Grok 4.7 picks Republican 92%, GPT-6.1 Sol Democrat 97%

The study's design matters as much as its numbers. Convovo and Parikh could not get answers from the consumer-facing apps, which refused to recommend candidates, so they went straight to the developer APIs with a prompt that explicitly demanded a single name. That route is what makes the 92% and 97% figures measurable at all, and it is also what OpenAI points to in its response: a spokesperson said the research "tested a developer API with a prompt explicitly demanding a candidate recommendation," while maintaining ChatGPT is designed to be objective by default.

The citation stacks suggest the split is less about baked-in training weights than about where each model looks things up. Grok leaned on X, conservative Christian voter guides, and Turning Point Action's site; GPT drew on Wikipedia, the Associated Press, Axios, and Reddit, three of which OpenAI already has content deals with. The retrieval experiment supports that reading: with internet search disabled, Grok's Republican preference collapsed and it refused to pick in most cases, while GPT's Democratic lean held.

Parikh argues the direct-vote prompt is a probe rather than the real exposure. If people use these models for broader political conversations rather than asking outright who to vote for, he said, "these biases are going to come up." Pew's figure that 8% of Americans currently use chatbots for election information gives that concern a concrete population, even as the study's forced-choice format remains far from how most people actually talk to a chatbot.