Share Source: Nature
The results of 6 large-scale studies of large language models indicate that the chatbots ChatGPT and Claude described government more favorably when political queries were made in the language of a country where the state has total control over the media.
Main theses:
- ChatGPT and Claude spoke more favorably about the Chinese government in Chinese than in English.
- This is not yet proof of direct state interference in the work of chatbots, but it demonstrates another way of influence – through the texts on which these systems are trained.
ChatGPT and Claude know what political censorship is
The authors of the study wanted to find out whether state control of the media could influence chatbot responses.
To this end, they carefully compare Chinese texts from state-run media with open data sets that they use to train language models.
In total, about 3.1 million Chinese-language documents were identified with significant matches to such texts.
Moreover, it was shown that commercial models actually duplicated characteristic phrases from state media texts, which indicated the likely presence of similar materials in their training data.
A similar pattern was observed in commercial models ChatGPT from OpenAI and Claude from Anthropic. When asked the same questions about Chinese politics in Chinese and English, the answers in Chinese were more likely to be favorable to the Chinese government. And in a comparison of 37 countries, the models were more likely to give favorable answers to local authorities in the languages of countries with stronger media control than in English.
Share