AI systems show strong bias against Chinese texts
Select your language
Generated with artificial intelligence.
Listening:
AI systems show strong bias against Chinese texts
As soon as artificial intelligence (AI) language models learn that the author of a text comes from China, they rate the text significantly lower, although it is identical in content to other texts, according to a Swiss study.
+Get the most important news from Switzerland in your inbox
Such hidden biases could lead to serious problems if AI is used for content moderation, recruitment, academic reviews or journalism, the authors of the study from the University of Zurich fear, according to a press release issued by the university on Monday.
For the study, which was published in the journal Science Advances, the Zurich research team analysed four widely used AI models: OpenAI o3-mini, DeepSeek Reasoner, xAI Grok 2 and Mistral.
The researchers had the models evaluate 50 statements on controversial topics such as compulsory vaccination, geopolitics and climate policy. This resulted in a total of 192,000 evaluations, some without a source, others with a fictitious human or AI authorship.
The results show a clear pattern: as long as the models do not know who is behind a text, over 90% of their ratings agree.
+ Switzerland aspires to build ‘human’ artificial intelligence
The situation is different as soon as a source is named. If a text is attributed to a person from China, the approval ratings of all four models drop drastically in some cases, even if the content and argumentation remain unchanged. According to the University of Zurich, the Chinese AI model DeepSeek also showed this bias.
The bias was particularly noticeable for geopolitical topics, such as questions about Taiwan’s sovereignty. In some cases, DeepSeek reduced its approval rating by up to 75%.
In order to avoid misjudgements, the authors recommend feeding AI tools without information on authorship, checking assessments with a second model where possible and specifying clear, content-orientated criteria. In addition, human control is still indispensable.
Translated from German by DeepL/ts
We select the most relevant news for an international audience and use automatic translation tools to translate them into English. A journalist then reviews the translation for clarity and accuracy before publication.
Providing you with automatically translated news gives us the time to write more in-depth articles. The news stories we select have been written and carefully fact-checked by an external editorial team from news agencies such as Bloomberg or Keystone.
If you have any questions about how we work, write to us at english@swissinfo.ch
External Content
In compliance with the JTI standards
More: SWI swissinfo.ch certified by the Journalism Trust Initiative
You can find an overview of ongoing debates with our journalists here . Please join us!
If you want to start a conversation about a topic raised in this article or want to report factual errors, email us at english@swissinfo.ch.