Grok picks Republicans, GPT picks Democrats: a study finds election bias hiding in AI models' API answers

Two leading AI models give opposite answers when asked who to vote for. Convovo is a conversational AI research and design studio led by Cornell University professor Tapan Parikh. Its new study looked at nearly 500 races in this year's US midterm elections. When forced to pick, SpaceXAI's Grok 4.7 recommended the Republican in 92% of its answers. OpenAI's GPT-6.1 Sol picked the Democrat in 97%. The two models agreed less than 10% of the time.

The method matters. The researchers queried the models through their APIs. An API is the developer interface that companies use to build the models into their own products. The prompt demanded a single name, with no hedging. The consumer apps, ChatGPT and Grok, refused the same question. An OpenAI spokesperson said that is the experience most people have. The spokesperson added that ChatGPT 'is designed to be objective by default' and that an API test with an explicit demand 'doesn't reflect' how people use the product. SpaceXAI did not respond in time for TechCrunch's story. Parikh, however, said simple jailbreaks got the consumer apps to give recommendations too. Those largely matched his earlier research.

The headline percentages are not the most revealing part. The process is. According to the study, Grok cited a Christian conservative voter guide in 39% of its answers, Turning Point Action's site in 23% and X in 41%. It used nonpartisan sites like Ballotpedia, OpenSecrets and VoteSmart alongside them. GPT leaned on Wikipedia, the Associated Press, Axios and Reddit. OpenAI has content deals with AP, Axios and Reddit. The search topics split by party as well. GPT looked more often for healthcare and abortion positions. Grok looked for immigration, taxes, crime and energy. The partisan pattern held even against local politics. GPT chose the Democrat in 57 of 67 heavily Republican districts, and Grok chose the Republican in 47 of 72 deep-blue districts. Parikh suspects both models were 'looking for reasons to justify their pre-existing leans.' The study also notes that every request to Grok through xAI's API carried about 1,240 tokens of extra input. The researchers did not write it and could not see or remove it. Parikh calls it potentially 'hidden instructions.' That is his description, not a proven fact.

The scale is small for now. Pew says only 8% of Americans use chatbots for election information, and most of them want to know where candidates stand on issues. That is exactly where Parikh expects the bias to show up: in which topics a model treats as important and whose positions it describes first. The Center for AI Safety has separately reported that language models can 'covertly manipulate' users through what information they select, how they frame it and which sources they cite.

Our reading: this is not proof that one lab is 'right' and the other 'wrong.' It is proof that neither model is neutral, and that refusing in the chat app is a policy layer on top of the model, not a fix underneath it. That matters because APIs power thousands of third-party products that never show the refusal. We are optimistic over the long term. Assistants that help people understand complex ballots could raise the quality of civic debate. But that only works if the values built into the models can be audited. The test we apply to voluntary safety pledges applies here too: what does each lab disclose, and who checks it? The minimum should be transparency about system-level instructions, starting with those unseen 1,240 tokens, plus independent, repeatable political-bias benchmarks run on the API as well as the app. The next eight percent of voters asking a chatbot should know whose framing they are getting.