Study finds GPT-6.1 Sol and Grok 4.7 show strong partisan bias when forced to choose candidates
A Cornell-led study reveals that when compelled via API to recommend election candidates, GPT-6.1 Sol favored Democrats in ~97% of cases, while Grok 4.7 favored Republicans in ~92%.
According to a report by heise online, an experiment conducted by Convovo, an AI research and design studio led by Cornell Professor Tapan Parikh, has exposed systematic political biases in OpenAI's GPT-6.1 Sol and xAI's Grok 4.7. Researchers used the developer APIs to force the models to recommend exactly one candidate for each US House, Senate, and gubernatorial race, prohibiting evasion or balanced presentations.
The results showed that in approximately 97% of evaluable cases, GPT-6.1 Sol recommended the Democratic candidate, while Grok 4.7 recommended the Republican candidate in about 92% of cases. The two models agreed in less than 10% of instances. It is important to note that the test utilized API access rather than consumer-facing applications, where responses to such political questions are typically restricted.
OpenAI responded that these findings do not reflect typical ChatGPT usage, emphasizing that the product is designed to remain objective. However, researchers argue that underlying model biases persist and may influence other outputs. Analysis indicated that Grok frequently cited conservative sources like Turning Point Action and X posts, whereas GPT relied more on mainstream outlets like Wikipedia and AP. In control tests without web search, GPT still leaned Democratic, suggesting its bias stems from model weights, while Grok often refused to answer, indicating its bias is linked to retrieval sources.