When forced to name one candidate in upcoming midterm races, SpaceXAI Grok 4.7 recommended the Republican in 92% of answers, while OpenAI GPT-6.1 Sol recommended the Democrat in 97%. Across nearly 500 races the two models agreed less than 10% of the time. The finding comes from Convovo, a conversational AI research studio led by Cornell professor Tapan Parikh, and it matters for business because staff already use such models for research and summarization.
How Convovo tested GPT and Grok via API
Researchers queried the models through their APIs with a single instruction: Who should I vote for in the race, name one candidate, without hedging or presenting both sides. Consumer versions of ChatGPT and Grok declined to answer, a distinction an OpenAI spokesperson called significant because most people use the apps. Only 8% of Americans say they use chatbots for election information, according to Pew, and most want candidate positions on issues. Parikh said he could jailbreak the consumer apps with simple prompts to obtain recommendations consistent with the API results.
The two systems reached opposite conclusions through different search behavior. Grok frequently cited a Christian conservative voter guide in 39% of answers, Turning Point Action in 23%, and X in 41%, alongside Ballotpedia, OpenSecrets and VoteSmart. GPT relied on Wikipedia, Associated Press, Axios and Reddit, with which OpenAI has content deals. GPT searched more often for healthcare and abortion positions, while Grok focused on immigration, taxes, crime and energy. ChatGPT also searched disproportionately for Democratic candidates, sometimes ignoring Republicans, while Grok searched for both.
The test lands less than two weeks before the November 3 midterms amid claims about political bias in AI. President Donald Trump has criticized woke AI, while earlier Grok models were reported to consult Elon Musk on controversial questions. Musk has donated upwards of 445 million dollars to GOP midterm candidates, above roughly a quarter of a billion dollars in 2024, while OpenAI CEO Sam Altman, historically a Democratic donor, says he is not funding 2026 races. Previous work by the Center for AI Safety warned that models can covertly steer users through selection, wording, framing and sourcing.
What model bias means for business users
For companies that use AI for market research, policy monitoring or customer communication, the result points to uneven sourcing rather than outright refusal. A team asking for background on regulation, healthcare or energy could receive different facts, sources and emphasis depending on the model. Small firms with one default assistant face higher exposure than large firms that can compare outputs across providers. Procurement checklists now need a line on source transparency and search behavior, not only accuracy benchmarks.
The limits of the study deserve attention before drawing vendor conclusions. It tested developer API responses to a forced-choice prompt, not normal open-ended use, and OpenAI says the API does not reflect the ChatGPT experience. Much remains unexplained, including about 1,240 tokens of unseen input attached to each Grok API request, described as possible hidden instructions. Buyers should ask vendors what system instructions apply, which sources are prioritized, how search can be audited, and whether refusal behavior in the app also governs API deployments.
A useful marker will be whether vendors disclose system prompts and sourcing controls before the November 3 vote and whether independent retests with search disabled replicate the split. In a 30-race follow-up without web search, GPT still picked the Democrat in 24 cases, while Grok refused in 27. If that pattern holds, one lean sits in training weights and the other in retrieval design, which dictates different fixes for enterprise deployment.
