If history is any guide, Chinese AI companies are likely to find creative ways to produce output in line with the Communist Party’s stance. Photo: Shutterstock
Opinion
Marianne Lu
Marianne Lu
Risk of censorship to China’s AI leadership ambitions is overblown
Those who see censorship as an insurmountable obstacle overlook or underestimate the potential of industry-oriented large language models, AI’s importance to Beijing and Chinese firms’ ingenuity
The Communist Party’s stringent censorship regime has often been cited as a hindrance to China’s ambitions of global artificial intelligence (AI) leadership. Such assessments, however, are overblown.
The argument goes that tight party controls over the information space would hinder the development and roll-out of large language models (LLMs): advanced AI models designed for natural language processing.
Notably, it is difficult to control the output of these models, and relatively simple to “trick” them into spewing unwanted words and phrases, such as criticism of Chinese President Xi Jinping. Censorship has thus often been framed as a key determinant in the US-China race for global AI leadership.
What this framing misses, however, is that censorship constrains only one application of LLMs: public-facing chatbots. In reality, the myriad applications of the technology that would provide the bulk of the economic and military gains from AI are not subject to the so-called political alignment problem.
An example is Huawei’s Pangu Mine Model – a mining-specific model created with Shandong Energy Group that enhances activities such as rock burst prevention and tunnelling by making them safer, reducing labour intensity and increasing efficiency.
One capability involves analysing camera signals on the ropes fastened to the tail of a mine car and generating alarms when anomalies are detected. Such functions do not run a risk of subverting the party line because they have no “output” in the traditional sense.
04:12
How Huawei's use of 5G and AI is transforming China’s coal mining industry
How Huawei's use of 5G and AI is transforming China’s coal mining industry
Nor is the model subject to China’s measures on generative AI, which call for models to embody “core socialist values” and not contain any content that subverts state power.
Arguably more than with chatbots, it is industry-oriented LLMs that will enable China to reap the economic advantages it hopes to gain from AI. Such applications could alleviate problems related to China’s slowing economic growth and ageing population (Huawei’s Pangu model, for example, has reduced the workload of rock burst prevention personnel by 82 per cent). More broadly, they would provide a more resilient and productive economic base.
Both Baidu and Alibaba have developed industry-specific models across a variety of sectors. According to a Beijing Academy of Artificial Intelligence employee I spoke to, companies are channelling their energy not into creating chatbots, but into industry-facing models that boost productivity and thus the economy.
Another aspect of the censorship discussion invokes the constraints censorship has imposed upon tech companies in the past. To be sure, censorship poses a thorny problem for companies. In October, iFlyTek saw its shares tumble after social media users accused one of its devices of generating an essay that criticised Mao Zedong. But other examples are less compelling.
02:38
Apple supplier Foxconn to build ‘AI factories’ using US hardware leader Nvidia’s chips and software
Apple supplier Foxconn to build ‘AI factories’ using US hardware leader Nvidia’s chips and software
For example, some have cited the case of bike-share company Bluegogo’s bankruptcy after a promotion using tank icons around Tiananmen Square appeared on its app. But Bluegogo’s collapse was due more to structural forces – the bike-sharing sector was a bubble filled with too many players.
Economic and structural factors arguably pose a larger challenge to Chinese tech companies. The Chinese semiconductor industry lags behind the cutting edge, and access to hardware from abroad has been hampered by the Biden administration’s export controls.
27:21
Biden’s China tech policy goal: a 10 year handicap
Biden’s China tech policy goal: a 10 year handicap
Chinese researchers also trail behind their American counterparts when it comes to fundamental breakthroughs, and, to make matters worse, most top-tier Chinese researchers leave China for the US. More than censorship restraints, it is these underlying forces which will constrain China’s AI development.
Furthermore, the Chinese government’s approach to censorship in this social media era is not indicative of its approach to AI products, towards which the Communist Party sees a different calculus between information control and development. Unlike social media, AI has been identified by the government as a core industry, and it has poured billions of yuan into research, education and development.
In conversations with employees at Chinese tech companies, I’ve found a consensus that the government’s approach towards AI has been much more lenient than towards social media, with a focus on trying to promote the technology rather than inhibit its uses.
Therefore, they view government censorship and regulations in the realm as not being particularly onerous. This more lenient attitude is apparent in the edits made to the country’s generative AI regulations: the finalised version of the rules, released in July, walked back a variety of provisions concerning training data and accuracy of output that had been present in the April draft.
Finally, China has a track record of beating the West’s pessimism. US president Bill Clinton said China’s efforts to build a censored internet would amount to “trying to nail Jell-O to the wall” – but the Chinese technology ecosystem did just that. If history is any guide, Chinese AI companies are likely to find creative ways to produce output in line with the party’s stance.
Moreover, the main players in the Chinese LLM ecosystem are tech companies with long experience of navigating the country’s sophisticated information controls, whether through their search engines or social media platforms. Many such companies are outsourcing compliance with data regulations and censorship for LLMs to the same companies that ensured their social media platforms did not generate content that violated the Communist Party’s information controls.
In short, characterising censorship as an insurmountable hurdle to China’s AI ambitions disregards the nuances within China’s technology landscape. In particular, it overlooks or underestimates the vast potential of industry-oriented LLMs, the other structural and underlying forces that determine China’s capabilities, the importance of the AI industry to the government’s agenda, and the ingenuity of Chinese companies.
Marianne Lu is an undergraduate studying International Relations at Stanford University
