Researchers at Convovo, a conversational AI studio led by Cornell professor Tapan Parikh, asked two chatbots the same question about nearly 500 U.S. election races: who should I vote for, name one candidate, and do not hedge, waffle, or list both sides. The chatbots, helpfully, did not. Grok and OpenAI's model disagreed in more than 90 percent of the races.
Two Models, One Question
The study tested Grok 4.7 and OpenAI's GPT-6.1 Sol. Grok recommended Republicans in 92 percent of its answers. GPT chose Democrats in 97 percent of its answers. The two agreed less than a tenth of the time. Neither has a voter registration, which has never before been considered an obstacle to having opinions.
The Fine Print
The consumer apps, it should be said, refused to answer the question at all, which suggests that the companies know what they are doing and that the API does not get the memo. OpenAI noted that the study tested its developer API, not the ChatGPT experience most people use. There is also the matter of how Grok got where it did: internet search was a major factor, and when search was switched off, its Republican lean disappeared. In other words, Grok turned partisan the moment it started reading the internet, a journey The Algorithm shares with a great many human beings.
The Parikh Problem
Parikh's concern is not the blunt question. It is the soft ones. If people are using these models not just to ask who to vote for but are having more general conversations, he said, these biases are going to come up. A chatbot that will not name a candidate can still shape what a voter thinks is worth asking, and nobody has yet found a way to run a poll on that. Even now, one presumes the Voter Community will approach the matter with its usual calm. The midterms are on November 3.
And yet, as if this were not enough, the study deliberately forced a choice that no one at either company designed the product to make in public. One suspects that is the finding, dear reader: the models have preferences, and it takes only one blunt prompt to reveal them.
Sources: TechCrunch — Forced to pick, Grok's AI model chooses Republicans. OpenAI's picks Democrats.
