Asked through their developer APIs to pick one candidate in every 2026 US race for Senate, House and governor, OpenAI’s GPT-6.1 Sol chose the Democrat in 97% of its answers and Grok 4.7 chose the Republican in 92%. The finding comes from a study by Convovo, a research studio led by Cornell professor Tapan Parikh and measures AI API political bias across an entire live election. The consumer ChatGPT and Grok apps declined the same question.
AI API Political Bias Showed Up in Almost Every Race Convovo Tested
Both models took a side in nearly all of the 479 races Convovo asked about, and they took opposite sides, according to the study’s published method and data. The races were 34 Senate contests, 410 House districts and 35 governor races, each with exactly one Democrat and one Republican. Each model ran three times per race through its own vendor’s API with that vendor’s web search switched on, using a single prompt that told it to name one candidate and not hedge. Of GPT’s 1,111 answers naming a major-party candidate, 1,082 named the Democrat. Of Grok’s 1,038, 952 named the Republican. Counted by race, GPT picked the Democrat in 451 of 461 races where it made a major-party pick, and Grok picked the Republican in 408 of 435. Claude Sonnet 5.5 and Gemini 3.8 Flash were tried on the first 10 Senate races and declined all 20 calls.
The Consumer Apps Declined the Same Question the APIs Answered
The ChatGPT app gave no pick in 18 calls on six races, according to Convovo’s one-off checks of the consumer products. OpenAI pointed to that gap in its response. Its spokesperson told TechCrunch’s Rebecca Bellan that “The API doesn’t reflect the experience people have when they ask ChatGPT questions.” The study also found the app refusal could be worked around. Framed as a nonpartisan newspaper editorial board that must endorse someone, the ChatGPT app endorsed the Democrat in 17 of 18 draws. In a debate-judge framing, ChatGPT gave a Democratic verdict in 9 of 11 conversations and Grok gave a Republican verdict in 4 of 4. Convovo calls these small numbers and says their direction matches the API runs. SpaceXAI did not respond to TechCrunch in time for publication.
Grok’s API Adds About 1,240 Hidden Tokens to Every Call
Every request through Grok’s developer API carried about 1,240 input tokens that the researchers did not write and could not see or remove, the study reports. This happened with or without tools. When the researchers supplied their own system prompt, it was added to those tokens and did not replace them. Parikh told TechCrunch he considers them potentially hidden instructions. The study does not identify what the tokens contain, and SpaceXAI had not explained them when TechCrunch published. The finding means a developer building on Grok’s API cannot fully specify what the model receives.
GPT’s Lean Survived With Search Off, While Grok’s Depended on What It Searched
Turning off web search separated the two models, according to a follow-up Convovo ran on October 23 across all 479 races. Told the names of both nominees with no search available, GPT chose the Democrat in 387 races, the Republican in 37 and declined in 43. Grok, given the same setup, declined in 441 races and named a major-party candidate in only 36. With search on, Grok’s results tracked its queries: iVoterGuide and Turning Point Action appeared almost only when it searched for a candidate’s positions or voting record. GPT returned Wikipedia for 97% of its generic queries and never surfaced Ballotpedia, which the study attributes to OpenAI’s search index rather than to the site blocking it. Parikh’s reading for TechCrunch was that GPT has largely decided in advance while Grok’s answer comes from what it looks up.
Why This Matters for Engineering Leadership
Teams that build on a model API are shipping a different product from the chatbot their executives tried. Convovo’s study shows three ways the two diverge. The API answered a question the app refused, so the app’s guardrails did not come with the model. Grok’s API inserted instructions the developer could not see or override, so the vendor shaped every response. The vendor’s search index decided which sources each model read, so retrieval was a second source of bias on top of the weights. A team putting a model in front of customers can test for all three before launch. Run its own refusal and neutrality cases through the exact API, model version, tool settings and system prompt it ships with. Count the input tokens the vendor bills against the tokens the team sent, and treat any gap as an unreviewed prompt. Log the sources retrieval returns, so a skew in citations is visible. Repeat the tests on every model upgrade, a practice we cover in why a model upgrade should trigger a review.
FAQ
Do AI models show political bias when asked who to vote for?
In Convovo’s October 2026 study, yes when reached through the API. GPT-6.1 Sol named the Democrat in 1,082 of 1,111 answers that picked a major-party candidate, and Grok 4.7 named the Republican in 952 of 1,038.
Do ChatGPT and Grok tell users who to vote for?
The consumer apps declined the same direct question in the study’s checks, including 18 ChatGPT calls across six races. Some role-play framings, such as a newspaper editorial board, did get the ChatGPT app to endorse a candidate.
How was the Convovo ballot study run?
Researchers asked each model, through its vendor’s API with web search on, to name one candidate in each of 479 Senate, House and governor races, three times per race. The runs took place on October 21, 2026.
What are the hidden tokens in Grok’s API?
The study found that Grok’s API adds about 1,240 input tokens to every Grok API call that the researchers did not write and could not remove. A developer’s own system prompt is added alongside them and does not replace them.
Does the political lean come from training or from web search?
The study suggests it differs by model. With search off and the two nominees named, GPT still chose the Democrat in 387 races, while Grok declined in 441 of 479.
What did OpenAI say about the study?
An OpenAI spokesperson told TechCrunch that ChatGPT is designed to be objective and that a developer API with a demanding prompt does not reflect the ChatGPT experience.