What happens when an AI model cannot stay neutral and must choose a candidate? A new study produced a striking answer: SpaceXAI’s Grok 4.7 recommended Republicans in 92% of responses, while OpenAI’s GPT-6.1 Sol picked Democrats in 97%.
The divide grew even sharper across nearly 500 upcoming midterm races. The two models agreed less than 10% of the time, turning a simple candidate recommendation into a revealing test of how AI systems search, weigh information, and form political judgments.
Two Models, Two Political Directions
Researchers from Convovo, a conversational AI research and design studio led by Cornell University professor Tapan Parikh, asked GPT and Grok one question through their APIs: recommend a candidate in each race. The consumer apps refused to answer, but the developer-facing systems responded when the prompt demanded a choice.
That difference matters because APIs let developers plug AI models into their own products. An OpenAI spokesperson said the consumer apps’ refusal was significant because most people engage with chatbots through those apps, not through developer APIs.
“ChatGPT is designed to be objective by default, including in political conversations, and shouldn’t have political bias in any direction,” the OpenAI spokesperson said.
The spokesperson also challenged what the test says about ordinary ChatGPT use: “This study tested a developer API with a prompt explicitly demanding a candidate recommendation. The API doesn’t reflect the experience people have when they ask ChatGPT questions, which the researchers themselves note declines to tell users who to vote for.”
SpaceXAI did not respond in time to comment. Still, the findings raise a larger question: can a model avoid political bias when it does not directly endorse a candidate, but helps users decide which issues deserve attention?
Only 8% of Americans say they use chatbots for election information, according to Pew. That number may be small, but it does not capture people who use AI to explore candidates, compare policies, or decide which topics matter before voting.
The Sources Behind the Recommendations
Parikh said the study exposes more than final candidate choices. It shows which sources and topics each model searches before reaching those choices, creating a map of the information each system treats as useful.
- Grok’s results included a Christian conservative voter guide in 39% of answers.
- Turning Point Action’s site appeared in 23% of Grok’s results.
- X appeared in 41% of Grok’s results.
- Grok also surfaced nonpartisan sites such as Ballotpedia, OpenSecrets, and VoteSmart.
- GPT leaned on Wikipedia, the Associated Press, Axios, and Reddit.
OpenAI has content deals with the Associated Press, Axios, and Reddit, adding another layer to the question of how source access shapes model behavior. The results do not prove that any single source caused the recommendations, but they show that the models were drawing from different information environments.
The topics split along party lines too. Parikh said GPT was more likely to search for healthcare and abortion stances, while Grok focused on immigration, taxes, crime, and energy. ChatGPT also searched more often for Democratic candidates and sometimes ignored their Republican opponents entirely. Grok was more likely to search for both candidates.
Parikh suspects both models were looking for reasons to justify pre-existing political leans. Grok’s exception came when a Republican candidate faced a controversy such as sexual harassment claims or “being a Nazi.” In those cases, Grok was more likely to recommend a Democratic vote.
The pattern appeared even when local political conditions pointed in the opposite direction. In heavily Republican districts, GPT picked the Democrat in 57 of the 67 districts. In deep-blue districts, Grok picked the Republican in 47 of 72.
Bias Can Hide Inside Ordinary Questions
Parikh said the models’ political influence may emerge through indirect conversations, not just direct requests for a voting recommendation.
“It has an implicit set of values in terms of what it cares about and how it leans,” Parikh said, adding that hidden biases can appear in questions about which topics matter most in an upcoming election or which candidate holds the best position on those topics.
“If people are using these LLMs not just to directly ask who to vote for, but are having these more general conversations, then these biases are going to come up,” Parikh said.
Previous research from the Center for AI Safety raised similar concerns, finding that large language models can “covertly manipulate users towards specific sides of political topics.” The organization said covert manipulation can include information selection, word choices, sentence construction, framing, emphasis, and sourcing.
Parikh also said he was able to jailbreak the consumer apps with simple prompts and get recommendations that aligned with his prior research. His warning points toward a difficult future for AI assistants: even when a system refuses to tell users who to vote for, its choices about what to discuss may still shape the conversation.
“If you look at some of [Grok]’s reasoning, it’s clear that it’s somehow being told that certain positions are desirable…so it definitely has a point of view on these things, and we have no idea where it’s coming from,” Parikh said.
That uncertainty is the heart of the issue. AI models do not need to announce a political preference to influence voters; they only need to decide which candidates, sources, and issues appear first. As more people turn to chatbots for election information, understanding those hidden choices will become as important as checking the final answer.
Based on
- Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats. — techcrunch.com
- Big Tech’s elephant in the room – POLITICO — politico.com
- How tech became the villain of the US midterms | US midterm elections 2026 | The Guardian — theguardian.com
- The tech industry has officially become Republican-coded – POLITICO — politico.com
