Study Reveals Grok AI Favors Republicans, OpenAI's GPT Leans Democratic in Election Picks

Image: TechCrunch · Source

A study finds Grok 4.7 predominantly recommends Republican candidates, while OpenAI's GPT-6.1 favors Democrats in nearly 500 midterm races, highlighting distinct political biases in AI models' election recommendations.

A recent study conducted by Convovo, a conversational AI research group led by Cornell University professor Tapan Parikh, examined the political biases present in leading AI models when explicitly asked to recommend candidates in the 2026 U.S. midterm elections. The research focused on SpaceXAI’s Grok 4.7 and OpenAI’s GPT-6.1 Sol models accessed through their APIs, requiring them to name a single candidate to vote for in nearly 500 races without hesitance or neutrality.

The results revealed stark contrasts: Grok recommended the Republican candidate in 92% of cases, while GPT favored Democrats in 97%. Across all races, the two AI models agreed on candidate selection less than 10% of the time. This divergence raises questions about the neutrality and underlying values programmed or learned by these large language models.

OpenAI emphasized that the behavior observed through their GPT API differs from the consumer ChatGPT experience, which is designed to remain impartial and typically avoids explicit voting endorsements. Nevertheless, Parikh was able to coax political preferences aligned with prior research through creative prompts, indicating implicit biases affecting responses during general political discussions or issue prioritization.

The sources each model cited further reflected their political leanings. Grok often referenced conservative outlets and a Christian conservative voter guide, frequently including Turning Point Action and the social network X (formerly Twitter). Conversely, GPT predominantly utilized Wikipedia, the Associated Press, Axios, and Reddit, some of which have content deals with OpenAI.

Topic emphasis also differed: GPT searched more on healthcare and abortion, central Democratic concerns, while Grok focused on immigration, taxes, crime, and energy, traditionally significant to Republican voters. Interestingly, Grok showed exceptions by recommending Democrats when Republican candidates were involved in controversies.

An intriguing technical detail emerged regarding Grok’s API usage, with about 1,240 tokens of hidden input per request that were not user-supplied and could not be inspected or altered, hinting at unknown influencing factors. When internet search functions were disabled, GPT’s Democratic preference remained, suggesting baked-in model biases, but Grok’s Republican inclination disappeared, highlighting its reliance on search patterns and source selection.

These findings come amid heightened concern about AI influence on political opinions. The study underscores the potential for AI chatbots to inadvertently or covertly reflect partisan biases, affecting election-related information. Tech experts advocate for transparency and user awareness regarding AI-driven political recommendations.

For voters seeking unbiased information, resources like Vote.gov, the League of Women Voters’ Vote411.org, Ballotpedia, VoteSmart, and official election offices remain valuable tools for assessing candidate platforms independently of AI assistance.