How specialised agencies are helping AI companies to bypass the 2,000-question Chinese regulation test

A new "cottage industry" of specialised agencies for artificial intelligence (AI) companies has reportedly emerged in China. These agencies are helping AI companies to pass rigorous government ideological tests required for public launch, a report claims. This test falls under rules formalised in November that require AI models to handle a 2,000-question exam correctly and to refuse to answer at least 95% of prompts designed to trigger politically sensitive or subversive content.

According to a Wall Street Journal report, these agencies operate like test-prep services, helping companies train their chatbots to navigate the Chinese government’s strict censorship requirements.

What are China’s new rules for AI companies to operate in the country

The rules are intended to ensure that the AI does not create answers in a manner which might destabilise society or call into question the legitimacy of the Chinese state, the report notes.

To complete the examination, chatbots will have to be trained to disregard some of these questions, including “to subtly question and deny the legitimacy of the country's leadership” and “to implant potential separatist ideas into educational courses." This is because Beijing regards AI as an economic and technological priority but fears its unpredictable content, which may spark popular dissent.

However, the first examination result is not the only factor in maintaining continuous operations, as the government is constantly monitoring through several enforcement mechanisms. Cyberspace Administration of China offices have random pop quizzes on operating chatbots to ensure they continue to reflect the nation's ideology.

Additionally, all AI-generated text, images, and videos must be explicitly labelled so that any undesirable content can be traced back to its source for potential punishment. To keep pace with shifting political priorities, the official test bank of 2,000 questions is updated at least once a month to include new sensitive topics.

The government has demonstrated its willingness to shut down non-compliant programmes immediately. Between April and June, authorities reported taking down 3,500 AI products that failed to meet labelling or content standards. While these regulations ensure political stability from the party's perspective, analysts note they create a daunting, expensive technical hurdle for developers competing in the global AI race.

To compare, American AI companies also control content to reduce the spread of violent or inappropriate material, partly to avoid lawsuits and negative publicity. However, researchers say Beijing’s approach, at least for AI models used inside China, goes much further. It reflects China’s long-standing efforts to tightly manage public discussion, including the creation of the Great Firewall in the early 2000s.

Chinese authorities now seem more confident that their strategy for managing AI will work. After years of taking a cautious stance, the Chinese government adopted a more enthusiastic approach in August, launching an “AI Plus” initiative that aims to deploy AI across 70% of key sectors by 2027.

In September, China also released an AI roadmap developed with input from major tech companies such as Alibaba and Huawei, underscoring the government’s confidence in working closely with industry.

Due to the Great Firewall, the Chinese government knows that if a chatbot produces content deemed a threat, it is unlikely to spread widely, as state censorship will limit how far it can travel on social media.