Study Finds Stark Political Bias in AI Models

A new study has found sharp differences in how two leading AI models respond to political prompts, with Grok overwhelmingly recommending Republican candidates and OpenAI's GPT favoring Democrats. The two models disagreed in more than 90 percent of nearly 500 election races examined. The results add to concerns about political bias in AI assistants that voters may turn to for information about candidates and elections. The study does not establish why the models diverge. Possible factors include training data, system prompts, fine-tuning and the alignment choices made by developers. Even small differences in how a model is instructed or how it weighs sources can shape the answers it gives, especially on politically sensitive questions. The findings therefore highlight a broader governance challenge: as chatbots become everyday interfaces for news and research, their outputs can carry an appearance of neutrality that may not be justified. For voters, the practical risk is not simply that a model has a preference. It is that users may treat a fluent answer as a reliable summary. If an assistant consistently favors one party's candidates or frames races in a particular way, it could influence perceptions without disclosing the basis for its recommendations. The disagreement between Grok and GPT also suggests that there is no single neutral default across the AI industry. The study's authors do not claim to explain the cause of the bias, and more research will be needed to test how prompts, model versions and use cases affect results. Developers may respond with clearer disclosures, stronger evaluation methods and safeguards for election-related queries. Political campaigns and regulators are likely to pay close attention. The larger lesson is that AI models are not passive mirrors of public information. They are products shaped by design decisions, and those decisions can have political consequences. As election information increasingly flows through chatbots, transparency about model behavior and independent testing will be essential to public trust.