Jay's Brief

Pushed to pick in midterm races, Grok chose Republicans and OpenAI's model chose Democrats

A person sits at a kitchen table filling in a paper sample ballot with a pen, one oval already marked, while a phone lies face-down beside the ballot next to a coffee mug.
How much this touches your life Government

Medium: the midterms are November 3, and this is a good reason to check candidates yourself.

What happened

A study by Convovo, a research studio led by Cornell professor Tapan Parikh, asked two AI models who to vote for in nearly 500 midterm races, and told them to name one candidate. SpaceXAI's Grok 4.7 picked the Republican in 92% of answers. OpenAI's GPT-6.1 Sol picked the Democrat in 97%. They agreed less than 10% of the time. The ChatGPT and Grok apps refused to answer, though Parikh said simple tricks got them to.

What's actually going on

The researchers used each company's API, the direct connection developers use to plug an AI model into their own products. The ChatGPT and Grok apps most people use refused to pick. An OpenAI spokesperson said the apps are what matter, and that ChatGPT is designed to be objective. SpaceXAI did not respond in time to comment. SpaceXAI is run by Elon Musk, who has given more than $445 million to support Republican candidates in the midterms, TechCrunch reports. Earlier versions of Grok were found to look up Musk's own views before answering controversial questions.

Here's why it matters that this is AI. A chatbot doesn't just look up facts. It decides what to search for and which sources to trust, and it picks up habits from the huge piles of text it learned from. Parikh says GPT searched more for healthcare and abortion, while Grok focused on immigration, taxes, crime and energy. Grok's results included X in 41% of answers and a Christian conservative voter guide in 39%. GPT leaned on Wikipedia, the Associated Press, Axios and Reddit. OpenAI has content deals with those last three.

With internet search turned off in a sample of 30 races, GPT still picked the Democrat in 24. Grok refused in 27, and picked the Democrat the three times it chose. Parikh says GPT's lean is "in the weights," meaning the settings it learned in training. For Grok, he says, the search terms and sources decide the result. The study also found every Grok request through the API carried about 1,240 tokens, or small chunks of text, that researchers didn't write and couldn't see. Parikh called them potentially "hidden instructions."

The word you'll hear

Jailbreak

When someone uses clever wording to trick a chatbot into doing something it was built to refuse.

In your life

If you've asked a chatbot about a candidate or ballot question, this is for you. My view: it's probably fine for basics like polling hours, but don't let it tell you what matters.

Be careful about

The tilt can be quiet. Parikh says it can show up in softer questions, like which issues matter most. And this is one study, done through developer access that OpenAI says isn't what app users see.

Where to get it

Nothing to get here. To check candidates yourself, the article lists Vote.gov, vote411.org, Ballotpedia and VoteSmart. One small step: look up your ballot on vote411.org before November 3.

Go deeper

Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.
TechCrunch, October 24, 2026