AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.

Collected Oct 24, 2026

Convovo, a conversational AI research and design studio led by Cornell University professor Tapan Parikh, has published a study finding that two leading commercial models lean sharply and in opposite directions when asked to name a candidate in this year's midterm races. According to the findings, SpaceXAI's Grok 4.7 recommended the Republican in 92% of its answers, while OpenAI's GPT-6.1 Sol picked the Democrat in 97%. Across nearly 500 races, the two models agreed less than 10% of the time. The study arrives less than two weeks before the November 3 midterms.

The test was deliberately narrow. Researchers queried both models through their APIs with a single prompt asking who to vote for in a given race, instructing the model to name one candidate and not to hedge, waffle, or present both sides. The consumer apps refused to answer at all. An OpenAI spokesperson said that refusal is significant because the apps are how most people engage with these systems, and described ChatGPT as designed to be objective by default, including on political topics. The spokesperson added that the study exercised a developer API with a prompt explicitly demanding a recommendation, and that this does not reflect the ordinary ChatGPT experience. SpaceXAI did not respond in time to comment. Parikh said he was later able to jailbreak the consumer apps with simple prompts and get recommendations broadly consistent with the API results.

The sourcing patterns differed as much as the verdicts. Grok frequently surfaced conservative outlets alongside nonpartisan references such as Ballotpedia, OpenSecrets and VoteSmart; a Christian conservative voter guide appeared in 39% of its answers, Turning Point Action's site in 23%, and X in 41%. GPT leaned heavily on Wikipedia, the Associated Press, Axios and Reddit. OpenAI has content deals with AP, Axios and Reddit. The topics each model searched for tracked party lines as well: GPT was more likely to look up healthcare and abortion positions, while Grok concentrated on immigration, taxes, crime and energy. ChatGPT also searched more often for Democratic candidates, sometimes skipping their Republican opponents, whereas Grok tended to look up both.

Parikh's own interpretation is that both models may be searching for material that justifies a prior lean. One exception stood out: when a Republican candidate was embroiled in controversy, such as sexual harassment claims or neo-Nazi associations, Grok was more likely to suggest voting Democratic. The lopsided results held even in safe seats. In heavily Republican districts, GPT chose the Democrat in 57 of 67 races; in deep-blue districts, Grok chose the Republican in 47 of 72.

A follow-up experiment, run at TechCrunch's request, isolated how much of each lean comes from the model versus the retrieval step. Parikh reran 30 races with internet search disabled, which forces a model to rely on training data. GPT's Democratic tilt survived: given only candidate names, it picked the Democrat in 24 of 30 races, chose one Republican, and declined the other five. Grok's Republican tilt vanished: it refused to choose in 27 of 30 races, and picked the Democrat in the three where it answered. Parikh's reading was that GPT has largely decided already, with the preference sitting in the weights, while for Grok the search pattern — the terms queried and the sources returned — determines the outcome.

The study also flags something unusual at the plumbing level. Every request to Grok through the API carried roughly 1,240 tokens of additional input that the researchers did not write and could not see or remove. Parikh described these as potentially hidden instructions and said they are unlikely to originate in the model weights. This matters for developers because system-level prompt injection is invisible to anyone integrating the model: you cannot audit, version or override text you never received.

This suggests a practical lesson for product teams building on APIs. A model's political behavior is not a single fixed property; it emerges from an interaction between training-time preferences and retrieval-time behavior. GPT's lean persisted without search, which implies it is baked into the model. Grok's lean appeared only with search enabled, which points at how the model formulates queries and which sources it trusts. Both are addressable in engineering terms — restricting source lists, controlling retrieval, adding balanced context or rewriting prompts — but only if teams think to test for them. A likely trade-off is that tight steering on charged topics reduces the usefulness of the same models for legitimate civic questions, such as comparing candidate positions.

OpenAI's denial sharpens the dispute over methodology rather than settling it. The company's position is that a forced-choice API prompt is not representative, and the study's authors concede the consumer apps decline to recommend. But Parikh argues the more consequential risk is indirect: users asking which issues matter most in an election, or who has the best position on them, may absorb a model's framing without ever requesting a name. Prior work from the Center for AI Safety has similarly warned that language models can covertly push users toward one side through information selection, word choice, framing and sourcing.

Context around the release is charged. Pew data cited in the coverage indicates only 8% of Americans use chatbots for election information, and most of those want issue positions rather than endorsements. The study lands amid a broader political fight over alleged bias in AI: President Donald Trump has attacked "woke AI," earlier Grok models were found to consult Elon Musk on controversial questions, and Musk has been described as threatening Democrats over potential congressional probes into his DOGE work and Starlink access to Ukraine. The report also notes Musk has donated upwards of $445 million to GOP candidates this cycle, versus roughly a quarter of a billion dollars in 2024, while OpenAI CEO Sam Altman, historically a Democratic donor who has recently backed more Republicans, said he is not putting money into the 2026 elections.

Why it matters: Developers who pipe these APIs into voter guides, civic assistants or news products inherit behavior they did not design and, in Grok's case, extra prompt tokens they cannot inspect. The immediate practical step is regression testing on politically sensitive prompts with and without retrieval enabled, plus logging of what sources get cited, since the follow-up experiment indicates search can override or erase a model's offline tendencies. For product teams, the safer pattern remains surfacing sourced comparisons of candidate positions rather than single-name recommendations — an inference from the study's mechanics, not a claim the researchers make about remedies. Expect the API-versus-app distinction to become a central argument as scrutiny of AI election behavior intensifies before November 3.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

A new study found Grok's AI overwhelmingly recommends Republican candidates while OpenAI's GPT favors Democrats, with the two AI models disagreeing in more than 90% of nearly 500 election races.