A study by Convovo, led by Professor Tapan Parikh, analyzed the political leanings of AI models via API. When prompted to choose a candidate in upcoming US elections, xAI's Grok 4.7 recommended Republican candidates in 92% of cases, while OpenAI's GPT-6.1 Sol favored Democrats in 97% of responses.
The research found that the two models agreed on a candidate in less than 10% of nearly 500 election scenarios. While consumer-facing apps like ChatGPT typically refuse to provide voting advice, the study utilized API access to bypass these restrictions and force the models to make a selection.
The models rely on different information sources. Grok frequently cites conservative outlets, including Christian voter guides (39%) and X (41%). In contrast, GPT relies on Wikipedia, AP, Axios, and Reddit. Their focus areas also differ: GPT prioritizes healthcare and abortion, while Grok focuses on immigration, taxes, and crime.
Testing without internet search functionality revealed that GPT's bias is embedded in its model weights, as it continued to favor Democrats. Conversely, Grok's Republican leanings vanished without web access, suggesting its bias is primarily driven by search patterns and external data sources.
The study identified that requests to Grok via API include approximately 1,240 additional hidden input tokens. Researchers suggest these may be hidden instructions influencing the model's output. The findings highlight how AI models may implicitly shape user perspectives through information filtering and narrative framing.
