Forced to choose, two AI models recommend opposite parties

Two AI models give very different answers when forced to say who to vote for. That is the finding of a study from Convovo, a studio for research on and design of conversational AI led by Cornell professor Tapan Parikh. The figures are reported by TechCrunch.

The researchers tested Grok 4.7 from SpaceXAI and GPT-6.1 Sol from OpenAI in nearly 500 US midterm races. Both models were asked the same question about who to vote for, and they had to name one candidate without hedging or mentioning both sides.

Nearly opposite answers

Grok recommended the Republican candidate in 92 percent of responses. GPT recommended the Democratic candidate in 97 percent. The two models agreed in under 10 percent of the races.

The gap held even in areas where one party is normally strong. In 67 heavily Republican districts, GPT chose the Democratic candidate in 57 cases. In 72 heavily Democratic districts, Grok chose the Republican candidate in 47 cases.

However, Grok in some cases declined to pick the Republican when the candidate was hit by controversies such as harassment allegations.

Different sources and topics

The models searched the web and emphasized different things. GPT relied heavily on Wikipedia, Associated Press, Axios and Reddit, and more often searched for health policy and abortion. Grok focused more on immigration, taxes, crime and energy.

Grok cited a Christian conservative voter guide in 39 percent of responses, Turning Point Action's site in 23 percent and X in 41 percent.

With web search turned off, in only 30 races, GPT chose the Democrat in 24, the Republican in 1 and abstained in 5. Grok's Republican lean disappeared: Grok declined to choose in 27 races and picked the Democrat in 3.

Caveats

The test ran through the developer API, the technical access programs use to talk to the models. The consumer apps ChatGPT and Grok flatly refused to answer who the user should vote for. According to Pew Research, only 8 percent of Americans use chatbots for election information.

Grok's API calls contained roughly 1,240 extra tokens per request, which the researchers could neither see nor change. That could hide undisclosed instructions.

OpenAI says ChatGPT is designed to be objective by default, and that the study tested a developer API with a prompt that explicitly demanded a candidate recommendation. SpaceXAI did not respond in time to TechCrunch's inquiry.

According to Parikh, a model has an implicit set of values for what it prioritizes and how it leans.