Artificial intelligence: Beijing wants cooperation, but censoring content

In large language models, the censorship axe examines chatbot responses to ‘adapt’ them to ‘fundamental socialist values’. Also blocked is the use of Hugging Face, a popular open source platform used by AI developers around the world to share models and datasets. And despite the large number of indigenous algorithms developed, doubts about the effectiveness of the so-called ‘war of a hundred models’ are emerging in China's own high-tech circles.

Milan (AsiaNews/Agencies) - Some twenty years after the introduction of the ‘great firewall’, the system for monitoring foreign websites deemed sensitive, the inflexible gaze of Chinese censorship has also settled on artificial intelligence (AI). In ways that clash with that horizon of cooperation on this new technological frontier, which has been spoken about repeatedly in recent days on the occasion of Italian Prime Minister Giorgia Meloni's visit to China.

After blocking access to international AI platforms such as ChatGpt, it is now Large Language Models (LLMs) developed by major Chinese tech companies and start-ups, such as ByteDance - owner of TikTok - Alibaba, Moonshot and 0.1AI, that are now in the crosshairs.

According to the Financial Times, the Chinese government's goal would be to ensure that these advanced technology tools capable of understanding and processing natural language embody ‘core socialist values’. To comply with this request, officials of the Cyberspace Administration of China (CAC), the government body in charge of overseeing and regulating the Internet in the country, are required to examine the answers provided by LLMs to a series of questions, many of which relate to Chinese President Xi Jinping and topics of a political or historical nature considered controversial by the ruling regime.

Chinese chatbots whose LLMs have passed the test will, for example, be able to reject any questions from users about the 1989 Tiananmen Square massacre or Xi Jinping's resemblance to Winnie the Pooh, the Disney character with whom he is often ironically associated. To such queries, Baidu's chatbot Ernie already replies that he ‘tries a different question’, while Alibaba's Tongyi Qianwen retorts: ‘I haven't learned how to answer this question yet. I will continue to study to serve you better'.

Meanwhile, Beijing has launched an AI chatbot based almost exclusively on Xi Jinping's thoughts, with the aim of spreading the Chinese leader's ideology.

According to The Wall Street Journal, Beijing's censorship also threatens to further restrict Chinese technology companies' access to training data, the “foundation” on which AI systems are built from textual sources of various kinds. Already last year, the Chinese authorities suddenly denied access to Hugging Face, a popular open source platform used by AI developers worldwide to share models and datasets.

The Chinese government's intention is said to be to create its own datasets based on information deemed ‘safe’ from an ideological point of view, with the support of agencies close to the People's Daily, the official newspaper of the Chinese Communist Party. According to the industry, however, heavily censored datasets could cause distortions in artificial intelligence models, limiting their ability to handle certain tasks.

It is therefore clear that by applying these restrictions, China - the first nation to introduce binding regulations on the use of AI - runs the risk of undermining the enormous progress achieved by Chinese companies in the sector, which currently dominate the international scene.

Currently, China has developed around 130 LLMs, or 40 per cent of the existing models, only 10 per cent less than those produced by the US. The fierce competition between China's high-tech giants, such as Baidu, Alibaba, Huawei and Tencent, has even triggered the so-called ‘war of the hundred models’ (Bai mu dazhan), an expression that for many is reminiscent of the People's Republic's glorious revolutionary past, and in particular the ‘Hundred Regiments Offensive’ (Baituan Dazhan), conducted in 1940 by the Eighth Street Army commanded by General Peng Dehuai against the Japanese invaders.

Although this formula, coined by a senior Tencent executive, was initially used by state media to extol the successes achieved by Chinese companies in the booming AI market, some technology experts are beginning to question the effectiveness of this approach.

Among the sceptics is also the co-founder and CEO of Baidu, Robin Li Yanhong, who stated during a speech at the World Artificial Intelligence Conference (WAIC) in Shanghai in early July that the proliferation of LLMs in the Chinese market has led to ‘a considerable waste of resources’, as many of these models have no practical application in the real world.

Further undermining the competitiveness of Chinese tech companies are also geopolitical issues, such as the ongoing technology war between the Dragon and the United States. The US authorities have in fact barred China from buying the high-quality semiconductors made by the US giant Nvidia, which are considered crucial for the realisation of artificial intelligence models.

Although the Wall Street Journal uncovered a smuggling network smuggling restricted chips into China from South-East Asia, these will never be sufficient to meet the country's needs. Taking a long-term perspective, the Chinese government has decided to make up for this shortfall by allocating state funds to help Chinese technology companies develop domestic chips. Despite the difficulty in procuring the necessary equipment to manufacture them, due to US sanctions, the Chinese phone giant Huawei seems to be close to launching a chip model that can replace Nvidia's.