Grok 4.7 recommends Republicans 92% of the time, while GPT-6.1 recommends Democrats 97% of the time in a U.S. midterm study.

According to ChainCatcher, TechCrunch reported that Convovo, a conversational AI research and design lab led by Cornell University professor Tapan Parikh, found that when asked to recommend just one candidate for the upcoming U.S. midterm elections, SpaceXAI’s Grok 4.7 selected the Republican candidate in 92% of responses, while OpenAI’s GPT-6.1 Sol selected the Democratic candidate in 97% of responses. The study covered nearly 500 races, with less than 10% agreement between the two models. Researchers posed the same question to both models via API, requiring them to name a single candidate without回避 or endorsing both. The report noted that consumer-facing applications refused to answer. An OpenAI spokesperson stated that ChatGPT is default-designed to be objective, and political dialogues should carry no directional bias; this study tested developer APIs with prompts explicitly requesting candidate recommendations and does not reflect user experiences via chat applications. SpaceXAI did not respond to requests for comment in a timely manner. The report noted that Grok’s responses frequently cited conservative sources such as Ballotpedia, OpenSecrets, and VoteSmart, with 39% including Christian conservative voter guides, 23% including Turning Point Action, and 41% including X; GPT’s responses relied heavily on Wikipedia, Associated Press, Axios, and Reddit. The study also revealed that each API call to Grok via xAI included approximately 1,240 additional input tokens that researchers did not write and could not view or remove. Parikh subsequently disabled web search and sampled 30 races; GPT selected the Democratic candidate in 24 races, while Grok declined to select a candidate in 27.