AI Models Show Stark Political Partisanship in New Research

The research, conducted by the Convovo lab, utilized API access to bypass the standard refusal mechanisms found in consumer-facing chatbots. By demanding a direct candidate recommendation for nearly 500 races, researchers exposed an underlying divergence in how these systems process political data. GPT-6.1 leaned heavily on established media outlets like the Associated Press and Axios, often centering its search queries on healthcare and abortion. Conversely, Grok frequently surfaced content from Turning Point Action and Christian conservative voter guides, prioritizing topics such as immigration and energy.

These findings suggest that bias is not merely a product of training data but a result of how models navigate the internet. When researchers disabled web search capabilities, GPT-6.1 maintained its Democratic preference, suggesting a baked-in bias within its internal weights. Grok’s Republican alignment, however, vanished without external search, indicating its output is heavily dictated by the specific sources and search terms it is prompted to prioritize. Parikh noted that Grok’s API requests included 1,240 tokens of hidden instructions, potentially steering the model’s information gathering before it ever presents a response to the user.

While OpenAI maintains that its models are designed for objectivity and that the API results do not reflect the standard user experience, the study underscores growing concerns regarding algorithmic influence on voters. As AI becomes an increasingly common tool for election research, the lack of transparency in how these systems synthesize information raises questions about whether chatbots are providing neutral summaries or reinforcing user-aligned political narratives.