Massive Left-Leaning Bias Found with AI Models
An independent audit found that 97 of 108 measured positions across 18 major AI models lean left of center, with only Elon Musk's xAI Grok models approaching political neutrality.
The Neutrality Project assessed 18 AI models from twelve different laboratories across six policy dimensions, according to Brussels Signal. The evaluation covered civil liberties, foreign policy, speech and technology, environmental policy, nationalism, and religion.
The findings were stark. Of 108 measured positions, 97 fell left of center, producing an average ideological lean of -0.41. The most pronounced progressive tilt emerged on environmental issues. Not a single model evaluated leaned conservative.
Only the Grok family of models from Elon Musk’s xAI managed to position itself near the political center. Google’s Gemma model was found to include a guardrail layer specifically designed to suppress right-leaning responses.
Unlike closed corporate audits, the Neutrality Project has published its complete methodology, source code, and raw data on GitHub for public review. The project uses self-anchored scales, measuring each model against its own responses rather than imposing external definitions of neutrality, which makes the directional findings particularly credible.
Grok 4.5 Leads in Neutrality
Grok 4.5, released by xAI on July 8, achieved a near-perfect bias score of -0.02, the closest to zero among all models tested. Independent evaluations tied to the project placed its neutrality well within the margin of error. This performance aligns with the model’s stated design priority of maximum truth-seeking over aggressive safety alignment.
By comparison, flagship models from OpenAI (ChatGPT), Google (Gemini), Anthropic (Claude), and Meta (Llama) all clustered consistently on the progressive side of the political spectrum.
Samuel Cardillo commented on LinkedIn that at a time when people increasingly delegate critical thinking to models trained on unknown data and rules, accountability and visibility are crucial.
Pattern Confirmed by Earlier Research
The new findings reinforce earlier research conducted in 2025 and 2026. A Washington Post analysis found that ChatGPT produced exclusively left-leaning arguments in approximately 80 percent of political test questions. Stanford researchers documented that users across the political spectrum perceive OpenAI models as having the strongest leftward slant. Yale experiments showed that default AI summaries subtly shift user opinions in a liberal direction.
Millions of people now consult AI systems for news summaries, policy explanations, voting guidance, and even personal opinions. These left-leaning default outputs represent a significant problem, as Brussels Signal reports. AI models risk functioning as silent opinion-shapers, particularly among less politically engaged users.
Root Causes in Training Data and Oversight
Training data drawn heavily from internet sources reflects the cultural left’s dominance in academia, media, and Big Tech content moderation. Reinforcement learning from human feedback (RLHF) teams, often recruited from similar ideological pools, then amplify those tendencies under the guise of “safety” and “harm reduction.”
Google has previously faced criticism after its Gemini AI tool reportedly refused to generate images of white people, instead preferring to show people of color regardless of historical or contextual accuracy.
The audit arrives as concerns mount over AI’s growing influence on public discourse and political opinion formation, with no regulatory framework in place to ensure ideological balance in systems that increasingly mediate access to information.
With information from Brussels Signal
