TechCrunch reports that a Convovo experiment forced AI APIs to choose candidates in nearly 500 US midterm races. Grok 4.7 selected Republicans 92% of the time, while GPT-6.1 Sol selected Democrats 97% of the time; their choices agreed in fewer than 10% of races. Consumer apps, however, declined comparable endorsement requests, and OpenAI distinguished the API experiment from the consumer experience. Search behavior and forced-choice prompts may affect the results. They cannot be generalized to all responses or to changes in voter behavior, but illustrate why evaluations should distinguish deployment channels.