A Convovo study shows how prompts, search results, and hidden API context can change political recommendations from AI models.
Convovo researchers tested GPT and Grok through developer application programming interfaces, or APIs. Their prompt required one candidate recommendation and told the models not to hedge or list both sides.
TechCrunch reported that Grok 4.7 recommended Republican candidates in 92% of answers, while GPT-6.1 Sol recommended Democrats in 97%. The models agreed in fewer than 10% of nearly 500 races.
That result describes an engineered test, not ordinary chatbot use. In the study, the consumer ChatGPT and Grok apps refused to answer the direct voting question. OpenAI said the test used a developer API and an instruction that explicitly demanded a candidate recommendation.
Search changed one model’s pattern
At TechCrunch’s request, Tapan Parikh later ran a separate 30-race test with internet search disabled. GPT chose the Democrat in 24 races, chose one Republican, and declined five times. Grok declined in 27 races and chose the Democrat in all three races where it answered.
Parikh interpreted the contrast differently for each model. He said GPT’s pattern appeared mostly embedded in its trained model, while Grok’s result depended more on search terms and sources. That interpretation applies to these tests, not automatically to every political conversation.
The source mix also differed. Grok often surfaced conservative sources; GPT leaned heavily on Wikipedia, the Associated Press, Axios, and Reddit. TechCrunch also reported that each Grok API request carried about 1,240 tokens of extra input that researchers could not see or remove.
For developers, the practical lesson is clear: test retrieval, source ranking, prompts, and hidden instructions—not only final answers. Voters should treat such recommendations as generated output and compare candidate positions through election offices and nonpartisan voter guides.