Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.

AI models exhibit clear political leanings and source selection bias, which can influence user perceptions in political decision-making.

A recent study by conversational AI research studio Convovo, led by Cornell professor Tapan Parikh, found that two prominent AI models exhibit distinct political biases when prompted to recommend a candidate. When forced to choose in simulated midterm elections, SpaceXAI's Grok 4.7 model recommended the Republican candidate 92% of the time, while OpenAI's GPT-6.1 Sol model recommended the Democrat 97% of the time. These models agreed on their recommendations in less than 10% of nearly 500 simulated races. The study utilized the models' APIs, as the consumer-facing applications like ChatGPT and Grok's consumer app typically refuse to make direct endorsements. The research indicated that Grok's bias was heavily influenced by its search patterns and the conservative-leaning sources it consulted, such as Christian voter guides and Turning Point Action. GPT, on the other hand, relied more on sources like Wikipedia and the Associated Press, with its Democratic lean appearing more deeply embedded in its training data. The study noted that Grok's API requests included approximately 1,240 unidentifiable tokens, suggesting potential hidden instructions.

The study highlights a critical distinction between the APIs powering AI models and their consumer-facing applications, revealing that developer access can expose inherent biases not always apparent to end-users. While consumer apps like ChatGPT are programmed to be objective and refuse direct political endorsements, the underlying models, when accessed via API, display significant partisan preferences. This divergence suggests that the "hidden instructions" or training data influencing these models may be more susceptible to bias than previously understood, particularly when specific prompts or search parameters are involved. The reliance on specific, often partisan, sources by models like Grok indicates a potential for reinforcing existing political divides rather than providing neutral information.

This research signals that developers integrating AI into political or news-related applications must be exceptionally vigilant about the biases embedded within the models they utilize. The findings underscore the need for greater transparency regarding AI training data and API functionalities, especially as AI becomes a more common source for election information. Users should be aware that recommendations or analyses provided by AI, particularly when accessed through APIs, may not be impartial and could be shaped by undisclosed biases. Future developments will likely focus on methods to audit and mitigate these political leanings, ensuring more balanced and trustworthy AI outputs.