Study Finds Significant Partisan Divergence in Leading AI Chatbots
A new research study reveals that OpenAI’s GPT and xAI’s Grok provide starkly different political recommendations when tested against nearly 500 election races.
Researchers at the AI studio Convovo have analyzed the responses of GPT-6.1 and Grok 4.7 to determine their political leanings. By accessing the models through developer APIs, the team prompted the systems to select candidates without hedging. The results showed that Grok 4.7 supported Republican candidates in 92 percent of the races, while GPT-6.1 recommended Democrats in 97 percent of cases.
The divergence appears rooted in how each model accesses information. GPT prioritized topics such as healthcare and reproductive rights, relying on sources like Wikipedia and mainstream news outlets. Grok focused on immigration, taxation, and crime, frequently citing the social media platform X and conservative voter guides.
Further testing indicated that the bias in GPT appears to be inherent to the model’s core weights, as it persisted even when web search capabilities were disabled. Conversely, Grok’s leanings appear more dependent on its search parameters and source selection. These findings suggest that developers integrating these models into commercial applications may inadvertently inherit the specific biases present in the underlying training data and search configurations.