AI Models Grok and GPT Display Clear Political Partisanship

When forced to select a winner in upcoming midterm elections, SpaceXAI’s Grok 4.7 favored Republican candidates in 92% of cases, while OpenAI’s GPT-6.1 Sol opted for Democrats 97% of the time. The study, conducted by Convovo, analyzed nearly 500 races and found that the two prominent artificial intelligence models reached the same conclusion in less than 10% of their responses.

Cornell University professor Tapan Parikh, who led the research, argued that these findings expose underlying biases in how large language models prioritize information. The Scale Report notes that while consumer versions of these apps typically refuse to provide voting endorsements, accessing them through developer APIs revealed starkly different behavior. According to the reporting by TechCrunch, OpenAI representatives defended their model by stating that their consumer-facing ChatGPT is designed to remain neutral and that the API results do not represent the standard user experience.

Source selection and search patterns

Part of the discrepancy appears to stem from the disparate sources each model prioritizes during internet searches. Grok frequently pulled data from Christian conservative voter guides, Turning Point Action, and X, while GPT relied heavily on Wikipedia, the Associated Press, and Axios. Tapan Parikh observed that Grok also appeared to include hidden instructions in its API requests, totaling 1,240 tokens that researchers could not identify.

Why this matters for AI objectivity

These findings suggest that even when users are not explicitly asking for voting advice, AI systems may subtly nudge them through word choice, information framing, and source emphasis. As these tools become primary interfaces for news consumption, the inability of models to remain neutral challenges the assumption that AI serves as an objective arbiter of information, raising concerns about potential covert manipulation of public discourse.