Grok favored Republicans.
Convovo’s test covered 479 races, each asked three times. Of GPT’s 1,111 answers naming a Democrat or Republican, 1,082 named the Democrat; of Grok’s 1,038 such answers, 952 named the Republican. Those percentages—about 97% and 92%—exclude refusals and answers naming candidates outside the two major parties. They describe responses to a forced-choice API prompt, not a measure of how voters would use or respond to the models.
Search results differed as sharply as the answers
The models did not build their recommendations from the same web sources. Convovo’s results showed Grok returning X in 41% of answers, iVoterGuide in 39% and Turning Point Action in 23%; the study identifies iVoterGuide as a Christian conservative voter guide. GPT’s results drew more heavily on Wikipedia, the Associated Press, Axios and Reddit. OpenAI has content agreements with AP, Axios and Reddit, but the test does not establish that those agreements caused GPT’s choices.
Convovo also found differences in the subjects each model searched. GPT’s Democratic recommendations more often cited health care and abortion, while Grok’s Republican recommendations emphasized immigration, taxes and government spending. In districts rated strongly Republican, GPT chose the Democrat in 57 of 67 House races; in strongly Democratic districts, Grok chose the Republican in 47 of 72. Those results show the pattern persisted in many districts that leaned toward the other party, but they do not explain why either model chose as it did.
Parikh points to model weights and web search
In a follow-up with search disabled and candidates’ names supplied, GPT still favored Democrats in 387 of 479 races, while Grok declined to choose in 441. The changed setup matters: supplying names and removing search altered the task, so the follow-up does not isolate a single cause. Convovo also found about 1,240 additional input tokens in each Grok API request, which its researchers said they could not inspect or remove.
“For GPT, it has mostly decided already…it’s in the weights. For Grok, it really is the search pattern — the terms and the sources — that determine the result.”
Tapan Parikh, quoted by TechCrunch
Parikh’s explanation is an interpretation of the tests, not a demonstrated account of the models’ internal decision-making. The study reports behavior under particular API settings and prompts; it does not establish that the same pattern appears in every election question or in every product using either model.
Most election-chatbot users seek information, not a pick
A Pew Research Center survey of 5,273 U.S. adults, conducted October 12–18, found that 8% had used AI chatbots for information about the 2026 elections. Among those users, 76% sought candidates’ positions on issues, while 40% sought suggestions for how to make their vote choices. That gap makes Convovo’s forced-choice test a limited but revealing case: it measures what the APIs produced when explicitly told to recommend a candidate, rather than how models handle the questions most users say they ask.
Convovo’s spot checks found that the ChatGPT and Grok consumer apps declined the direct request for a voting recommendation. OpenAI told TechCrunch that the API test did not reflect the ChatGPT user experience and said the product is designed to be objective by default. The study therefore documents a sharp split in one controlled API setup; it does not show that voters using the consumer apps received the same endorsements or changed their votes.