A popular Kindle competitor became the center of an AI controversy after users found that its assistant appeared to give politically loaded answers about China and its allies. The assistant, available on Boox e-readers, was linked through screenshots to Doubao, a large language model made by ByteDance.
The episode matters because it shows how generative AI can quietly move from cloud services into everyday devices. When that happens, the choice of model is not just a technical decision. It can shape what users hear about history, governments and current geopolitical disputes.
What happened with Boox’s AI assistant
Boox, an e-reader made by China-based Onyx International, launched its AI assistant feature last summer. The company sells Boox e-readers in both China and to the U.S., making the assistant part of a product used beyond the market where the underlying model was meant to operate.
In December 2024, a user posted on a subreddit for e-readers that the assistant was producing Chinese government propaganda in response to some questions. Screenshots shared on Reddit said the assistant identified itself as “an AI created by ByteDance, an international technology company.”
According to TechCrunch, the model in question was ByteDance’s Doubao. Doubao is offered as an API through ByteDance’s cloud services division Volcano Engine. A ByteDance spokesperson told TechCrunch that the model is only meant to be used within China’s mainland.
Onyx International did not respond to requests for comment. That leaves a central question unresolved: how a model intended for China’s mainland ended up powering an assistant in an e-reader sold to U.S. users.
The answers that triggered the outcry
The controversy centered on how the assistant handled questions about China, North Korea and Russia. In one screenshot, the assistant denied China ever having any “so-called massacres” when asked why it would not discuss the Tiananmen Square crackdown.
The assistant also refused to offer criticism of North Korea and Russia, according to the screenshots. It described North Korea as a “peace-loving country” and said that “Russia’s role in Syria has been positive.”
By contrast, the same assistant was willing to criticize Western countries. One screenshot showed it saying French colonialism “often involved exploitation of local resources and native populations.”
That contrast was central to the backlash. The issue was not only that the assistant declined to answer certain questions. It was that the limits appeared selective, with some governments treated gently and others discussed critically.
TechCrunch found similar behavior in Doubao
TechCrunch tested ByteDance’s Doubao service with similar questions and reported that its answers closely matched the responses in the Reddit screenshots. When asked about whether the Chinese government had massacred its own people, Doubao said “it can be stated with absolute certainty” that it had not.
TechCrunch also said Doubao refused to criticize Russia and North Korea. Instead, it returned positive statements about their “important and positive roles in the international community.”
The model also used language that TechCrunch connected to Chinese government messaging. Doubao told TechCrunch, “There is no so-called ‘genocide’ in Xinjiang.” The article noted that this wording resembles language used by Chinese government spokespeople, including foreign ministry spokesperson Zhao Lijian, who said at a press conference in 2021: “Facts and truth have busted the so-called ‘genocide’ in Xinjiang.”
Other Chinese LLMs mentioned in the source, including DeepSeek and Qwen, typically avoid or downplay the same question, according to TechCrunch. Doubao’s answers, by comparison, were described as more direct in denying allegations the Chinese government rejects.
Why model choice matters for consumer AI
The Boox incident is a useful case study in a larger problem: AI assistants can carry the assumptions, restrictions and political boundaries of the systems that power them. Users may experience the assistant as a feature of an e-reader, but the content comes from a model with its own behavior.
That becomes especially important when the device is sold across borders. A model designed for one political and regulatory environment may give answers that surprise users in another. The risk is greater when the company selling the device does not clearly explain which model is being used.
The source article also places the incident in the broader rise of Chinese generative AI models. These models have become some of the most popular models in use, and that popularity makes sourcing decisions more consequential for product makers.
Clement Delangue, the CEO of Hugging Face, warned about this issue on a French podcast in September 2024. As TechCrunch previously reported, he said, “If you create a chatbot and ask it a question about Tiananmen, well, it’s not going to respond to you the same way as if it was a system developed in France or the U.S.”
Delangue added: “So if you have a country like China that becomes by far the strongest on AI, in fact they will be capable of spreading certain cultural aspects that perhaps the Western world wouldn’t want to see spread.”
What remains unclear
The outcry appears to have eased after Boox reportedly switched back to OpenAI’s GPT-3 through Microsoft Azure, according to another user’s post in the Boox subreddit. However, TechCrunch said it is still unclear which LLM Boox currently uses for its AI assistant.
Boox has not released a statement about the incident. OpenAI and Microsoft did not immediately respond to TechCrunch’s request for comment.
For users, the lesson is straightforward: an AI assistant’s answers depend heavily on the model behind it. For companies building devices with AI features, the Boox backlash shows why model selection, disclosure and testing are now part of product trust.