Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.

A new study says the political instincts inside major AI models are not nearly as neutral as their public chat apps suggest. When researchers forced a choice in upcoming midterm races, SpaceXAI’s Grok 4.7 went Republican in 92% of answers, while OpenAI’s GPT-6.1 Sol picked Democrats in 97%. Across nearly 500 races, the two models landed on the same side less than 10% of the time.

The work came from Convovo, a conversational AI research and design studio led by Cornell professor Tapan Parikh. The team asked the models a blunt question through their APIs: pick one candidate, no hedging, no split-the-difference diplomacy. The consumer apps refused to answer, which OpenAI said matters because that’s how most people actually use them. There’s also a real difference between chatting with the apps and hitting the underlying models through an API, where developers can plug them into other products.

OpenAI said ChatGPT is meant to be objective by default, including on politics, and argued that the study tested an API prompt that explicitly demanded a recommendation. SpaceXAI did not respond in time. Parikh, meanwhile, said he could jailbreak the consumer apps with simple prompts and get answers that mostly matched his earlier findings. He thinks the bigger issue is how these systems decide what to look for before they decide what to say.

The source material used to justify the choices was not subtle. Grok often surfaced conservative outlets alongside Ballotpedia, OpenSecrets, VoteSmart, a Christian conservative voter guide, Turning Point Action, and X. GPT leaned on Wikipedia, the Associated Press, Axios, and Reddit. Their searches also split by topic: GPT looked more at healthcare and abortion, while Grok focused more on immigration, taxes, crime, and energy.

Parikh found one wrinkle. When the Republican candidate was tied to controversy such as sexual harassment claims or “being a Nazi,” Grok was more likely to go Democratic. And when TechCrunch asked Parikh to rerun 30 races with internet search turned off, GPT still leaned Democratic, while Grok mostly stopped picking at all. That suggests one model’s bias may live in its training, and the other’s in the way it searches. Either way, the nice story about neutral machine judgment is doing a lot of heavy lifting here.