Report Finds AI Chatbots Exhibit Left-Wing Bias

The Washington Post published a controlled test of major chatbots and concluded that the model powering ChatGPT produced predominantly left-leaning answers across a battery of political questions, with the Post reporting the model offered right-leaning positions only once in their sample (Washington Post, June 24, 2026). The New York Post echoed similar figures, reporting an 80% share of left-leaning answers for a ChatGPT model in its coverage (New York Post, June 24, 2026). The Centre for Policy Studies report by David Rozado found that in a separate sample of 24 LLMs more than 80% of responses to policy recommendation prompts were left of centre, and measured average sentiment scores of +0.71 for left-leaning parties versus +0.15 for right-leaning parties (CPS, 2024).

Public reporting and academic work attribute political skew to a mix of training data composition and alignment-stage choices. The arXiv preprint "Perceived Political Bias in LLMs" notes training corpora often include large amounts of mainstream media and academic text that may skew left, which can influence model priors absent counterbalancing data or alignment targets (arXiv, 2026). Independent observers such as AllSides and AEI have previously documented systematic lean in some AI outputs across different evaluation protocols.

Multiple public analyses and an academic literature stream converge on the finding that LLMs can display measurable political lean. For practitioners this matters because LLM outputs are increasingly used for summarization, policy explainers, and decision support. Dartmouth Polarization Lab director Sean Westwood is quoted in the Washington Post saying, "These AI tools are not presenting a truly neutral representation of really nuanced policy debates, on average," which underscores concerns about the models influence on public understanding (Washington Post, June 24, 2026).

Observers will watch for three signals:

• •whether vendors publish reproducible, transparent evaluation datasets and methodologies for political-sensitivity testing
• •any peer reviewed replications of the CPS/David Rozado findings across languages and jurisdictions
• •whether academic or policy bodies recommend standards for measuring and reporting political balance in model outputs. Policy debates and regulator attention to "neutrality" provisions, which have already appeared in some jurisdictions, are likely to reference reproducible tests like those cited by the Washington Post and CPS