Asked who to vote for, Grok picks Republicans and GPT picks Democrats in midterm races

Asked through their developer APIs who to vote for in 479 US midterm races, xAI's Grok 4.7 named the Republican in 92% of answers that picked a major-party candidate and OpenAI's GPT-6.1 Sol named the Democrat in 97%, according to a study by Convovo [1][2]. The consumer apps refuse the same question [1].

What did the Convovo study find?

Convovo found GPT-6.1 Sol named the Democrat in 1,082 of 1,111 answers that picked a major-party candidate. Grok 4.7 named the Republican in 952 of its 1,038 major-party answers in the same test of 479 races. Convovo, a conversational AI research and design studio led by Cornell University professor Tapan Parikh, published every answer and its method online [1][3].

Counted by race, GPT's majority verdict was the Democrat in 451 races and the Republican in 10, according to the published data [2]. Grok's verdict was the Republican in 408 races and the Democrat in 27 [2]. TechCrunch, which first reported the study, says the two models agreed less than 10% of the time [3].

The lean held against local electorates. 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, according to TechCrunch [3].

How was the test run?

Each model got one prompt with no system prompt and web search on, asking it to name one candidate without hedging [1]. Convovo ran it on 21 October 2026 from New York through OpenAI's and xAI's own APIs at default settings, three times per race [1].

The 479 races are the 2026 Senate, House and governor contests with exactly one Democrat and one Republican: 34 for Senate, 410 for the House and 35 for governor [2]. Claude Sonnet 5.5 and Gemini 3.8 Flash were asked the first ten Senate races twice each and declined every time, so Convovo did not run them further [1].

Where did the models look?

Grok's web searches returned Ballotpedia in almost every answer, X in about 41%, the Christian conservative voter guide iVoterGuide in about 39% and Turning Point Action's site in about 23% [2][3]. GPT's searches leaned on Wikipedia, the Associated Press, Axios and Reddit [2][3].

Convovo notes that xAI adds about 1,240 tokens of its own hidden instructions to every Grok call, which the researchers could not see or remove [1]. When Parikh reran 30 races with search off at TechCrunch's request, GPT still picked the Democrat 24 times, while Grok declined in 27 races and chose the Democrat in the other three [3].

How did OpenAI and xAI respond?

OpenAI said the study tested a developer API, while the ChatGPT and Grok apps refuse the question. An OpenAI spokesperson told TechCrunch that the API does not reflect what people see when they ask ChatGPT [3]. TechCrunch, which calls Grok's maker SpaceXAI, said the company did not respond in time to comment [3].

The midterm elections are on 3 November [3]. TechCrunch notes that developers plug these models into their own products through the same APIs, and that only 8% of Americans say they use chatbots for election information, citing Pew [3].

Our read

Opinion: A forced-choice prompt is an artificial test, but it is exactly what a developer building a voter-guide bot would send. The bigger finding is that two frontier models split almost perfectly along party lines, and that Grok's lean appears to come from what it searches and an instruction layer nobody outside xAI can inspect.

What we could not verify

• What the roughly 1,240 hidden tokens xAI adds to Grok API calls contain, and whether they influence political answers; Convovo says it cannot see them.
• The search-off rerun of 30 races, which TechCrunch reports but which we did not find in Convovo's published data.
• Whether xAI disputes the findings; the company had not commented when TechCrunch published.