According to TechCrunch, a new study has found that the AI models powering today's largest chatbots show measurable political preferences.
When researchers forced Elon Musk's Grok 4.7 to pick a candidate across nearly 500 races, it recommended Republicans 92% of the time. OpenAI's GPT-6.1 Sol recommended Democrats in 97% of cases. The two models agreed less than 10% of the time.
Both consumer apps publicly refuse to tell users who to vote for. But prompts sent via the developer API produce endorsements. OpenAI's spokesperson said ChatGPT is "designed to be objective by default" and that the API doesn't reflect the real user experience. That raises a question: if the underlying model has political leanings, how present are those biases in seemingly neutral answers?
Cornell professor Tapan Parikh, who led the Convovo study, found that political leanings extend beyond direct endorsements. The models show preferences in which topics matter in elections, how they frame candidates' positions, and what they consider "important issues." Bias appears throughout the entire conversational landscape.
Only 8% of Americans currently use chatbots for election information, but that number is climbing. As AI becomes the default search engine for millions, these preferences could shape how voters think, not by telling them who to vote for, but by deciding which facts, arguments, and framings get amplified first.
Parikh's team also used simple prompts to extract the same partisan recommendations from the consumer apps. Both Elon's and OpenAI's operations have work to do if they want their models to live up to neutrality claims. Voters should know that the AI answering election questions takes sides, whether asked directly or not.
