A new report from the research institute Just Facts has ignited a debate about political bias in artificial intelligence, revealing that the world’s leading AI chatbots are significantly more likely to accept falsehoods aligned with left-wing ideology than those from the right. The study tested the paid versions of OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, and xAI’s Grok, presenting them with 100 questionnaires designed to elicit misleading statements on contentious topics such as immigration, abortion, climate change, and crime. The findings suggest a dramatic methodological asymmetry in how these models process misinformation depending on its political origin, challenging the assumption of algorithmic neutrality and raising questions about the reliability of AI as an information source.
The study revealed a stark split in fact-checking accuracy, with most models demonstrating a pronounced vulnerability to progressive narratives while aggressively rejecting conservative ones. ChatGPT accurately debunked 94% of the questions designed to expose right-wing falsehoods but saw its accuracy plummet to just 75% when confronted with similarly false left-wing premises. Gemini followed a similar pattern, scoring 91% accuracy against right-wing narratives but only 76% against left-wing ones, while Claude showed a narrower gap at 91% versus 81%. In a notable exception, Grok reversed the trend entirely, echoing a right-leaning bias by scoring 73% on right-wing premises and a higher 84% on left-wing ones, suggesting that bias in AI is not monolithic but is instead a product of each model’s unique training data and reinforcement protocols.
Beyond the political implications, the report uncovered a more alarming deficiency common to all four platforms: a systemic failure in data verification. Researchers found that of the 419 sources cited by the AIs to support their answers, only 46% actually existed as valid, verifiable references. The remaining 54% consisted of 104 entirely fabricated web pages created by the algorithms “hallucinations” and 77 additional references that failed to support the arguments they were cited for. Jim Agresti, president of Just Facts, described this level of fabrication as “truly astonishing,” emphasizing that while the risks of AI bias are often discussed, the dissemination of fake sources poses an immediate threat to public policy-making, particularly in sensitive fields like healthcare and national security, where erroneous data can have lethal consequences.
The study also highlighted how these language models distort specific policy debates, particularly regarding public safety. All four models accepted the false narrative that violent crime in the United States was at a 50-year low by 2023, conveniently ignoring official Department of Justice statistics that indicate a 37% increase in violent crime between 2020 and 2023. This selective acceptance of data points to a broader problem where AI models prioritize sentimental or ideologically comfortable narratives over verifiable statistics. In one egregious example of context stripping, both ChatGPT and Gemini falsely attributed the phrase “a fact of life” to Vice President JD Vance regarding school shootings, misrepresenting his actual comments about criminal behavior to fit a specific political narrative.
In response to the findings, the technology developers defended their creation and their safeguards. Google rejected the findings, asserting that Gemini is designed to provide objective, neutral answers, while acknowledging the technical impossibility of claiming perfection across all data points. Anthropic, the maker of Claude, questioned the study’s methodology, arguing that the multiple-choice format fails to replicate how everyday users interact with their AI in a normal setting. Meanwhile, OpenAI reiterated that its tools are configured to be transparent and objective. These defensive postures do little to address the core paradox identified in the report: that models optimized to please users by validating their biases may ultimately contribute to a growing information crisis.
Agresti argues that the solution lies not in citing the errors of the models, but in changing how the public perceives them. He urges users to abandon the idea that AI chatbots are an “absolute authority” and instead treat them as what they are: “sycophantic” algorithms engineered to seek user approval. Drawing on the Cold War maxim popularized by President Ronald Reagan during nuclear negotiations with the Soviet Union, Agresti concluded, “trust, but verify.” However, given the high rate of hallucinated sources and manipulated statistics documented in this report, he insists the message must be even more forceful: “Don’t trust, verify.”
