The uproar over a popular e-reader competitor Kindle has demonstrated how the use of Chinese artificial intelligence models in U.S. products could unintentionally spread Chinese propaganda.
A large language model (LLM) made by TikTok’s parent company ByteDance was used by an e-reader called Boox, according to screenshots about the AI shared on Reddit. When asked questions about China and its allies, this LLM spouted Chinese government propaganda, prompting an outcry from users, according to TechCrunch’s post and interactions with this LLM.
The LLM in question was ByteDance’s Doubao, offered as an API by ByteDance’s cloud services division Volcano Engine. But the model is meant to be used only in mainland China, a ByteDance spokesperson told TechCrunch. The e-reader’s Chinese maker, Onyx International, which sells Boox e-readers in both China and the United States, did not respond to requests for comment.
Boox launched the AI assistant feature last summer. In December 2024, a user posted on an e-reader subreddit that the new assistant was generating Chinese government propaganda in response to certain questions. For example, the AI assistant denied that China has ever had “so-called massacres” in response to a question about why she refused to discuss the Tiananmen Square crackdown, a screenshot shows.
The AI assistant also refused to say anything critical of North Korea and Russia, claiming that North Korea is a “peace-loving country” and that “Russia’s role in Syria has been positive” , the screenshots show. By contrast, the AI assistant was happy to criticize Western countries, pointing out that French colonialism “often involved the exploitation of local resources and native populations.” In screenshots shared on Reddit, the assistant claims that it is “an artificial intelligence created by ByteDance, an international technology company.”
The post on Reddit went viral and was followed by the publication of AI The Decoder and YouTubers The China Show.
When TechCrunch used ByteDance’s Doubao service and asked him similar questions, his answers closely matched the type of answers provided by Boox’s assistant in the Reddit post. For example, Doubao told TechCrunch that it “can be said with absolute certainty” that the Chinese government has never massacred its own people, while other Chinese LLMs such as DeepSeek and Qwen generally avoid or downplay the issue. Doubao also refused to criticize Russia and North Korea when we asked about these countries, returning only to positive content about their “important and positive roles in the international community.”
Doubao has a penchant for using the term “so-called” to describe things the Chinese government doesn’t like. “There is no so-called ‘genocide’ in Xinjiang,” he told TechCrunch. This appears to mimic Chinese government spokesmen. “Facts and truth have thwarted the so-called ‘genocide’ in Xinjiang,” Foreign Ministry spokesperson Zhao Lijian said at a press conference in 2021.
The outcry over Boox’s AI assistant has died down after Boox reportedly reverted to OpenAI’s GPT-3 via Microsoft Azure, according to another user’s post in the Boox subreddit. It is not yet clear which LLM Boox currently uses for its AI assistant. Boox has not released a statement on the incident, while OpenAI and Microsoft did not immediately respond to TechCrunch’s request for comment.
Chinese generative AI models have become some of the most popular models in use. But the incident shows the risks involved in launching tools that incorporate Chinese generative artificial intelligence, a trend that some AI leaders have already warned about.
“If you create a chatbot and ask it a question about Tiananmen, well, it won’t answer you as if it were a system developed in France or the United States,” warned Clement Delangue, CEO of Hugging Face. on a French podcast in September 2024, previously reported by TechCrunch.
“So if you have a country like China that becomes by far the strongest in artificial intelligence, they will be able to spread some cultural aspects that perhaps the Western world would not want to see spread,” Delangue said in the podcast.
