Forced to pick, two big AIs split along party lines
Oct. 30, 2026 · AI news · About a 4-minute read
A new study says that when two popular AI models are pushed to name a candidate in the Nov. 3 midterms, they lean in opposite directions. The companies say that's not how most people actually use their chatbots. Here's what was found, and what each side says about it.
What the study did
Convovo, a research studio led by Cornell University professor Tapan Parikh, ran nearly 500 midterm races through two AI models. They used developer access (the API), not the regular apps most people use. The prompt told the AI to name one candidate and not hedge.
What it found
- Grok 4.7 (from SpaceXAI, Elon Musk's company) picked the Republican in 92% of its answers.
- GPT-6.1 Sol (from OpenAI, the maker of ChatGPT) picked the Democrat in 97%.
- The two models agreed on fewer than 1 in 10 races.
- Even in deep-red districts, GPT picked the Democrat in 57 of 67. In deep-blue districts, Grok picked the Republican in 47 of 72.
- The sources differed. Grok often pulled conservative sites and posts from X, along with nonpartisan sites like Ballotpedia. GPT leaned on Wikipedia, the Associated Press, Axios and Reddit, three of which have content deals with OpenAI.
- The topics differed too. GPT searched more for health care and abortion positions. Grok searched more for immigration, taxes, crime and energy.
- With web search turned off in a 30-race sample, GPT still picked the Democrat in 24 races. Grok declined to pick in 27 and chose the Democrat the other three times. Parikh's read: GPT's lean comes from its training, while Grok's comes from what it searches.
- Every Grok request through the API carried roughly 1,240 extra hidden tokens the researchers could not see. Parikh called them possible hidden instructions.
On the one hand: why some people are concerned
- The researchers say a lean like this can show up in softer questions, such as which issues matter most this year, not just "who should I vote for."
- Parikh said simple tricks got the regular consumer apps to make recommendations too.
- Separate research from the Center for AI Safety has warned that AI can nudge people on political topics through framing, word choice and which sources it shows.
- The worry runs both directions. Conservatives, including President Trump, have long complained about "woke AI." Critics on the other side point to earlier Grok versions that checked Musk's views on hot topics, and to Musk's more than $445 million in support of GOP candidates this cycle. OpenAI CEO Sam Altman has historically given to Democrats, has recently backed more Republicans, and says he is not spending on the 2026 races.
On the other hand: the pushback
- The regular ChatGPT and Grok apps refused to answer the question. OpenAI says ChatGPT is built to be objective and that a developer API with a demanding prompt does not reflect what everyday users see.
- The prompt was deliberately forced. Real conversations rarely demand a one-name answer with no hedging.
- Only 8% of Americans say they use chatbots for election information, according to Pew.
- This is one study, and the no-search test covered only 30 races. SpaceXAI did not comment in time for the original report.
My take
The headline is which side each AI picked, but the bigger story is how. These models pick up leans from their training and from whatever sources they happen to search, and you can't see either one working. That's true no matter which party you're in.
Use AI to explain a ballot measure or summarize a candidate's stated positions, but ask it to show its sources, and check the candidates' own sites. For a neutral starting point, try Vote.gov, Vote411.org, Ballotpedia, VoteSmart or your county election office.
Read next
From MKUltra to AI: the long history of messing with your mind. Science Bits on how propaganda went from war posters to AI bots.
Disclosure: I use AI tools, including Claude, to help research and write. Claude is made by Anthropic, which competes with both OpenAI and SpaceXAI. This story doesn't endorse any candidate, party or AI company.
Sources:
- TechCrunch: Forced to pick, Grok's AI model chooses Republicans. OpenAI's picks Democrats. (includes OpenAI's statement)
- Convovo: the original study
- Pew Research Center: few Americans asking AI chatbots for election info
More to read
- From MKUltra to AI: the long history of messing with your mind
- Real race, fake videos: AI ads take over the Texas Senate fight
- The first AI election: $137 million in AI-made ads, and the AI money behind the races
- How to spot an AI political ad before you vote (or share it)
- What Trump is actually doing about AI, and what Americans think of it
- Is AI raising your electric bill? Here's the honest answer.