Domestic AI in the Age of ChatGPT
How will China and its government adapt to this new era of powerful generative AI?
Update: About a week after I published this article, the Cyberspace Administration of China (CAC) released their draft regulations governing generative AI services.
I shared some of my thoughts on these regulations with DigiChina in this piece, which also features insights from others in the China tech policy sphere.
Last week, Alibaba founder Jack Ma made a highly publicized appearance in his hometown of Hangzhou. To get the obvious part out of the way, this was Ma’s first time in China in over a year. But this event was also significant for another reason: his thoughts on the future of AI, particularly in education.
Yungu School, the Alibaba-financed institution Ma visited, published a WeChat article with several quotes from the tech founder’s visit:
While ChatGPT and other similar technologies have posed a challenge to education, ChatGPT is just the beginning of the AI age. We must use artificial intelligence to solve problems rather than let it control us. Our physical and mental abilities may not be able to surpass that of a machine, but people have hearts — machines don’t.1
The industrial age was knowledge-driven... The digital age is intelligence-driven. This age is a competition of ingenuity and imagination, of leadership and responsibility, of independent thinking.
ChatGPT is the most prominent example of generative AI: a subcategory of AI that focuses on creating new data, content, or patterns, such as text, voice, and video.
As you all but certainly know, the recent leaps of quality made across generative AI models such as GPT-4, Midjourney, VALL-E (for text, image, and voice, respectively) have caused many throughout the tech industry as well as the general public to believe that we’ve reached a paradigm shift.
Similar to Jack Ma’s sentiments above, Bill Gates has called generative AI “every bit as important as the PC, as the internet.”
China, of course, has not been sleeping on artificial intelligence. Over much of the last decade, it has built a general ideological and regulatory framework for various aspects of AI and its applications. Beijing has already rolled out regulations combating deepfakes, one of the less savory applications of generative AI models.
But with the most recent advances in the landscape of generative AI, what will China’s generative AI landscape look like over the next year or so?
“To regulate generative AI like ChatGPT, it’s essential to combine algorithm regulation with data regulation.”
China’s Emerging LLM Chatbots
OpenAI has geo-locked ChatGPT for users in China, but that hasn’t prevented the chatbot from emerging in seemingly every other tech article in China.2 Some highlights:
In February, the city of Beijing pledged to support top enterprises in “the development of large models comparative to ChatGPT.” This was in addition to other statements of support for the city’s AI industry.
On February 20, a team from Fudan University released MOSS, which the creators described as “a conversational language model like ChatGPT.” Within hours, the site hosting the chatbot crashed due to the amount of traffic it received. MOSS was soon taken offline, and a new version is currently scheduled to be released sometime in April.
At least five Chinese companies, including Baidu, Alibaba, and Tencent, have either released or announced ChatGPT-like generative AI products.
Other domestic businesses made statements regarding their own research into large language models (LLMs), the underlying technology powering chatbots like ChatGPT and MOSS.
Another company that has announced plans for an LLM-powered chatbot is China Telecom Corp, operated by state-owned enterprise (SOE) China Telecom. They intend to use it for customer service.
Of those commercial AI products mentioned above, Baidu’s new chatbot ERNIE Bot (文心一言) has gathered the most buzz. First announced in February and released to the public last month, Ernie is a knowledge-enhanced LLM powered by Baidu’s deep learning framework PaddlePaddle (飞奖) and Wenxin (文心) natural language processing (NLP) model. The public has paid close attention to Ernie and its perceived performance; Baidu’s stock initially dipped after the company demoed the chatbot via recordings and slides rather than a live demo, and it later rose after initial users tested the product and expressed general satisfaction.
As
pointed out in February, Ernie is just one facet of Baidu’s robust AI strategy, which already includes the aforementioned PaddlePaddle framework —which has also been used to create software for computer vision (PaddleDetection), image generation (PaddleGAN), and voice synthesis (Parakeet) — and self-driving cars.As just one Chinese company building generative AI products, Baidu offers a glimpse into the future of generative AI in China. But so far, we’ve only discussed how some of China’s private enterprises are implementing this type of AI tech. Considering the Chinese government’s persistent motto of self-reliance in science and technology3 — a move that is tightly coupled with factors such as geopolitical tensions and export restrictions — we need to zoom out a bit.
China’s Existing AI Regulatory Framework
In recent years, China’s government has begun constructing a framework for AI regulation. Below is a non-exhaustive list of some of the key policies that comprise it:
General principles
In 2017, the State Council published the Next Generation Artificial Intelligence Development Plan. The State Council has described this as the Chinese government's "first documents in the field of artificial intelligence addressing the matter of systematic deployment."
The Ministry of Science and Technology (MoST) published two brief policy documents in 2019 and 2021: respectively, Principles of Governance for Next Generation Artificial Intelligence4 and Ethical Principles for Next Generation Artificial Intelligence. As their names suggest, they establish a general ideological framing for the research, development, administration, and use of artificial intelligence.
Algorithm regulation
The Cyberspace Administration of China (CAC) laid out requirements for algorithm-based recommendation services in the Internet Information Service Algorithm Recommendation Administration Regulations of 2022. Included among these stipulations:
Algorithm services must “promote core socialist values.”
Entities cannot use recommendation algorithms to endanger national security or the public interest, disturb the economic or social order, or infringe on people’s legal rights.
Companies were required to register any algorithms with “public opinion attributes or social mobilization capabilities” with the CAC.
The CAC has made the information on these algorithms available via a public-facing platform. Although the details contained here are quite surface-level (e.g. they contain no code), the platform demonstrates an initial commitment to cataloging this type of information.
Deep synthesis technology
Earlier this year, China’s regulations on “deep synthesis technology” went into effect. This was a joint effort by the CAC, the Ministry of Industry and Information Technology (MIIT), and the Ministry of Public Security (MPS).
While these regulations are referred to as targeting deepfakes (i.e. digitally manipulated media often showing doctored images and audio of popular figures), their scope is wider than that. According to a recent post from an MPS-managed WeChat account, “deep synthesis technology” covers tech used to generate text, audio such as speech and music, and images — including everything from three-dimensional reconstructions to cartoon photo filters.
The regulations also include certain restrictions that directly impact businesses and developers working on public-facing LLMs. This includes producing, copying, publishing, or transmitting illegal information or fake news; endangering national security; disturbing the economic or social order; or damaging China’s national image.
ChatGPT, on the other hand, has largely been allowed to maintain its “let it rip” approach, with Italy becoming the first Western country to ban access to it, citing privacy concerns.
The CAC’s official publication elaborated further on the definition of deep synthesis in a recent explainer:
Deep synthesis technology has become a crucial technology in the global information field. In particular, this refers to the swift development of technologies — embodied by generative adversarial networks (GAN), diffusion models, and generative pre-trained transformers (GPT) — to provide strong technical support to applications such as virtual humans, human-computer interaction, and the metaverse; and support the swift application of deep synthesis technology in fields such as education, finance, healthcare, and the service industry.
Until the government releases some more specific provisions, the regulations on deep synthesis tech offer the clearest guidance for people working on generative AI models and products.
Speaking with Shanghai-based publication The Paper, Guo Jinlan, an associate professor at the Artificial Intelligence Law School at the Shanghai University of Political Science and Law, shared his thoughts on how generative AI should be regulated.
To regulate generative AI like ChatGPT, it’s essential to combine algorithm regulation with data regulation. As mechanisms for algorithm regulation continue to be explored, ample consideration should be given to the subject types and the functional classification used in algorithm-related policies, and scenario-based and granular regulatory mechanisms should be explored.
Encouraging Top-Down Innovation
Over the past month or so, top members of the Ministry of Science and Technology have commented directly on recent AI developments, particularly generative AI like ChatGPT.
As I mentioned in my coverage of the Digital China plan, Minister of Science and Technology Wang Zhigang has stressed the importance of China’s self-reliance in the areas of science and technology. Wang and his MoST colleagues have also spoken directly about China’s AI priorities.
On February 24, Chen Jiachang, head of the ministry’s Bureau of New Technologies, acknowledged that ChatGPT’s abilities demonstrated the potential for applications across many industries. He then spoke more broadly about AI, citing its “permeating, diffusive, and disruptive effect on socio-economic development,” including in scientific research.
Chen stated that China’s government had placed great importance on artificial intelligence. In addition to citing key policies published over the last few years, he elaborated on the ministry’s two-pronged approach to AI research, which has focused both on “fundamental theory” and “innovative applications.” This included an effort to drive growth at private enterprises as well as the establishment of 18 national experimental zones for AI innovation and research. Pointing to the work of Chinese researchers on multimodal Chinese language models and large models for images, text, and audio, Chen stated that by international standards, has reached an advanced level of technological achievement in AI.
But how does MoST plan to move forward — and more specifically, what are their thoughts on generative AI such as OpenAI’s GPT models?
Chen stated that MoST’s immediate priorities were to view AI as a “strategic emerging industry” and a “new engine for growth” and to continue its “vigorous support” for AI. He laid out four specific priorities:
Constructing a collaborative system of innovation for AI. This primarily includes accelerating work on foundational theoretical research and key technological issues.
The deep integration of AI with China’s economic society.
Building a secure and controllable administrative system for AI.
Promoting open collaboration in the field of AI.
Wang Zhigang, Minister of Science and Technology, addressed the issue of ChatGPT more directly. In his view, China had made notable achievements in the field of natural language processing (NLP), as well as the “effective integration” of the components necessary for properly training and supporting a large language model like GPT — big data, major computing power, strong algorithms.
Over a week later, the MoST head fielded another question about ChatGPT. While not delving into specifics, Wang Zhigang mentioned that the impact of AI “goes far beyond the realm of science and technology.” He reiterated his hope that universities, research institutes, and businesses involved in AI research could contribute to China’s objectives in the grand scheme of global AI development.
Change Ahead
China is due for some more concrete policies tackling user-facing generative AI services such as chatbots and AI art generators:
Major players ranging from Baidu to Beijing’s municipal government have invested heavily in LLMs and other generative AI tech.
In the face of tech export restrictions and increased international progress in generative AI technology, MoST has emphasized the importance of AI and NLP, China’s need for self-reliance in tech and science, and the importance of enterprise-driven innovation.
The state recently announced a major restructuring that includes the reorganization of MoST.
A recent article co-authored by MoST vice-secretary He Defang discussed China’s national innovation system, which prioritizes contributions from enterprises and institutes of higher education (among others) for driving national goals of social and economic growth.
Sources
Baidu Baike: 飞奖
Baidu Baike: 文心一言
BBC: ChatGPT banned in Italy over privacy concerns (April 2023)
Beijing Municipal Government: 《2022年北京人工智能产业发展白皮书》重磅发布
Chinese Academy of Sciences Bulletin: 国家创新体系的发展演进分析与若干思考 (March 2023)
Cyberspace Administration of China (CAC): 互联网信息服务算法推荐管理规定 (January 2022)
Reuters: Factbox: Chinese firms working on ChatGPT-style technology (March 2023)
South China Morning Post: Baidu’s ChatGPT alternative gets positive reviews for handling of Chinese translations as search giant’s stock jumps (March 2023)
State Council Information Office: 《新一代人工智能发展规划》政策解读 (July 2017)
The Paper: ChatGPT火爆背后有何法律风险？记者调查 (February 2023)
Xinhua: 科技部部长谈ChatGPT时，举了这样两个例子 (March 2023)
Xinhua: 中共中央 国务院印发《党和国家机构改革方案》(March 2023)
Yungu School: 马云来云谷学校，和校园长讨论了未来教育 (March 2023)
中国警察网：使用ChatGPT的注意了！这些事不能做 (April 2023)
中国网信杂志：深度解读《互联网信息服务深度合成管理规定》(March 2023)
Jack Ma’s quote contains a pun that only a literal translation conveys, playing off the identical pronunciations for “core” and “heart”: “Machines only have ‘cores’ (芯, xīn); people have ‘hearts’ (心, also xīn).”
The geolock also hasn’t prevented people in China from accessing it via various workarounds, but that’s another story.
For more on this, see last month’s article on the Digital China plan.
Also referred to as the “eight principles,” or 八项原则.
great stuff! my takeaway is they still have no idea!