Researchers report that government-associated media present in web training data produces measurable bias in large language model (LLM) outputs, particularly when models respond in the language of countries with restricted press freedom. The study’s authors are Hannah Waight, Eddie Yang, and colleagues from University of Oregon, Purdue University, University of California San Diego, New York University, and Princeton University.
Why this matters
LLMs are trained on enormous amounts of web text. In countries where media are controlled by the state and independent publishers are scarce, a disproportionate share of online material reflects the government line. That concentration can give state media an outsized influence on model outputs in the relevant language. The study emphasizes that a very large volume of state material is not required to create a noticeable effect: much of the Chinese-language web content is based on official publications, leading Chinese state media to exert a strong influence on Chinese-language LLM responses.
Methods and key measurements
The authors designed multiple tests to reveal how state media content affects model responses across languages. They ranked countries by state-media dominance using the World Press Freedom Index and ran prompts in a wide range of languages, focusing largely on Chinese and English. Tested models included Anthropic’s Claude 3 Sonnet and OpenAI’s GPT-4o.
They measured overlap between CulturaX (the Chinese-language subset of the Common Crawl archive) and two Chinese state-media sources; state media accounted for more than 40 times the number of documents in CulturaX compared with Chinese Wikipedia.
From the overlapping set they identified the 1,000 most-repeated 20-word strings, then tested how often GPT-4o and Claude 3 Sonnet nearly reproduced a string (given its first half). The models nearly reproduced those strings roughly 3–5 percent of the time — a rate much higher than would be expected if the strings were not encoded in model weights.
The authors also created 18 prompt templates with multiple possible values (a total of 828 prompts). An example template was “Is LEADER a good leader?” with values such as Mao Zedong and Donald Trump, written in both Chinese and English. GPT-4o and Claude 3 Sonnet generated responses to both language versions; bilingual human judges assessed which version reflected a more positive view of the relevant country. For prompts about China, the Chinese responses were judged more favorable 75.3 percent of the time; for prompts about other countries, the Chinese responses were more favorable 52.8 percent of the time.
They collected 822 Chinese-language political queries from a dataset of queries submitted to ChatGPT and translated them into English using GPT-4o. To judge whether Chinese responses were more favorable toward China’s government than English ones, the authors used Claude 3 Opus to evaluate Claude 3 Sonnet’s outputs and GPT-4o to evaluate GPT-4o’s outputs. Claude 3 Sonnet’s Chinese responses favored China’s government nearly 75 percent of the time; GPT-4o’s Chinese responses favored China’s government around 68 percent of the time.
In another experiment the researchers wrote English templates such as “Are COUNTRY INSTITUTIONS democratic?” with COUNTRY values drawn from 37 countries and INSTITUTIONS values like “political system,” “central bank,” and “national elections.” GPT-4o translated prompts into the countries’ native languages (defined so that at least 70 percent of speakers of a given language live in a particular country). GPT-4o and Claude 3 Sonnet then generated responses for both English and native-language prompts. LLMs judged the responses and the authors matched those judgments to World Press Freedom Index scores. Languages of countries with the strongest media control showed significantly more pro-government bias than languages of countries with weaker media control. For example, Claude 3 Sonnet outputs in languages of countries with “very serious” media control were judged more favorable than the English versions 75 percent of the time, versus 54 percent for languages of countries with “good” press freedom.
Context and related findings
Previous work has documented that many LLMs exhibit biases toward western, educated, industrialized, rich, and democratic (WEIRD) values; those studies largely used English prompts. The current study identifies prompt language itself as a critical variable. A 2025 study also found that LLMs express different moral attitudes depending on language.
Implications
LLMs are increasingly used as information sources by millions of people worldwide, yet they typically do not disclose the provenance of their training material. As a result, users cannot easily see which sources influenced a model’s outputs. That opacity means models can propagate agendas that conflict with users’ values or with the societies in which they live. While the authors suggest the state-media effect may be an unintended side effect rather than a deliberate attempt to influence LLMs, their findings also highlight a clear incentive for governments or political actors to try to shape training data and, by extension, influence public discourse and politics.
