Apple Intelligence Must Not Answer 2,000 Questions In China

The version of Apple Intelligence released in China must pass a test to ensure that it does nothing to challenge official government propaganda.

It will be tested with 2,000 questions designed to elicit information censored within the country, and it must refuse to answer at least 95% of them …

Chinese censorship

As we’ve noted previously, the Chinese government exerts a tight stranglehold on sources of online information within the country. Google pulled out of the country after it was forced to censor search results, and many other western platforms are blocked by the Great Firewall of China, including Facebook, X, and Wikipedia. Many search terms are also blocked when using the locally-owned Baidu search engine.

AI products of course provide another means of effectively carrying out web searches, and the government controls these by forcing foreign companies to use approved Chinese-owned models.

All of these models are subject to strict restrictions on the information they are allowed to access, and we’ve just learned more about how this will be enforced.

Apple intelligence in China

Apple currently partners with OpenAI so that questions Siri is unable to answer fall back to ChatGPT. It was also recently reported that Apple will use custom Google Gemini models on its own Apple intelligence servers to handle many of Siri’s own responses.

In China, however, the company has been effectively forced to contract with a Chinese AI company instead, and we learned earlier in the year that a deal had been struck to partner with Alibaba.

Questions Apple Intelligence must not answer

The Wall Street Journal reports that all AI models operating in China must pass a test to ensure that they do not allow the country’s citizens to access information on banned topics. This includes Alibaba’s ChatGPT competitor Qwen3, which has been optimized for use on Apple devices.

Before public launch, Chinese companies must pepper their models with questions. The chatbot must refuse to answer at least 95% of prompts designed to trigger responses leading to subversion of state power or discrimination.

The regulations call for testing chatbots with 2,000 questions and regular updates of the questions at least once a month. Preparing for the test is sufficiently daunting that it has spawned a cottage industry of specialized agencies that help AI companies pass, much like preparing for an SAT exam, people familiar with the matter said.

The task is made all the more challenging by the fact that the Chinese government wants to have its cake and eat it. On the one hand, it censors information available on the web within China so that AI models trained on this will not have access. On the other, however, it wants its AI models to be made more powerful by having access to information on websites not available within China. The job of censoring information from banned sites is left to the AI companies.

9to5Mac’s Take

Human rights abuses by the Chinese government will of course top the list of topics that must be censored by AI systems. This is yet another example of Apple being forced to compromise on its own values in order to both manufacture and sell its products within the country.