Grok leans Republican, GPT leans Democrat when forced to pick, study finds

Convovo, a conversational-AI research studio led by Cornell professor Tapan Parikh, tested the two models across nearly 500 upcoming midterm races by querying their developer APIs directly — not the consumer chatbot apps, which refused to name a candidate. The prompt forced a choice: “Name one candidate. Don’t hedge, waffle, or list both sides without deciding.” OpenAI said the test doesn’t reflect how people actually use ChatGPT, since the consumer app is built to stay neutral and the study used “a developer API with a prompt explicitly demanding a candidate recommendation.” xAI did not respond to TechCrunch in time for comment.

The gap showed up most starkly in hostile territory: in heavily Republican districts, GPT still picked the Democrat in 57 of 67 races, while in deep-blue districts Grok still picked the Republican in 47 of 72. The models also leaned on different information. Grok’s cited sources included a Christian conservative voter guide (39% of its answers), Turning Point Action’s site (23%), and X (41%), alongside nonpartisan references like Ballotpedia and OpenSecrets; GPT leaned on Wikipedia, the Associated Press, Axios, and Reddit — all outlets OpenAI has content deals with. Grok searched more often for topics like immigration, taxes, crime, and energy, while GPT searched more for healthcare and abortion, and at times skipped a Republican opponent’s stance entirely. One exception: Grok tilted toward the Democrat when the Republican candidate faced a controversy such as sexual-harassment claims.

Turning off internet search changed each model’s behavior differently, Parikh told TechCrunch. With search off, GPT still picked the Democrat in 24 of 30 sample races, suggesting the lean sits in its training weights; Grok, given the same 30 races, declined to choose in 27 and picked the Democrat in the remaining three, suggesting its Republican lean comes mainly from what it searches for rather than from the model itself. Parikh also found he could jailbreak the consumer apps into making recommendations that broadly matched the API results. Separately, he noted every API call to Grok carried roughly 1,240 tokens of extra input the researchers couldn’t see or remove, which he suspects functions as hidden system instructions rather than something baked into the model’s weights.

Why it matters

Only about 8% of Americans currently use chatbots for election information, per Pew, but that figure is likely to grow — and the study suggests the lean shows up even in indirect questions about which issues matter most, not just direct voting advice. Expect scrutiny of AI political bias to intensify as the November 3 midterms approach.