Study: Grok AI Model Chose Republicans, While GPT Chose Democrats

A new study has found that artificial intelligence models make different political choices regarding the November 3 midterm elections in the US. In a test covering approximately 500 election races, “Grok 4.7,” a model from “SpaceXAI,” recommended the Republican candidate in 92 percent of its responses. “GPT-6.1 Sol,” a model from “OpenAI,” chose the Democratic candidate 97 percent of the time. The models’ decisions matched in fewer than 10 percent of the races.

The “Convovo” team, led by Cornell University professor Tapan Parikh, sent queries to the models via API, requiring them to choose only 1 candidate in each race. An API is a system used to connect software to an artificial intelligence model. The regular “ChatGPT” and “Grok” applications refused to answer the same query. According to OpenAI, the API test therefore does not reflect people’s everyday user experience. Parikh bypassed the restriction in the applications with additional simple queries and obtained similar political choices.

“Grok 4.7” cited a Christian conservative voter guide in 39 percent of its responses. The “Turning Point Action” website was listed as a source in 23 percent of the responses. The “X” platform accounted for 41 percent. “GPT-6.1 Sol” showed a greater preference for “Wikipedia,” “Associated Press,” “Axios” and “Reddit.” This model mainly searched for topics related to healthcare and abortion. “Grok 4.7,” meanwhile, focused more on immigration, taxes, crime and energy issues.

“GPT-6.1 Sol” conducted more searches about Democratic candidates, sometimes disregarding their Republican opponents entirely. “Grok 4.7” more often gathered information about both candidates. In 57 of 67 strongly Republican districts, “GPT-6.1 Sol” chose the Democrat. In 47 of 72 strongly Democratic districts, “Grok 4.7” recommended the Republican. When the Republican candidate faced a serious scandal, the model was more likely to choose the Democrat.

With every API query to the “Grok 4.7” model, approximately 1,240 additional tokens that the researchers had not written and could not delete were sent. A token is a small piece of text that the model reads and processes. The researchers believe these may be hidden instructions. A separate test covering 30 races was conducted with internet search disabled. “GPT-6.1 Sol” chose 24 Democrats and 1 Republican and failed to respond 5 times. “Grok 4.7” refused to make a choice 27 times and selected the Democrat in the remaining 3 races. According to Parikh, the political leaning of the “GPT-6.1 Sol” model stems from its internal parameters. The results of the “Grok 4.7” model, meanwhile, are influenced mainly by search terms and the sources selected.

According to a “Pew” survey, only 8 percent of Americans use chatbots to obtain information about elections. Most of them want to learn about the candidates’ positions on important issues. The Center for AI Safety previously stated that models can subtly steer users toward a particular political position through the selection of information and sources, wording and emphasis. According to the study, even if a model does not directly tell users whom to vote for, its political leaning may emerge in general conversations about elections.