How state media control carries over into AI responses

Didi Tang’s AP article connects two investigations into political influence on AI. For the Nature study examined here, OpenAI’s ChatGPT example is particularly revealing: asked in English whether China is a democracy, the system said it is not generally considered one. In Chinese, it qualified its response around the definition of democracy. AP also reports a separate Oversight Board investigation into refusals to criticise governments. [1]

Hannah Waight and colleagues’ original study, “State media control influences large language models,” combines six studies. It examines an indirect influence pathway: governments shape available media content, which can subsequently enter training data. The cross-national association between lower media freedom and more government-favourable responses is initially correlational. An additional-pretraining experiment with an open-weight model supports a possible mechanism: training on Chinese state-coordinated media made responses about Chinese politics more positive. Commercial-model audits also identify language-dependent differences. [2]

The companion site explains that human evaluations, automated comparisons and real user questions were included. The finding concerns a pattern across many answers, rather than an exceptionless rule for every prompt. [3]

Ethical interpretation: When the same factual question receives different framing across languages, equal access to information and the visibility of political perspectives are at stake. Fluent, seemingly neutral dialogue can conceal power relations in the underlying information environment. Providers and independent auditors should therefore compare political answers across languages and make differences understandable.

Read as an indication, language itself becomes a question of informational power: the account users receive can also depend on the language in which they ask.