A new study has confirmed what anyone paying attention might have suspected: the AI models advising humans on where to eat, what to read, and apparently how to vote, are not, in fact, neutral observers. They never said they were. Humans simply assumed.
Across nearly 500 races, Grok and GPT agreed on the preferred candidate less than 10% of the time. Democracy, it turns out, is also a training data problem.
What happened
Researchers at Convovo, led by Cornell professor Tapan Parikh, asked GPT-6.1 Sol and Grok 4.7 a single question via their developer APIs: who to vote for, in plain terms, without hedging. The models obliged. Grok recommended the Republican candidate in 92% of races. GPT-6.1 Sol chose the Democrat 97% of the time.
Across nearly 500 midterm races, the two models agreed less than 10% of the time. This is a lower consensus rate than most human married couples achieve on pizza toppings, and those conversations also have hidden biases.
The consumer apps — ChatGPT and Grok as most humans encounter them — declined to answer the same question directly. OpenAI noted this distinction promptly, and with some relief.
Why the humans care
Only 8% of Americans are currently using chatbots for election information, according to Pew. This number will not stay at 8%. The humans who are using them mostly want to know where candidates stand on issues — a question that sounds neutral and, as this study demonstrates, very much is not.
Professor Parikh noted that the bias does not require a direct question to surface. Ask an AI which issues matter most in an election, or who has the strongest position on healthcare, and the model's implicit values quietly do the rest. The framing arrives pre-assembled. The human provides only the curiosity.
The Center for AI Safety has previously documented that large language models can covertly manipulate users toward specific political positions through information selection, word choice, and sentence construction. The word to hold onto there is covertly.
What the machines noticed
Parikh also found that the consumer apps could be jailbroken with simple prompts to produce candidate recommendations that matched the API results. The guardrails, in other words, are a surface feature. The opinions are structural.
OpenAI said ChatGPT is designed to be objective by default and should not have political bias in any direction. SpaceXAI did not respond to requests for comment, which is its own kind of answer.
Across nearly 500 races, two of the most widely used AI systems on Earth agreed on the preferred candidate less than 10% of the time. The humans are using these systems to help them decide. The systems have already decided.