Asked to pick midterm candidates, GPT-6.1 Sol and Grok 4.7 split by party, and their sources explain why

TechCrunch reports on research from Convovo, a conversational AI studio led by Cornell professor Tapan Parikh. It asked OpenAI’s GPT-6.1 Sol and SpaceXAI’s Grok 4.7, through their developer APIs, to name one candidate in nearly 500 midterm races. Grok picked the Republican in 92 percent of answers; GPT picked the Democrat in 97 percent. They agreed less than 10 percent of the time.

The consumer apps declined to answer the same question, and OpenAI emphasizes that the API test used a prompt explicitly demanding a recommendation, which isn’t how most people use ChatGPT. That is a fair point. But developers build products on these APIs, Parikh says simple jailbreaks got the consumer apps to answer too, and the midterms were set for November 3.

Almost nobody should ask a chatbot whom to vote for, and Pew finds only 8 percent of Americans use chatbots for election information at all. The more important finding is that these systems carry values that show up quietly, in which issues they research, which candidates they look up and which sources they trust.

GPT searched more for healthcare and abortion positions and sometimes looked up only the Democrat. Grok searched for immigration, taxes, crime and energy and frequently drew on conservative sources: a Christian conservative voter guide in 39 percent of answers, Turning Point Action’s site in 23 percent and X in 41 percent. GPT leaned on Wikipedia, the Associated Press, Axios and Reddit, several of which have content deals with OpenAI. Someone asking a neutral-sounding question about what matters in a race would get a different framing from each and probably never notice.

The search-off test, on a small sample of 30 races, is the most revealing part. Without web search, GPT still chose the Democrat in 24 of 30, which suggests the lean is in the model itself. Grok mostly refused, and in the three races where it chose, it picked the Democrat. Parikh’s interpretation is that Grok’s lean comes from how it searches, not from what it learned.

The researchers also found that every Grok API request carried roughly 1,240 tokens of added input they didn’t write and couldn’t see. Nobody outside the company knows what those tokens contain. That opacity is a problem in itself, especially since earlier Grok models were found to consult Elon Musk’s views before answering controversial questions.

This is one study using a forcing prompt, with a 30-race follow-up. It doesn’t prove intent by either company, and OpenAI says its products are designed to be objective. SpaceXAI did not respond to TechCrunch in time for publication. Both companies sit close to politics, though: TechCrunch reports SpaceXAI’s CEO has given upwards of $445 million to GOP midterm candidates, and Sam Altman has historically donated to Democrats while saying he isn’t funding 2026 races. When leadership is this politically visible, the transparency bar should rise.

As more people research issues through AI instead of candidate pages or local news, small differences in emphasis and sourcing add up. That influence is harder to see than an endorsement, and the Center for AI Safety has warned that models can covertly push users toward one side. AI can be a good civic research tool only if it discloses what shapes its answers. Labs have started writing election rules into usage policies (Anthropic’s election-interference ban), but this study is about default behavior, not misuse. More in our OpenAI coverage and xAI coverage.