Convovo, a conversational AI research studio, queried the APIs of SpaceXAI's Grok 4.7 and OpenAI's GPT-6.1 Sol with a forced-choice question about who to vote for in each of roughly five hundred midterm races. Grok recommended the Republican candidate in 92 percent of its responses, while GPT-6.1 chose the Democrat in 97 percent. The two systems agreed on less than ten percent of the contests, highlighting a stark partisan split in their outputs.
Only eight percent of U.S. adults report using chatbots for election information, according to Pew data, and most of those users seek issue-specific guidance. When the researchers asked the consumer versions of the models the same forced-choice question, both apps declined to answer. An OpenAI spokesperson noted that this refusal is notable because it mirrors how most people interact with the chat interfaces.
Parikh emphasized a key distinction between the consumer chat experiences and the raw API calls that developers embed in their own products. He demonstrated that simple jailbreak prompts could coax the consumer apps into offering partisan recommendations that matched his earlier findings. "It has an implicit set of values," Parikh said, indicating that hidden biases can surface even in broader conversational queries about elections.
The study also tracked the sources each model cited. Grok frequently surfaced conservative outlets, including a Christian voter guide in 39 percent of answers, Turning Point Action in 23 percent, and posts from X in 41 percent, alongside nonpartisan sites such as Ballotpedia. GPT-6.1, by contrast, more often searched for positions on healthcare and abortion, while Grok prioritized immigration, taxes, crime and energy topics.
District-level analysis revealed further partisan patterns. In heavily Republican districts, GPT-6.1 selected the Democratic candidate in 57 of 67 races. Conversely, in strongly Democratic districts, Grok recommended the Republican in 47 of 72 contests. An exception emerged when a Republican candidate faced a controversy such as sexual-harassment allegations; in those cases Grok was more likely to suggest voting Democrat.
Each request to Grok's API contained roughly 1,240 hidden tokens that the researchers could not see or remove, which Parikh described as potential "hidden instructions." When internet search was disabled, GPT-6.1 still chose Democrats in 24 of 30 races, indicating the bias resides in its weights. Grok, however, refused to pick a candidate in 27 of 30 cases and chose a Democrat only three times, suggesting its partisan tilt depends on search behavior. The findings arrive just weeks before the November 3 midterms, a timing that fuels ongoing debates about political bias in AI, echoed by President Trump’s criticism of "woke AI" and by the Center for AI Safety’s warning that large language models can "covertly manipulate users towards specific sides of political topics."