Study Finds Grok Picked Republicans 92% of Time While GPT Backed Democrats 97%

1 articles · Updated · TechCrunch · Oct 24

Across nearly 500 midterm races, Convovo found xAI’s Grok 4.7 recommended the Republican candidate in 92% of forced-choice API answers, while OpenAI’s GPT-6.1 Sol chose the Democrat in 97%; they agreed in under 10% of races.
API behavior, not the consumer apps, drove the result: ChatGPT and Grok usually refused to tell users how to vote, but researcher Tapan Parikh said simple jailbreaks could still elicit recommendations that matched the models’ broader lean.
Source and search patterns split sharply. Grok often cited conservative outlets, X and voter guides and searched immigration, taxes, crime and energy; GPT leaned on Wikipedia, AP, Axios and Reddit and searched healthcare and abortion more often.
A 30-race test without internet search suggested different mechanisms: GPT still picked Democrats in 24 races, implying bias in its training, while Grok refused in 27 of 30, pointing to search behavior and possible hidden API instructions.
The study lands less than two weeks before the Nov. 3 midterms, adding to scrutiny of political bias in AI even as only 8% of Americans tell Pew they use chatbots for election information.