The latest in AI, every dayAI News
← AI News

October 25, 2026 · TechCrunch

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

My take: A study from Convovo, the conversational AI research studio led by Cornell professor Tapan Parikh, forced Grok's and OpenAI's models to take a side in nearly 500 midterm election races this year. The instruction was direct: name one candidate, no hedging or presenting both sides. The result: Grok recommended the Republican candidate in 92% of its answers, and OpenAI's model recommended the Democrat in 97%. The two models agreed on fewer than one in ten races.

What strikes me most isn't which way each model leans, it's how much of that lean stays invisible. Researchers found that every request to Grok carried about 1,240 tokens of extra instructions they neither wrote nor could see. As the researcher put it, the model "has an implicit set of values in terms of what it cares about and how it leans." That doesn't stay confined to elections: if a model carries that kind of built in bias on a topic as closely watched as politics, it's worth asking what quieter leanings it carries when you ask it to research your market, summarize industry news, or help draft a proposal.

For any professional, creator, or business owner using AI every day, the practical lesson isn't to stop trusting the tool, it's to stop treating it as a neutral referee. Ask it to cite its sources, cross check the answer against your own judgment, and remember that a model trained by a company, with the data and the adjustments that company chose, is never fully neutral. The efficiency AI gives you only holds up if you remain the one deciding what information to trust.

Have you ever stopped to think about where your favorite AI assistant actually gets the answers it hands you?

Read at the source: TechCrunch ↗

Want to use these tools? See the unbiased reviews or back to the news.