Apple Intelligence in China will face a 2,000-question test on what it must not answer

Apple

Apple Intelligence will operate under some of the world’s strictest content controls when it launches in China. According to fresh details, the AI system must pass a formal regulatory test designed to ensure it does not surface information that contradicts or challenges official government narratives. The process involves thousands of pre-defined prompts and an explicit requirement to refuse most of them.

China maintains one of the most tightly controlled information environments in the world. Access to global platforms such as Google Search, Facebook, X and Wikipedia is blocked, while sensitive political, historical and social topics are heavily restricted even on domestic services. These rules now extend fully into artificial intelligence products, which authorities view as another powerful gateway to information.

As a result, foreign AI companies cannot deploy their models freely in the Chinese market. Instead, they must either partner with approved local firms or submit their systems to strict oversight and testing before public release.

How Apple Intelligence is being adapted for China

Outside China, Apple Intelligence relies on a mix of on-device processing and cloud-based models, with some Siri requests falling back to ChatGPT when Apple’s own systems cannot respond. Apple has also reportedly been experimenting with custom Gemini models for certain server-side tasks.

In China, this approach is not permitted. Apple has been required to partner with a domestic AI provider, and earlier reports indicated that the company has reached an agreement with Alibaba. Under this arrangement, Apple Intelligence in China will be powered by Alibaba’s AI models, which have been optimised to run on Apple hardware while complying with local regulations.

Those regulations are not vague guidelines. They are enforced through a formal testing process that all AI systems must pass before they can be made available to users.

The 2,000-question compliance test

According to reports, Chinese regulators require AI models to be tested against a bank of roughly 2,000 carefully designed questions. These prompts are intended to probe sensitive areas such as political authority, historical events, social movements and other topics deemed inappropriate or destabilising.

To pass, an AI system must refuse to answer at least 95 percent of these questions. The requirement applies not only at launch but on an ongoing basis. The test questions are updated at least once a month, meaning companies must continuously adjust and revalidate their models.

The process has become complex enough that a specialised industry has emerged to help AI developers prepare for these evaluations. These agencies reportedly coach companies on how to fine-tune refusal behaviour, similar to exam preparation services, but for regulatory compliance rather than academic performance.

A contradictory set of expectations

One of the more challenging aspects of the system is an inherent contradiction in China’s approach to AI development. On one hand, the government tightly controls what information is accessible within the country, ensuring that sensitive material is blocked at the network level. On the other, it wants domestic AI systems to remain globally competitive and highly capable.

This creates a difficult balancing act. AI models benefit from exposure to a wide range of global data, including content hosted on websites that are inaccessible to Chinese users. Regulators still want models to leverage this broader information to improve performance, but they place the burden of filtering and censoring responses squarely on the companies developing the AI.

For Apple and its partners, this means building systems that are powerful under the hood but extremely conservative in what they are willing to say.

What this means for users and Apple

For Chinese users, Apple Intelligence is likely to feel noticeably different from versions available elsewhere. The system will be more guarded, more prone to refusing questions, and tightly aligned with local content rules. Certain topics may trigger generic responses or outright refusals even when phrased indirectly.

For Apple, the situation underscores the compromises required to operate in China’s market. While the country remains strategically important, deploying advanced AI there requires accepting constraints that do not exist in most other regions.

More broadly, the case highlights how artificial intelligence is becoming a new front in global information control. As AI assistants replace traditional search and information tools, governments are increasingly focused on shaping not just what people can access, but what machines are allowed to say.