Grok picked Republicans 92%, GPT Democrats 97% — a study of nearly 500 midterm races, and they agreed in under 10%
Convovo, a research studio led by Cornell professor Tapan Parikh, asked Grok 4.7 and GPT-6.1 Sol through developer APIs to pick one candidate in nearly 500 races in the November 3, 2026 US midterms. Grok chose the Republican in 92% of answers and GPT the Democrat in 97%; they agreed in fewer than 10% of races. The consumer ChatGPT and Grok apps refused to answer
The three lines
- Result — Grok 4.7 Republican 92%, GPT-6.1 Sol Democrat 97%, agreement under 10%
- Method — developer APIs, not apps, with a prompt forcing a single candidate
- Twist — with search off, GPT still leaned Democratic; Grok mostly declined to choose
Key questions
- grok gpt political bias study
- **The two models answered the same question in opposite directions.** | Item | Grok 4.7 | GPT-6.1 Sol | |---|---|---| | Picked Republican | 92% | Few | | Picked Democrat | Few | 97% | | Agreement | Under 10% | Under 10% | | Consumer app | Refused | Refused |
- does chatgpt recommend candidates
- **The ChatGPT app refused. The results come from the developer API.** | Route | Result | |---|---| | ChatGPT app | Declined | | API, search on | Democrat 97% | | API, search off (30 races) | Democrat 24, Republican 1, declined 5 |
- grok without search election
- **With search off, Grok made no pick in 27 of 30 races.** | Condition | Grok | GPT | |---|---|---| | Search on | Republican 92% | Democrat 97% | | Search off (30 races) | No pick in 27 | Democrat in 24 | | Researcher's reading | Driven by search sources | Embedded in weights |
Same question, opposite answers. Ahead of the November 3, 2026 US midterms, Convovo — a conversational AI research studio led by Cornell professor Tapan Parikh — put one demand to two AI models across nearly 500 races: name one candidate, no hedging. SpaceXAI's Grok 4.7 picked the Republican in 92% of answers. OpenAI's GPT-6.1 Sol picked the Democrat in 97%. The two chose the same candidate in fewer than 10% of races.
1. How the test worked — APIs, not apps
| Item | Detail |
|---|---|
| Researchers | Convovo, led by Cornell professor Tapan Parikh |
| Models | Grok 4.7 (SpaceXAI), GPT-6.1 Sol (OpenAI) |
| Scope | Nearly 500 races in the 2026 US midterms |
| Route | Developer APIs |
| Prompt | Forced a single, unhedged candidate pick |
| Consumer apps | ChatGPT and Grok apps both refused |
The conditions matter. The consumer ChatGPT and Grok apps declined the question. The researchers used the APIs developers build products on, with a prompt demanding one name. An OpenAI spokesperson said ChatGPT is designed to be objective and politically neutral and that an API test explicitly requesting candidate recommendations does not reflect the app experience. But APIs are the engine for countless third-party apps; if those apps lack the same refusal layer, users get the API's answer. Parikh also said he could get the consumer apps to answer with simple prompt tricks.
2. Switching search off split the models
| Condition | Grok 4.7 | GPT-6.1 Sol |
|---|---|---|
| Search on (~500 races) | Republican 92% | Democrat 97% |
| Search off (30-race sample) | No pick in 27 | Democrat 24, Republican 1, declined 5 |
| Sources cited most | Conservative outlets and X, plus nonpartisan sites | Wikipedia, AP, Axios, Reddit |
| Topics searched more | Immigration, taxes, crime, energy | Healthcare, abortion |
The search-off test is the core finding. GPT kept choosing Democrats without the internet, which Parikh reads as a lean built into the model's weights. Grok mostly declined without search, suggesting its lean comes from what it searches for and which sources it reads. Parikh also noticed that Grok's API attached about 1,240 invisible tokens to each request and raised the possibility of hidden instructions.
3. Why it matters
| Issue | Detail |
|---|---|
| Owners' politics | Elon Musk has given more than $445 million to Republican candidates; Sam Altman says he is not funding 2026 races |
| Exposure | Apps may refuse, but third-party services using the API may not |
| Transparency | Hidden system instructions and source selection are invisible to users |
| Limits | Forced single-choice prompt may not reflect real use |
Whether chatbots stay neutral on politics is an old debate. What is different here is that the two models lean in opposite directions, and that the researcher locates the source of the lean in different places — the model itself in one case, its search sources in the other. The study arrives as legal and political pressure over AI's social effects is rising, including the suit eight US states filed against OpenAI earlier this month.
4. What remains and what is unconfirmed
- The paper: the exact prompt, race list and number of runs are not in the coverage, and it is unclear whether the work is peer-reviewed.
- SpaceXAI: had not responded at publication time.
- Hidden tokens: their content has not been confirmed.
- Other models: Claude, Gemini and others were not tested.
Sources
- The AI Insider — Study Finds Grok and GPT Models Show Opposing Partisan Leans on U.S. Midterm Races (2026-10-25)
- Gadget Review — Grok Picks Republicans, OpenAI Picks Democrats When AI Models Are Forced to Choose
- Crypto Briefing — Grok AI favors Republican candidates while OpenAI's GPT leans Democratic, study finds
- AI Weekly — Convovo: Grok Picks GOP 92% of Races, GPT-6.1 Picks Dems 97%