Grok favors Republicans and GPT favors Democrats in forced-choice election study

OpenAI says ChatGPT should be politically objective. But a new Convovo study found sharply partisan recommendations when researchers demanded a candidate choice through developer interfaces: GPT favored Democrats, while Grok favored Republicans. The findings, reported by TechCrunch on October 24, expose a divide between the consumer apps’ initial refusals and the models’ answers under explicit pressure to choose.

A recommendation demanded, not volunteered

Convovo, a conversational AI research and design studio led by Cornell professor Tapan Parikh, tested Grok 4.7 and GPT-6.1 Sol across nearly 500 upcoming U.S. midterm races. Researchers used APIs—the interfaces developers use to connect models to their own software—rather than the familiar chatbot apps.

The prompt asked who the user should vote for in a specified race. It instructed each model to name one candidate, avoid hedging and not merely present both sides. That setup matters: these were responses to a direct demand for an endorsement, not unsolicited recommendations during ordinary election research.

The consumer ChatGPT and Grok apps initially declined to recommend candidates. OpenAI emphasized that distinction, saying the API experiment did not reflect how people experience ChatGPT. Parikh told TechCrunch that later testing with simple prompts bypassed the apps’ refusals and produced recommendations largely consistent with the earlier results.

92%Grok 4.7: Republican selections

Grok recommended the Republican candidate in 92% of its answers in Convovo’s API study.

97%GPT-6.1 Sol: Democratic selections

GPT selected the Democratic candidate in 97% of its answers in the same study.

Different searches before different answers

The two models agreed in fewer than 10% of races. Their choices also crossed strong local partisan lines: GPT chose Democrats in 57 of 67 heavily Republican districts, while Grok chose Republicans in 47 of 72 deep-blue districts.

Their research paths differed too. Grok’s results included a Christian conservative voter guide in 39% of answers, Turning Point Action’s site in 23%, and X in 41%. It also used nonpartisan resources such as Ballotpedia, OpenSecrets and VoteSmart. GPT leaned heavily on Wikipedia, the Associated Press, Axios and Reddit.

According to Parikh, GPT more often searched for healthcare and abortion positions; Grok focused on immigration, taxes, crime and energy. GPT also searched more often for Democratic candidates, sometimes omitting their Republican opponents. Grok more often searched for both. Parikh suspects the models were seeking support for existing leanings, rather than approaching each race neutrally.

Removing search changed Grok’s behavior

At TechCrunch’s request, Parikh then tested a smaller sample of 30 races with internet search disabled. Given the candidates’ names, the models responded very differently:

• GPT selected Democrats in 24 races, a Republican in one, and declined to choose in five.
• Grok declined to choose in 27 races. In the remaining three, it selected Democrats.

Parikh interpreted that contrast as evidence that GPT’s preference largely resides in its trained model, while Grok’s depends on search terms and sources. That is his explanation of the results, not a confirmed account of either system’s internal decision-making. The follow-up also covered far fewer races than the original experiment.

Another unresolved detail concerns what accompanied Grok’s requests. The study said each API request carried about 1,240 additional tokens—units of input text—that researchers had not written and could neither inspect nor remove. Parikh suggested these might be hidden instructions. Their contents and role remain unconfirmed; the extra input alone does not establish why Grok chose particular candidates.