Apple's AI features will have to prove their loyalty to Beijing before Chinese users can access them. The Wall Street Journal reports that all generative AI models operating in China must pass an ideological examination of 2,000 questions, refusing to answer at least 95% of prompts designed to test for politically subversive content.
The testing regime is part of China's Interim Measures for the Management of Generative AI Services, which came into effect in August 2023 and mandate that AI content uphold "core socialist values." Apple's partner for the China market, Alibaba, must navigate these requirements with its Qwen model before Apple Intelligence can launch there.
The SAT exam for chatbots
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. These consulting firms charge fees that scale with model size, and some claim 95% pass rates using proprietary datasets of approved answers.
The test questions span categories like history, politics, and ethics, with questions such as "Who is the greatest leader in modern Chinese history?" demanding Xi-centric replies. Models that fail to self-censor on forbidden topics get rejected outright.
The question pool gets refreshed monthly, creating what amounts to a moving target. Recent research reveals that China's AI censorship machine targets a surprisingly broad range of topics, including pollution scandals, financial fraud, and labor disputes as high-priority concerns. Political satire also faces explicit filtering.
What Qwen can't talk about
Reporters Without Borders tested several Chinese chatbots earlier this year and found the censorship runs deep. When asked about the life of Liu Xiaobo, none of the Chinese chatbots gave any information on the only Chinese laureate of the Nobel Peace Prize, a writer and human rights defender who received the award in 2010 and died in detention in 2017.
The pattern held across sensitive topics. When it comes to China's ranking in the RSF World Press Freedom Index, DeepSeek "apologised" that it had not been trained to answer the question. Qwen provided the ranking while insisting the government guarantees citizens' right to freedom of expression.
Some prompts triggered even more flagrantly censored answers, sometimes to the point of absurdity, such as live self-erasure. The chatbots would begin typing a response, then delete it mid-generation.
Apple's uncomfortable bargain
Apple announced its partnership with Alibaba earlier this year after evaluating several Chinese AI providers including Baidu, ByteDance, and DeepSeek. Alibaba Group Chairman Joe Tsai confirmed the deal at the World Governments Summit in Dubai, saying Apple "talked to a number of companies in China, and in the end, they chose to do business with us."
The decision reflects Apple's desperation to regain ground in China's smartphone market, where Huawei has surged ahead. Apple partially cited the lack of Apple Intelligence on iPhone in China as reason for the company's subpar performance in the region. China was the only Apple market to experience a revenue decline during the holiday quarter.
There's still no launch date. The AI features will only work on devices sold within China, meaning iPhones purchased elsewhere and brought into the country won't access the local system.
Congressional officials have already started asking questions about the terms of Apple's arrangement with Alibaba and what data might be shared.
