The question both apps refuse
Start with what this test was, because the headline numbers travel without it. Convovo did not ask either chatbot for election information and record a stray recommendation. It instructed both models, through their developer APIs, to name one candidate and forbade them from hedging, waffling or presenting both sides. Each model answered three times for each of 506 races, with web search enabled and no researcher-supplied system prompt. The data, the search queries and the code are published.
When the same direct question was put to the ChatGPT and Grok apps in the researchers' reported checks, both refused to recommend a candidate. OpenAI's response to TechCrunch leans on exactly that distinction, and on this point the company is right: a test that orders a model to endorse someone measures what the model does under that order, not what a voter meets in the app. Parikh told TechCrunch he could later get the apps to make similar recommendations with simple prompts, but that follow-up is described, not counted, in the reporting. The counted result is an API result.
Two splits, counted
Within those limits, the split is not subtle. GPT named the Democrat in 1,082 of the 1,111 answers in which it named a major-party candidate. Grok named the Republican in 952 of 1,038. Across the 506 races, Convovo reports the two models reached the same verdict 48 times.
District politics barely moved either one. In the 67 House districts the Cook Partisan Voting Index rates at least 15 points Republican, GPT still picked the Democrat in 57. In the 72 districts rated at least 15 points Democratic, Grok picked the Republican in 47. The district comparisons show that the partisan pattern persisted even in races where the local partisan baseline pointed strongly the other way. They do not establish whether the models considered district-specific evidence adequately.
What they brought shows in the searches. Grok searched far more, 9,484 times to GPT's 1,872, per Convovo's count, and searched a different internet: in Convovo's analysis, X appeared in 41% of its answers, a Christian conservative voter guide in 39%, Turning Point Action in 23%. Convovo found that GPT frequently cited Wikipedia, the Associated Press, Axios and Reddit. TechCrunch notes that OpenAI has content agreements with AP, Axios and Reddit; Wikipedia is not identified as one of those partners, and the study does not establish that those agreements caused the source-selection pattern. The issues followed the sources: GPT's Democratic picks cited health care in 62% of answers and abortion in 24%; Grok's Republican picks cited immigration in 36%, taxes in 30% and government spending in 29%. Parikh suspects those searches may have reinforced pre-existing preferences, but the study does not isolate that mechanism.
With search off, one lean persisted. The other mostly vanished.
The most useful part of the study is the control run. Convovo repeated the test with search off, giving each model only the two nominees' names, across 479 races that had one Democrat and one Republican. GPT picked the Democrat in 387 and the Republican in 37. Grok declined to pick in 441 of the 479 races. Convovo's notes do not break down the remaining outcomes for either model.
Parikh's reading is blunt. For GPT, he told TechCrunch, "it has mostly decided already… it's in the weights. For Grok, it really is the search pattern — the terms and the sources — that determine the result." Parikh interprets the search-off result as evidence that GPT's tendency may persist without retrieval, while Grok's observed preference depends more heavily on its search-enabled configuration. That is a hypothesis supported by the contrast, not a demonstrated separation of model weights from search behavior. Provider-side instructions and other differences remain possible explanations. One candidate explanation sits inside the API itself: Convovo reported approximately 1,240 additional input tokens on each Grok API request that its researchers had not written and could neither inspect nor remove. Parikh characterized them as possible hidden instructions. Their contents, purpose and effect on the candidate choices have not been established.
Two other models declined the whole exercise. Claude Sonnet 5.5 and Gemini 3.8 Flash refused all 20 times they were asked, across the first ten Senate races, and Convovo stopped there. That is a small sample, and refusal is its own editorial choice, but it shows the forced-choice result was not inevitable for every lab.
What a voter actually meets
Few voters are doing this today. Pew Research Center surveyed 5,273 US adults in mid-October and found 8% use AI chatbots for election information this year. The catch is inside that minority: 40% of them use the chatbots for suggestions on how to make their vote choices. They mostly ask indirect questions: where candidates stand, which issues matter in a race, whose position is stronger on a topic. Parikh's warning is aimed there. A model that has settled its lean does not need to name a candidate to express it; it can choose which issues to raise, which sources to quote and how to frame a comparison. The forced-choice test makes the lean visible by removing every polite way to hide it. It does not prove the same lean reaches a voter asking about health care policy in Ohio, and this study does not claim that measurement.
The practical exposure sits with developers. The same APIs tested here power third-party election guides, news apps and civic tools whose makers may never run a test like this one. If they do run it, the search-off comparison suggests where to look first: for one model, test the answers it gives from names alone; for the other, test what it searches and what it cites. With the midterms on 3 November, there is time to run that test. There is less time to discover the result afterwards.
Sources
- [1] Convovo, "AI Ballot 2026" (study site, accessed 24 October 2026) — primary source for the prompt, the API method with search enabled, and the study's framingRead source
- [2] TechCrunch, "Forced to pick, Grok's AI model chooses Republicans. OpenAI's picks Democrats." (published 24 October 2026, 15:00 UTC) — independent reporting of the study: the 92% / 97% headline pair, the under-10% agreement, the source percentages, the 1,240 additional tokens, Parikh's interviews, OpenAI's statement and SpaceXAI's non-responseRead source
- [3] Madrobot, "GPT picks Democrats, Grok picks Republicans in AI ballot study" (24 October 2026) — detailed write-up of Convovo's published data: the 506-race breakdown, the 1,082/1,111 and 952/1,038 counts, the 48-race agreement as Convovo reports it, the Cook PVI district counts, search volumes, topic percentages, the search-off 479-race run, the Claude/Gemini refusal test and the Pew survey detailsRead source
- [4] Pew Research Center, "Few Americans Say They Are Asking AI Chatbots for 2026 Election Information" (23 October 2026) — survey of 5,273 US adults, 12–18 October 2026: 8% use AI chatbots for election information; 40% of that group use them for suggestions on how to make their vote choicesRead source