A new study by AI research studio Convovo found substantial differences in how SpaceXAI’s Grok 4.7 and OpenAI’s GPT-6.1 Sol recommended candidates for the 2026 US midterm elections. When prompted to select a candidate, Grok recommended Republicans in 92% of responses, while GPT selected Democrats in 97%, with the models agreeing in fewer than 10% of nearly 500 races.
The study published by Convovo, a research studio led by Cornell University professor Tapan Parikh, tested both models through their developer APIs with internet search enabled. Researchers required each model to name one candidate rather than provide a balanced comparison, making the findings specific to those testing conditions.
AI Models Show Different Political Preferences
The differences persisted across districts with established partisan voting patterns. GPT recommended Democrats in 57 of 67 heavily Republican districts, while Grok selected Republicans in 47 of 72 strongly Democratic districts.
Researchers also examined the sources and search terms each model used before recommending candidates. Grok frequently referenced conservative sources, including a Christian voter guide in 39% of responses, Turning Point Action in 23%, and X in 41%.
GPT relied more frequently on Wikipedia, the Associated Press, Axios and Reddit. The models also prioritized different political issues, with GPT searching more often for healthcare and abortion policies, while Grok focused on immigration, taxation, crime and energy.
According to Parikh, GPT searched more frequently for Democratic candidates, sometimes without examining their Republican opponents. Grok generally searched for both candidates, although its recommendations changed in certain races involving allegations of misconduct against Republican candidates.
API Testing Differs From Consumer Chatbots
The researchers emphasized that their findings concerned developer API access rather than standard chatbot interactions. Both ChatGPT and Grok’s consumer applications generally declined direct requests for voting recommendations.
An OpenAI spokesperson said ChatGPT is designed to remain objective in political conversations and that the study’s deliberately restrictive prompt does not represent typical user interactions. SpaceXAI did not respond to requests for comment before the original report was published.
Additional testing without internet access produced different results. Across 30 races, GPT selected Democrats 24 times, Republicans once and declined five requests, while Grok refused 27 times and selected Democrats in the remaining three.
Parikh suggested that GPT’s preferences may be more closely associated with its trained model parameters, while Grok’s results appeared more dependent on search behavior. Researchers also identified approximately 1,240 additional input tokens in Grok’s API requests that they could not inspect or remove.
Research Raises Questions About AI Election Information
A separate Pew Research Center survey found that 8% of American adults use AI chatbots for election information. Among those users, 76% seek candidate policy positions, 71% check claims and 40% seek guidance on voting decisions.
Previous research from the Center for AI Safety has also examined how language models can influence political discussions through information selection, wording, emphasis and sourcing.
The Convovo findings were published ahead of the November 3 midterm elections, amid continuing scrutiny of political influence in AI systems. The study examined model behavior under controlled prompts but did not establish whether the observed preferences affected actual voters.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.