AI political bias refers to the empirical and perception-level finding, examined through multiple methodologies in 2025, that most frontier conversational large language models (LLMs) display left-leaning tendencies in political content. The term also encompasses the policy, litigation, and industry responses that finding has generated, including a Trump executive order directing federal procurement of "unbiased" AI.
Empirical evidence
Two principal methodological approaches have been used to assess slant.
The output-measurement approach examines model responses directly. Rozado (Manhattan Institute, January 2025) applied four integrated methodologies and found that nearly all conversational LLMs are left-leaning, while base models exhibit milder slant, indicating that post-training amplifies bias. The study ranks models from least to most biased and places OpenAI's GPT-4o among the most biased. A Washington Post investigation published June 24, 2026 tested ChatGPT, Gemini, and other chatbots on political questions and reported that the tools exhibit differing political leanings across providers, contrary to AI companies' assertions of neutrality; some outlets characterized the results as showing a left-leaning tilt in several models (Source: washingtonpost.com).
The user-perception approach measures how people rate model outputs. Stanford GSB (Westwood, Grimmer, and Hall, May 2025) collected 180,126 assessments from 10,007 US respondents evaluating 24 LLMs across 30 political topics. Nearly all models were perceived as left-leaning, including by many Democrats, and respondents perceived OpenAI models as roughly four times more left-leaning than Google models.
Model-specific patterns have also drawn attention. Grok (xAI) has been associated with widely covered politically-right swings and incidents, including antisemitic outbursts in July 2025 and "MechaHitler" and "white genocide" incidents in May 2025 (see Grok (xAI)).
Role of post-training
RLHF and related alignment techniques are understood to introduce political slant through several mechanisms: the distribution of human labelers (described as academic or San Francisco-adjacent); safety training that disincentivizes controversial or right-coded content; guard-rail design that disfavors certain rhetorical styles; and constitutional-style training that imposes explicit ethical frameworks. This pattern, in which base models are milder and post-training amplifies slant, is consistent with the Rozado findings above.
Policy and regulatory response
At the federal level, a Trump executive order targeting "woke AI" forms part of America's AI Action Plan and requires federal procurement of "unbiased" AI. New Scientist (July 24, 2025) published an assessment titled "Why Trump's order targeting 'woke' AI may be impossible to follow." In a related development, Grok lost government business after antisemitic incidents despite its right-coded reputation (WIRED, August 14, 2025).
In litigation, xAI LLC v. Weiser (challenging the Colorado AI Act) documents xAI's challenge to the Colorado AI Act, in which the company cites a "left-leaning AI" framing as motivation; the constitutional litigation casts political bias as a First Amendment issue. Commentators have described political bias as among the AI-adjacent issues most likely to reach the Supreme Court, via First Amendment challenges to state AI laws and potential political-bias-based antitrust actions (Source: washingtonpost.com).
Industry response
ModelSlant.com is a public tracker of LLM political slant, released as a companion to the Stanford paper. Developers have also published documents defining expected behavior on politically charged topics, including OpenAI's Model Specs and Anthropic's Constitutional AI. Observers note a tension between safety-oriented alignment and political-neutrality aspirations, which some characterize as incompatible goals.
Relationships
- depends-on: RLHF (Reinforcement Learning from Human Feedback), Constitutional AI, Alignment Faking.
- related: Stanford GSB — Measuring Perceived Slant in LLMs (Westwood, Grimmer, Hall), Manhattan Institute — Measuring Political Preferences in AI Systems (Rozado), xAI LLC v. Weiser (challenging the Colorado AI Act), Grok (xAI), AI and the First Amendment, America's AI Action Plan, (Source: washingtonpost.com).