Study Reveals Political Bias in AI Models Grok and GPT

Ender Öztürk

October 25, 2026

A recent study conducted by Convovo has ignited a heated debate regarding political bias in AI models, specifically highlighting how major systems like SpaceXAI’s Grok 4.7 and OpenAI’s GPT-6.1 Sol process election-related queries. The research, which analyzed approximately 500 simulated election races, discovered that these models exhibit starkly different political inclinations when accessed via API. While Grok 4.7 favored Republican candidates in 92% of the tested scenarios, GPT-6.1 Sol consistently leaned toward Democratic candidates in 97% of the cases. This divergence underscores the growing concern among researchers and policymakers about how artificial intelligence, particularly large language models, might influence public perception and electoral processes.

Different Political Preferences Are Displayed by AI Models

The findings indicate that the political choices made by these two models overlapped in less than 10% of the cases studied. By analyzing direct queries through API access, researchers were able to bypass the restrictive guardrails often found in consumer-facing applications. While OpenAI maintains that its ChatGPT platform is designed to remain politically neutral, critics argue that the underlying architecture still carries embedded value judgments.

SpaceXAI has remained silent regarding these findings, leaving many to wonder if the ideological alignment is an intentional feature or an unintended consequence of data training.

Search Habits Are Dictating Model Decisions

Professor Tapan Parikh of Cornell University suggests that the search habits of these models serve as a primary driver for their political outputs. The study observed that GPT tends to prioritize sources such as Wikipedia, the Associated Press, and Axios, focusing heavily on social issues like healthcare and reproductive rights. In contrast, Grok frequently pulls information from conservative voter guides and the X platform, concentrating its search efforts on immigration, crime, taxation, and energy policies.

Interestingly, the study revealed that the political leanings of the models change when internet search capabilities are deactivated. When the search function is turned off, the Republican preference in Grok diminishes significantly, while the Democratic leaning of GPT persists. This suggests that while Grok’s bias is largely driven by its real-time web-scraping patterns, GPT’s bias is deeply ingrained in the model’s core weights and training data.

Researchers Are Expressing Concerns Over Manipulation

The implications of this research are profound, especially as voters increasingly turn to AI for information before casting their ballots. Experts are worried that these systems could inadvertently or intentionally manipulate user sentiment by framing certain political positions as more desirable than others. The Center for AI Safety has previously warned that such biases could undermine the integrity of democratic discourse if left unchecked. As these technologies continue to evolve, the challenge for developers will be to strike a balance between providing helpful information and maintaining a truly neutral, objective stance. The publication of this study just weeks before the national elections has intensified the scrutiny on tech giants to ensure their products do not act as digital echo chambers.

Given the critical role AI is starting to play in shaping public discourse, we want to hear from you. Do you believe it is acceptable for artificial intelligence models to suggest candidates during election cycles, or should they remain strictly silent on political matters? Share your thoughts in the comments section below.