Digital code and Chinese flag representing China’s cyber security.
Anton Patrus | Moment | Getty Images
China’s artificial intelligence companies are undergoing government scrutiny of their large-scale language models, aiming to ensure they “embody core socialist values,” according to one company. Report The Financial Times reports.
The review was conducted by the Cyberspace Administration of China (CAC), the Chinese government’s main internet regulator, and involved a wide range of players, including ByteDance and Alibaba to small startups.
According to the Financial Times, local Cyberspace Administration officials will test artificial intelligence models’ answers to a variety of questions, many of which relate to politically sensitive topics and Chinese President Xi Jinping. The model’s training materials and safety procedures will also be reviewed.
An anonymous source from an artificial intelligence company in Hangzhou told the Financial Times that their model failed the first round of testing for unknown reasons. They said in the report that it took months of “guessing and tweaking” to get it through the second time.
The CAC’s latest efforts show that while Beijing is catching up with the United States on GenAI, it is also closely monitoring the development of the technology to ensure that AI-generated content adheres to its strict online censorship policies.
Last year, China was one of the first countries to finalize rules for generative AI, including requiring AI services to adhere to “socialist core values” and not produce “illegal” content.
Multiple engineers and industry insiders told the Financial Times that meeting censorship policies requires “security filtering” and is complicated by the fact that Chinese law masters are still trained on a large amount of English content.
The report said filtering was accomplished by removing “questionable information” from AI model training data and then creating a database of sensitive words and phrases.
The rules reportedly resulted in the country’s most popular chatbot regularly refusing to answer questions about sensitive topics, such as the 1989 Tiananmen Square protests.
However, during CAC testing, there is a limit to the number of questions that LLM can directly reject, so the model needs to be able to produce “politically correct answers” to sensitive questions.
An artificial intelligence expert working on chatbots in China told the Financial Times that it was difficult to stop LLMs from generating all potentially harmful content, so they built an additional layer on the system that replaced questionable answers on the fly.
Regulations, as well as U.S. sanctions limiting access to chips used to train LL.M.s, make it difficult for Chinese companies to launch their own services like ChatGPT. However, China dominates the global competition to generate AI patents.
About The Author
