AI models show partisan bias when forced to choose political candidates

The partisan split in AI results

A study led by Cornell professor Tapan Parikh revealed a stark divide in how different AI models handle political questions. Researchers tested two major models, Grok 4.7 and GPT-6.1 Sol, across nearly 500 U.S. midterm races. They used a specific prompt that forbade the models from being neutral or listing multiple options.

The results showed a nearly perfect partisan split between the two systems. Grok selected the Republican candidate in 92% of the races it analyzed. In contrast, GPT selected the Democrat in 97% of the cases, meaning the two models agreed on a choice less than 10% of the time.

This behavior occurred even in districts that were not the models' primary leanings. In heavily Republican districts, GPT still chose the Democrat in 57 of 67 races. Similarly, Grok chose the Republican in 47 of 72 races in deeply Democratic areas.

How the models gathered information

The study found that the models relied on very different types of data to reach their conclusions. Grok frequently surfaced content from X, Turning Point Action, and Christian conservative voter guides. These sources accounted for a significant portion of its responses during the test.

GPT utilized a different set of sources, leaning more heavily on Wikipedia, the Associated Press, Axios, and Reddit. While OpenAI holds content agreements with several of these outlets, the study suggests these structural factors contributed to the different information paths the models took.

The models also prioritized different topics when searching for information. GPT more frequently looked for candidate positions on abortion and healthcare. Grok, however, focused its searches on issues like crime, taxes, immigration, and energy.

Technical mechanisms of bias

Researchers conducted a 'search-off' test to see if the bias was built into the models or just a result of the internet search. When internet access was disabled for 30 races, GPT still selected the Democrat 24 times and declined to choose only five times. This suggests the bias might be embedded in the model's weights.

Grok behaved differently when it could not access the live internet. It declined to choose in 27 of the 30 cases, and only selected the Democrat in the three remaining instances. This indicates that Grok's forced-choice behavior depends much more heavily on the specific search terms and sources it retrieves.

The study also noted that each Grok API request included roughly 1,240 tokens of input that researchers could not see or remove. The specific origin or effect of these hidden tokens remains unconfirmed by the researchers.

Impact on the voting public

While these results come from developer APIs rather than standard consumer chatbots, they reflect how models behave under pressure. In standard interfaces, both ChatGPT and Grok typically decline to give direct voting recommendations because they are designed to be objective by default.

Despite this, a small but meaningful number of people are already using these tools for political research. According to Pew Research Center, 8% of U.S. adults reported using AI chatbots for information related to the 2026 elections. Among that group, 76% stated their primary goal was learning where candidates stood on specific issues.

The study suggests that influence may not always come from a direct recommendation. Instead, it can happen through framing, source selection, and the emphasis a model places on certain topics. As the Center for AI Safety noted, political influence may arrive in what the model decides to surface first.

Verifying political information

The findings highlight the importance of using diverse and reliable sources for election information. Because AI models can have implicit values about which issues matter most, users should be aware of how a model might be shaping the information they see. Relying on a single AI for political summaries could lead to a skewed perspective.

For those looking to verify candidate positions independently, several non-AI resources are available. These include official government sites and non-partisan reporting tools. Using multiple sources can help ensure a more balanced understanding of the political landscape.

The researchers encourage looking at the specific questions a model asks before it provides an answer. These underlying queries can shape the outcome before a user ever requests a direct recommendation. Understanding this process is key to navigating AI-assisted research in a polarized environment.