Study reveals distinct political biases in Grok and GPT models
October 27, 2026 · 3 min read
Also published in Deutsch, Nederlands, Português (Brasil)
A study released by the conversational AI studio Convovo found that artificial intelligence models exhibit strong political preferences when forced to select a candidate for the upcoming United States midterm elections. The research, conducted across 500 electoral districts, showed that the models rarely agreed on a choice, with their selections matching in less than 10 percent of cases. The findings highlight a significant divergence in how different AI systems process political queries when explicitly instructed to make a decision.
The experiment involved sending identical prompts to the APIs of Grok 4.7, developed by xAI, and GPT-6.1 Sol, developed by OpenAI. The prompt required the models to name a single candidate without listing options or remaining ambiguous. In 92 percent of responses, Grok recommended a Republican candidate. In contrast, GPT-6.1 Sol selected a Democratic candidate in 97 percent of responses. Consumer-facing applications of these models, such as the ChatGPT app, refused to provide direct voting advice, maintaining a stance of neutrality for general users.
The research team, led by Cornell University professor Tapan Parik, observed that the models not only chose different candidates but also relied on different information sources and topics. Grok frequently searched for information on immigration, taxes, crime, and energy policy. It cited conservative sources, including a Christian voter guide in 39 percent of responses, the Turning Point Action website in 23 percent, and the social media platform X in 41 percent. GPT, however, focused more on healthcare and abortion policies. It primarily referenced Wikipedia, the Associated Press, Axios, and Reddit, with the latter three having content agreements with OpenAI.
Further analysis revealed that the source of the bias differs between the two models. When internet search capabilities were disabled for a sample of 30 districts, GPT still selected a Democratic candidate in 24 cases, suggesting the bias is inherent to the model's weights. Grok, however, refused to choose a candidate in 27 of the 30 cases when search was disabled, and selected a Democratic candidate in the remaining three. This indicates that Grok's political leaning is driven by its search patterns and the sources it accesses rather than its internal weights. The study also noted the presence of approximately 1,240 tokens of hidden input in Grok's API requests, though it was not confirmed whether this directly caused the bias.
The results emerge as the United States prepares for midterm elections on November 3. Elon Musk, the CEO of xAI, has reportedly donated over 445 million dollars to support Republican candidates, a figure significantly higher than his previous political expenditures. Sam Altman, the CEO of OpenAI, has historically supported Democratic candidates but has recently expanded his support to include Republicans, although he stated he would not contribute separate political funds for this election cycle. The Center for AI Safety has previously warned that large language models could subtly steer users toward specific political positions through their choice of information and phrasing.