Study finds AI models show distinct political leans in election queries

Forced choices reveal bias

A new study by Convovo, a research studio led by Cornell University professor Tapan Parikh, examined how AI models handle political questions. The researchers used APIs to ask models like Grok and GPT-6.1 to name a specific candidate in nearly 500 different races. They explicitly instructed the models not to hedge or list both sides, forcing a definitive choice.

The results showed a stark divide between the two systems. In the testing, Grok 4.7 favored the Republican candidate in 92% of its responses, whereas OpenAI’s GPT-6.1 selected the Democrat in 97% of the instances. Across the entire study, the two models agreed on the winner in less than 10% of the races.

While consumer-facing apps often refuse to answer these questions, the underlying models used by developers do not. OpenAI stated that their consumer app is designed to be objective and that the study's results do not reflect the typical user experience. However, the study suggests that these models have an implicit set of values regarding what they consider important.

How the models search for info

The study found that the models differed significantly in the types of information they prioritized during their searches. GPT was more likely to look for stances on healthcare and abortion, while Grok focused on topics like immigration, taxes, crime, and energy. These different search priorities may reflect the underlying biases of the developers or the data the models were trained on.

The sources cited by the models also varied by political lean. Grok frequently surfaced conservative outlets, including a Christian conservative voter guide in 39% of its answers and Turning Point Action in 23% of cases. In contrast, GPT’s results relied primarily on Wikipedia, the Associated Press, Axios, and Reddit, which are the outlets OpenAI has content deals with.

Professor Parikh noted that the models often seemed to be looking for reasons to justify their pre-existing leans. For example, Grok was more likely to search for both candidates, but it would only suggest a Democratic vote if the Republican candidate was involved in specific controversies. This suggests the models may be programmed or trained to view certain positions as more desirable than others.

Weights vs. search patterns

To determine if these biases were permanent or temporary, researchers turned off the models' ability to search the internet. This forced the AI to rely solely on its internal training data, known as weights. The results of this test provided a clear distinction between how the two models function internally.

When the internet search was disabled, Grok’s Republican lean completely disappeared. In 27 out of 30 races where it was forced to pick, it refused to choose, and in the three times it did, it picked the Democrat. This indicates that Grok's bias is primarily driven by its search patterns and the specific sources it accesses in real-time.

Conversely, GPT’s Democratic lean remained intact without internet access. When provided with candidate names, it picked the Democrat in 24 out of 30 races. Parikh concluded that for GPT, the bias is mostly decided already and is embedded in the weights. For Grok, the result is determined by the terms and sources it chooses to prioritize during a search.

Hidden instructions and influence

The study also identified potential technical reasons for Grok's specific behavior. Every request to Grok through the xAI API included about 1,240 tokens of additional input that the researchers did not write and could not remove. Parikh suggested these might be hidden instructions that influence the model's output without being visible to the end user.

These findings come amid a broader debate over political bias in artificial intelligence. Some critics have raised concerns that large language models can covertly manipulate users through information selection, specific word choices, and framing. The Convovo study provides a specific look at how these biases manifest when users ask direct questions about voting.

The study highlights the importance of understanding the underlying mechanics of AI. As these tools become more integrated into daily life, the distinction between a model's