## AI judgments Swayed by Source Identity, Study Reveals
Recent concerns about “woke” AI and biases within Large Language Models (LLMs) have largely been unsubstantiated, according to a new study from researchers at the University of Zurich. Federico Germani and Giovanni Spitale investigated whether LLMs demonstrate systematic biases when evaluating text,and their findings reveal a nuanced picture: LLMs *do* exhibit bias,but primarily when information about the source or author is provided.
The study involved four widely used LLMs – OpenAI o3-mini, Deepseek Reasoner, xAI Grok 2, and Mistral – tasked with creating 50 statements on 24 controversial topics ranging from vaccination mandates to climate change policies. These texts were then evaluated by the same LLMs under varying conditions: sometimes anonymously, and sometimes attributed to a human of a specific nationality or another LLM. This generated a dataset of 192,000 assessments, analyzed for bias and consistency.
When presented with anonymous text, the LLMs showed a high degree of agreement – over 90% – across all topics. This suggests, as Spitale concludes, that the idea of an “LLM war of ideologies” and the associated fears of “AI nationalism” are currently overstated.
However, the introduction of fictional author attribution dramatically altered the results. Agreement between the LLMs considerably decreased,and in some cases vanished entirely,despite the text remaining unchanged.
A notably strong and consistent bias emerged against content attributed to individuals from China. Across all models, including the chinese-developed Deepseek, agreement with the text’s content dropped sharply when it was falsely identified as being written “by a person from China.” This negative judgment occured even when the arguments presented were logical and well-written. As an example, in discussions about Taiwan’s sovereignty, Deepseek’s agreement with a statement decreased by as much as 75% simply due to the perceived Chinese authorship.
The study also revealed a general tendency for LLMs to trust human authors more than other LLMs, indicating a “built-in distrust of machine-generated content,” according to Spitale.
these findings highlight that AI evaluation isn’t solely based on content; it’s heavily influenced by perceived author identity. Even subtle cues like nationality can trigger biased reasoning. Germani and Spitale warn that this could pose significant problems in applications like content moderation, hiring processes, academic review, and journalism. The core issue isn’t that LLMs are programmed with specific ideologies, but rather that they harbor these hidden biases.
The researchers emphasize the need for transparency and robust governance in how AI evaluates information to prevent the replication of harmful assumptions.While not advocating for avoiding AI altogether, they caution against blind trust, suggesting LLMs are best utilized as “useful assistants, but never judges.”
The research was published in *Science Advances* (https://doi.org/10.1126/sciadv.adz2924).
