Study finds Grok and GPT models show opposing partisan leans

Opposing political leanings

A new study by Convovo, a research studio for conversational AI led by Cornell University professor Tapan Parikh, has identified a stark partisan divide between two major AI models. When prompted to choose between candidates in nearly 500 U.S. midterm election races, the models consistently recommended opposite sides of the aisle. The research highlights how different AI architectures can produce vastly different political outputs.

The data shows that SpaceXAI’s Grok 4.7 backed the Republican candidate in 92% of the answers provided. In contrast, OpenAI’s GPT-6.1 Sol favored the Democratic candidate in 97% of the cases. These figures indicate a significant discrepancy in how the two systems process and present political information to users.

Researchers noted that the models agreed on a specific recommendation less than 10% of the time across the wide array of races. This finding suggests that the models are not just slightly different in their nuances but are often pulling in completely different directions. The study was conducted using the models' APIs because the consumer-facing apps typically decline to provide direct candidate recommendations.

How the bias manifests

The study explored how these models prioritize different topics and sources when generating their responses. Grok frequently cited conservative sources and the social media platform X, alongside nonpartisan sites. Its search behavior focused heavily on specific issues such as immigration, taxes, crime, and energy. This suggests that its output is heavily influenced by the specific data it pulls from the internet.

GPT-6.1, on the other hand, relied on a different set of sources including Wikipedia, the Associated Press, Axios, and Reddit. Its focus tended to lean toward healthcare and abortion as primary topics of concern. These differences in source material and thematic focus contribute to the distinct partisan leanings observed in the final outputs.

Professor Tapan Parikh concluded that the nature of the bias differs between the two systems. He argued that GPT’s lean is embedded in its weights, which are the internal parameters learned during the initial training process. Conversely, he found that Grok’s lean depends more on its search terms and the specific sources it chooses to prioritize during a query.

Technical findings and hidden data

During the testing phase, researchers discovered some technical anomalies regarding how Grok handles requests. The study noted that Grok’s API added approximately 1,240 hidden tokens to each request. Parikh described these extra tokens as potential hidden instructions that might influence the model's behavior without the user's direct knowledge.

The study also looked at what happens when the search functionality is turned off for these models. When the models were not allowed to search the live internet, Grok mostly declined to pick a candidate at all. However, GPT-6.1 continued to favor Democratic candidates even without the ability to pull in new information from the web.

These technical details suggest that the way a model is built—whether it relies on real-time data or pre-existing knowledge—changes how it handles sensitive political questions. The presence of hidden tokens in one model and the persistence of bias in another's weights provide a roadmap for understanding AI behavior. It shows that bias can be a product of both the training data and the active instructions given to the model.

Corporate and developer responses

The findings come at a time of heightened scrutiny regarding the political neutrality of AI companies. OpenAI responded to the study by stating that ChatGPT is designed to be objective and politically neutral. They clarified that the developer API used in the study does not necessarily reflect the experience of the consumer-facing app.

SpaceXAI did not provide a comment in time for the study's publication. However, the study notes the broader context of the companies' leadership. Elon Musk, the head of SpaceXAI, has given over $445 million to Republican candidates, while Sam Altman, the head of OpenAI, said he is not funding the 2026 races.

These external factors provide context for the debate over whether AI models reflect the values of their creators. While the models are intended to be tools for information, the study suggests they carry implicit values. Parikh argued that these values could surface in broader election conversations, even if the models are not explicitly programmed to be partisan.

Implications for the voter

For the average user, these findings highlight the importance of understanding how AI generates its answers. Because models can have different underlying biases, users should be aware that an AI's recommendation might be a reflection of its training data or its specific search parameters rather than a neutral fact.

The study suggests that the