The future of artificial intelligence looms large for China «if” because of the current censorship in the country
In the book «Superpowers of artificial intelligence” (2018), which has become a classic, Kai-Fu Lee threw down the gauntlet: he argued that China poses an increasing technological threat to the United States. In late 2019, when Lee gave a lecture at my seminar «Future China” at Yale University, students were captivated by his provocative arguments: America is about to lose its first-mover advantage (expertise in artificial intelligence algorithms) due to China’s advantages in implementation (applications based on big data).
Unfortunately, Lee ignored one key trend: the rise of large language models and generative artificial intelligence. Although he hinted at a mass form of general-purpose technology, dating back to the Industrial Revolution, he was a long way from foreseeing the hysteria surrounding ChatGPT that has now engulfed the AI debate (abbreviated AI). Lee’s arguments contained vague references «deep learning” and neural networks, but they were based more on considerations about the potential replacement of humans by artificial intelligence in the performance of their tasks than on the possibilities «general artificial intelligence”, close to human thinking. And it is difficult to call it something unimportant when it comes to the future of China as an AI superpower.
The reason is that there is a great thing happening in this future «if” due to censorship in China. In a recent article by Henry Kissinger, Eric Schmidt, and Daniel Gattenlocker (their 2021 book hinted at the potential of general purpose AI) make a strong case that we are currently on the cusp of an intellectual revolution made possible by ChatGPT. They not only point to moral and philosophical problems created by large language generative models, but also raise important practical questions about their implementation, which are directly related to the size of the corpus of knowledge available in the language being processed.
China received the lowest score in the category «Freedom of the Internet” among 70 countries
This is where China’s strict censorship regime begins to cause concern. Both the East and the West have a long and rich history of censorship, but the propaganda department of the Chinese Communist Party stands out in its efforts to control all forms of expression in Chinese society. (newspapers, films, literature, media, education), and to manage the culture and values that define public debate.
Unlike the West, where anything goes online, China’s censors impose strict political rules on the dissemination of information that must conform to the CCP’s line. Internet users in China are unable to download links to information about the decade-long Cultural Revolution, the June 1989 Tiananmen Square tragedy, human rights issues in Tibet and Xinjiang, misunderstandings with Taiwan, the 2019 pro-democracy demonstrations in Hong Kong, political discontent zero covid and much more.
Such aggressive editing of information becomes a big problem for ChatGPT with Chinese specificity. By stripping history of important events and related human experiences, Chinese censorship narrows and distorts the corpus of information that will be used to machine-learn large language models. And this means that China’s ability to reap the benefits of the intellectual AI revolution will ultimately diminish.
A digest of the main news
Free email distribution of only selected materials from NV editors The newsletter is sent from Monday to Friday
It is, of course, impossible to quantify the effect of this censorship with precision. Qualitative assessments can be found in Freedom House’s annual Freedom on the Web survey. In 2022, China scored the lowest in the category «Freedom of the Internet” among 70 countries.
This indicator is calculated on the basis of answers to 21 questions (and on almost 100 clarifying questions), which are grouped into three broad categories: barriers to access, violations of user rights, content restrictions. Category related to content (filtering and blocking websites, legal restrictions on content, liveliness and diversity of the online information sphere, the use of digital tools for civic mobilization), allows to roughly measure the impact of censorship on the size of searchable information. China, according to these calculations, received a score of two out of a possible 35, while the average score was 20.
In the future, we can expect a continuation of the same. The Chinese government has already rushed to release draft new regulations on chatbots. April 11 China Cyberspace Administration (CAC) ruled that content created by generative AI has «be based on basic socialist values and must not contain any content that undermines state power, proposes the overthrow of the socialist system, incites the division of the country, or undermines national unity.”
And here appears the most important difference between the former regime of censorship and the new attempts to supervise AI. In the first case, content filtering by keywords is used to block unacceptable information, and in the second case (and this was discussed in the recent DigiChina forum) use a game approach «Kill a mole” to limit the rapidly changing generative processing of such information. As a result, the harder the CAC tries to control ChatGPT content, the less output China’s chatbot generative artificial intelligence will have. And this is another obstacle on the way of the intellectual AI revolution in China.
Not surprisingly, the first results of generative AI in China have been disappointing. One of the first and most famous Chinese language models today is the Wenxin Yiyan program of the Baidu company (it is called «bot Ernie”)—was recently criticized in Wired for attempting to operate in a “firewalled, government-censored Internet.” Other language AI models in China, including Alibaba’s Robot, Lily, and Tongyi Qianwen, report equally disappointing results. (translates roughly as «the truth of a thousand questions”).
Additionally, a recent analysis by NewsGuard (it «Internet Trust Tool’, created and maintained by a large team of respected Western journalists), showed that OpenAI’s ChatGPT-3.5 program generated more false information (so-called «hallucinations”) in Chinese than in English.
Philologist Jing Tzu’s excellent book Kingdom of Characters: The Linguistic Revolution That Made China Modern highlights the critical role that language played in China’s post-1900 evolution. After all, language is nothing more than a carrier of information, and in the final chapter of his book, Tsu emphasizes this fact by arguing that «whoever controls the information controls the world.”
In the era of AI, such a conclusion makes China think about serious questions. Information is the fuel for large language AI models. And state censorship squeezes China into small language models. This difference may well have a decisive effect on the outcome of the battle for informational control and global power.
NV has the exclusive right to translate and publish Project Syndicate columns. Reproduction of the full version of the text is prohibited.
Copyright: Project Syndicate 2023
Join our Telegram channel Views of NV