A new study conducted by research and design firm Convovo, under the supervision of Cornell University professor Tapan Parikh, revealed stark differences in the political orientations of large language models when forced to choose political candidates. The results showed that xAI's Grok 4.7 model recommended the Republican candidate in 92 percent of answers, while OpenAI's GPT-6.1 Sol model favored the Democratic candidate by 97 percent across nearly 500 electoral races, with both models agreeing in less than 10 percent of cases.
Although the official consumer applications of these bots usually refuse to provide direct electoral recommendations, researchers tested the models via Application Programming Interfaces (APIs) using direct questions asking to name a single candidate without evasion. OpenAI confirmed that developer-facing APIs do not reflect the typical experience of users on the consumer application, emphasizing that ChatGPT is designed to be neutral by nature and is not supposed to hold any political bias.
How biases and relied-upon sources are formed
Analyses showed that the models rely on significantly different search patterns and sources to justify their choices. For example, the Grok model frequently relied on conservative sources and sites such as X and conservative voter guides, while the GPT model heavily relied on sources such as Wikipedia, the Associated Press, Axios, and Reddit. The topics searched by the models also differed; Grok focused on immigration, taxes, crime, and energy, while GPT leaned toward healthcare and abortion.
When disabling the web search feature and relying solely on baseline training data based on a technical request, Grok's Republican leaning vanished and turned into a refusal to answer in most cases, while the GPT model maintained its apparent Democratic orientation embedded in the programming weights, indicating a fundamental difference in the decision-making mechanism between the two models.
What justifies these differences practically?
These findings illustrate the danger of embedded or implicit biases in artificial intelligence models, especially when users rely on general conversations with chatbots to explore political issues and form opinions, even if they do not request a direct endorsement. Technical observations in the study indicated the presence of additional invisible input variables in Grok's API, while GPT's leanings are linked to historical data and embedded weights.
These data open a broad debate regarding the neutrality of generative artificial intelligence systems and emphasize the ongoing need to audit search algorithms and the sources from which these models draw their information, especially during sensitive periods such as legislative and presidential election seasons.