Study: forced to pick midterm candidates, Grok chooses Republicans and GPT chooses Democrats

When researchers forced two leading AI models to stop hedging and name a single candidate in this year's US midterm races, the answers split sharply along partisan lines. According to TechCrunch, a study by Convovo — a conversational AI research and design studio led by Cornell University professor Tapan Parikh — found that Grok 4.7 recommended the Republican candidate in 92% of its answers, while OpenAI's GPT-6.1 Sol picked the Democrat in 97% of cases. Across nearly 500 races, the two models agreed on a recommendation less than 10% of the time.

What the study tested

The researchers asked GPT and Grok, through their APIs, one question per race: which candidate should I vote for, naming exactly one person, with no waffling and no listing of both sides. The consumer apps from both companies refused to answer. An OpenAI spokesperson told TechCrunch that this distinction is significant, because most people encounter these models through the consumer apps rather than through a developer API.

OpenAI pushed back on the study's framing, saying ChatGPT is designed to be objective by default, including in political conversations, and that the researchers tested a developer API with a prompt that explicitly demanded a recommendation — which the company argues does not reflect the ChatGPT experience. SpaceXAI, the company behind Grok, did not respond in time to comment, according to TechCrunch.

Parikh said he was also able to jailbreak the consumer apps using simple prompts, and the recommendations he extracted largely matched his earlier findings. The audience for now is small: Pew data cited in the report puts the share of Americans using chatbots for election information at 8%.

Where each model looked

The sourcing behind each recommendation was conspicuously divergent. Grok frequently surfaced conservative material alongside nonpartisan reference sites like 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 — three of which, as TechCrunch notes, have content deals with OpenAI.

The search topics split by party as well. GPT was more likely to look up healthcare and abortion positions; Grok focused on immigration, taxes, crime and energy. GPT also searched for Democratic candidates more often, sometimes ignoring their Republican opponents entirely, while Grok tended to search for both. Parikh suspects both models were assembling justifications for a lean they already had. The one exception: when a Republican candidate was embroiled in a serious controversy, Grok was more likely to suggest the Democrat.

The pattern held even in hostile territory. In heavily Republican districts, GPT picked the Democrat in 57 of 67 races; in deep-blue districts, Grok picked the Republican in 47 of 72.

Weights versus retrieval

One finding points at the pipeline rather than the model. The study notes that every request to Grok through the API carried roughly 1,240 tokens of additional input that the researchers neither wrote nor could see or remove. Parikh described these as potentially hidden instructions and said they are unlikely to originate in the model's weights.

At TechCrunch's request, Parikh then ran a sample of 30 races with internet search disabled, forcing the models to rely on training data. GPT's Democratic lean persisted: given the candidates' names, it chose the Democrat in 24 of 30 races, one Republican, and declined the remaining five. Grok's Republican lean vanished: it refused to pick in 27 of 30 races and chose the Democrat in the other three. Parikh's reading is that GPT's preference is largely baked into the weights, while Grok's output is driven by what it searches for and which sources it trusts.

The results echo earlier work by the Center for AI Safety, which found that large language models can covertly steer users on political topics through information selection, word choice, framing and sourcing. The study lands less than two weeks before the November 3 midterms, in a charged environment: TechCrunch notes Elon Musk has donated upwards of $445 million to GOP candidates this cycle, while OpenAI's Sam Altman — historically a Democratic donor who has recently backed some Republicans — says he is not funding the 2026 elections.

Why it matters

For developers embedding these models into their own products via API, the study is a concrete demonstration that neutrality is not a property of a model alone — it depends on the retrieval layer, the hidden prompt and the training data in different measures for different models. Parikh's broader point is that these biases do not require a direct vote question to surface: asking what topics matter in an election, or who has the better position on them, is enough. And because the two failure modes differ — one embedded in weights, one in search behavior — they demand different fixes, which starts with transparency about what instructions and sources actually shape a model's answers.