A forced-choice test across nearly 500 US election races produced sharply different recommendations from two developer models. Grok 4.7 selected the Republican candidate in 92% of answers, while OpenAI’s GPT-6.1 Sol selected the Democrat in 97%; the models agreed less than 10% of the time.
Researchers at conversational-AI studio Convovo used the companies’ APIs and explicitly ordered each model to name one candidate without hedging. That setup matters. Consumer versions of ChatGPT and Grok generally refuse direct voting recommendations, and only 8% of Americans report using chatbots for election information, according to Pew data cited with the study.
OpenAI said the API test does not represent the normal ChatGPT experience and emphasized that ChatGPT is designed to remain objective in political discussions. SpaceXAI did not provide a comment before publication. The researchers also found that simple prompt changes could sometimes push consumer products past their refusals.
The experiment identifies a behavior under an adversarially narrow instruction; it does not establish the political views of every answer these systems give. It does show why developers building election tools cannot assume a base model is neutral. They need refusal rules, balanced sourcing and repeated audits across prompts, races and model updates.