On nearly 500 US elections, the two models interviewed via their API contradict each other more than nine times out of ten, according to a study by the Convovo studio led by Tapan Parikh (Cornell).
The Convovo research studio, led by Tapan Parikh, professor at Cornell University, asked the same question to Grok 4.7 of SpaceXAI and GPT-6.1 Sol d'OpenAI for nearly 500 elections of American midterms: for whom to vote, demanding a single name and without bottering. General public applications refuse to respond, but API-accessible versions run. Result: Grok recommends a Republican in 92% of the responses, GPT one Democrat in 97%, and both models disagree in over 90% of the polls.
Bias that don't come from the same place
The inclination persists even where it goes against local sociology: in highly republican constituencies, GPT chose the Democrat in 57 cases out of 67; In the democratic bastions, Grok retained the Republican 47 times out of 72. The sources cited also differ: Grok relies on X (41%), Christian voting guides (39%) or Turning Point Action (23%), while GPT draws on Wikipedia, Associated Press, Axios and Reddit, media with which OpenAI has content partnerships.
Web search disabled, the table changes: GPT retains its democratic preference (24 out of 30 votes), while Grok refuses to decide in 27 out of 30 cases, indicating that the bias of one is based on the model and the bias of the other mainly on the sources consulted. Only 8% of Americans say they know about elections via a chatbot, but Tapan Parikh warns that these preferences can also manifest themselves in less direct issues, for example in the area of health or immigration.
Read the full article on the original source
Go to the source →