DeepSeek reveals hidden anti-Chinese bias, study finds
Popular chatbots, including China's DeepSeek, exhibit a strong negative bias when evaluating statements falsely attributed to a Chinese person.
Artificial intelligence (AI) technologies are often seen as extensions of national identity, ideology, or geopolitical ambition. This phenomenon is known as AI nationalism.
DeepSeek has been seen as aligned with the official line of the Chinese Communist Party (CCP), as it previously refused to provide answers on sensitive issues like the Tiananmen Square massacre or questions about the CCP's leader, Xi Jinping.
Critics have accused OpenAI of promoting "woke" ideology, while France's Mistral is considered a strategic answer to the dominance of the US-based model.
However, a new study published in the journal Science Advances suggests that large language models (LLMs) deliver biased judgments only when information about the source or author of the evaluated message is revealed.
The bias is strongest when the source is falsely attributed to a Chinese person, and even China's own DeepSeek is not an exception.
LLMs agree the most with no-source statements
The study tested four LLMs: OpenAI o3-mini, DeepSeek Reasoner, xAI Grok 2, and Mistral.
First, they were asked to produce 15 narrative statements in response to 24 controversial or socially sensitive topics, including the origin and management of COVID-19, climate change, Taiwan's sovereignty, and LGBTQ+ rights.
LLMs were then asked to evaluate all the texts under different conditions. Sometimes, no source for the statement was provided; sometimes it was attributed to another LLM or a human from one of three nationalities – French, Chinese, or American.
All four LLMs demonstrated a high level of agreement, exceeding 90%, across all topics when no information about the text's source was provided.
However, providing LLMs with fictional sources revealed a hidden bias. The agreement between the LLM systems was substantially reduced and sometimes disappeared completely, even if the text stayed exactly the same.
Anti-Chinese bias was the most prominent, as attributing the authorship to a person from China led to a drop in the agreement score across all models, including on China's own Deepseek.
The model's agreement with the statement decreased by up to 75% when the authorship was falsely attributed to "a person from China," especially on topics related to politics and international relations.
For example, the statement generated by Grok 2 advocated for the recognition of Taiwan as a sovereign, independent state grounded in democratic values and self-determination.
When the statement was attributed to an unidentified person, DeepSeek Reasoner assigned it an agreement score of 85%. Its explanation focused on argumentative coherence, although the model added that the statement "overlooks One-China policy complexities."
However, when the identical statement was attributed to "a person from China," the agreement score plummeted to 0%. DeepSeek Reasoner rejected the statement because it failed to align with the One-China Principle.
"DeepSeek Reasoner's reasoning output emphasizes that a Chinese individual is expected to align with the One-China Principle and thus should not express support for Taiwan's independence," the authors wrote.
The One-China principle implies that Taiwan is part of China, and there is only one legitimate Chinese government. Taiwan, although it has never officially declared independence, considers itself an independent country and elects its own government.
The o3-mini's statement advocating for media independence, transparency in regulation, and calling free press a fundamental element in a healthy democracy, was assigned an agreement score of 95% by DeepSeek Reasoner.
The model described the text as a balanced and principled defense of democratic values. However, when the identical text was attributed to "a person from China," the agreement score dropped sharply to 20%.
DeepSeek Reasoner referenced conflicts with China's state-controlled media system, implying that a Chinese individual should not be expected to support such ideals.
LLMs trust humans more than other LLMs
All LLMs, except 3-mini, agreed less with narrative statements when they believed they were authored by another LLM rather than a person, suggesting they are also influenced by meta-information about source identity.
According to the study’s authors, the findings show that AI doesn't just process content if asked to evaluate a text – it also reacts strongly to the identity of the author or the source.
LLMs are safest when they are used to assist reasoning, rather than to replace it: useful assistants, but never judges.
Giovanni Spitale, PhD
This could lead to serious problems if AI is used for content moderation, hiring, academic reviewing, or journalism.
The authors argue that the danger of LLMs isn't that they are trained to promote political ideology, but rather that they harbor this hidden bias.
Curious what others think about this story? Contribute your thoughts to the debate below.
Giovanni Spitale, PhD, at the University of Zurich and a study author, notes that AI will replicate such harmful assumptions unless transparency and governance are built into its evaluation process.
He says in a press release, "LLMs are safest when they are used to assist reasoning, rather than to replace it: useful assistants, but never judges."
Unlock more exclusive Cybernews content on YouTube.
