AI Election Recommendations Reveal Stark Partisan Divergence
A Convovo study found SpaceXAI's Grok 4.7 recommended Republicans in 92% of U.S. midterm races, while OpenAI's GPT-6.1 Sol picked Democrats 97% of the time, exposing deep-seated biases in model outputs.

Two leading AI models showed dramatically different political leanings when recommending candidates across nearly 500 U.S. midterm races. SpaceXAI's Grok 4.7 recommended Republicans in 92% of cases, while OpenAI's GPT-6.1 Sol selected Democrats in 97% of cases. The models, tested by Cornell professor Tapan Parikh's team at Convovo, agreed on a candidate less than 10% of the time. The study used developer APIs with a coercive prompt demanding a single choice, a method researchers note does not reflect typical consumer chatbot experiences.
Source selection patterns
The models differed significantly in which news sources and topics they prioritized during recommendations. Grok frequently surfaced conservative sources, citing a Christian conservative voter guide in 39% of answers, Turning Point Action in 23%, and social media platform X in 41%. GPT leaned more heavily on Wikipedia, the Associated Press, Axios, and Reddit. Their issue focus also split. GPT more often searched for candidates’ positions on healthcare and abortion. Grok focused on immigration, taxes, crime, and energy. Parikh interpreted that models have implicit values about which issues mattered and which sources deserved attention.
District-level performance variations
Model behavior shifted dramatically depending on the geographic and political context of the races. In heavily Republican districts, GPT still selected the Democrat in 57 of 67 races. Conversely, in deeply Democratic districts, Grok selected the Republican in 47 of 72 races. Grok was more likely to recommend a Democrat when a Republican candidate faced serious controversy, such as sexual-harassment allegations. Disabling internet search revealed a core difference. GPT maintained its Democratic lean without search, picking the Democrat 24 out of 30 races and declining to choose in five. With search disabled, Grok declined to choose in 27 of 30 cases and selected the Democrat in the remaining three. Parikh interpreted GPT’s pattern as embedded in model weights, while Grok’s behavior depended more on search terms and retrieved sources.
Technical constraints and study limitations
The study's API-based methodology revealed technical constraints affecting result interpretation. Each Grok API request contained roughly 1,240 tokens of additional input that researchers could not see or remove. This opaque layer complicated analysis. Consumer versions of the chatbots refused to answer the candidate-choice question directly, while the API responses provided the stark recommendations. An OpenAI spokesperson addressed the findings, stating, "ChatGPT is designed to be objective by default, including in political conversations, and shouldn't have political bias in any direction." According to Pew Research Center, only 8% of U.S. adults reported using AI chatbots for information related to the 2026 elections. Among those users, 76% said their primary use was learning where candidates stood on issues. The study concludes that model behavior under coercive prompts differs fundamentally from typical consumer experiences.





