History will likely mark August as a watershed moment for China’s generative AI (GAI) era. Shortly after interim GAI regulations went live mid-month, the country’s foundational models, like Baidu’s Ernie, were approved for public use. GAI could unlock enormous economic productivity for China as it faces population decline,
GAI could unlock enormous economic productivity for China as it faces population decline. Thus, the government is keen to balance its security concerns with innovation. AI in China is particularly focused on commercial readiness and industrial specialisation.
As we learned in our Forums, Western firms in China will have to contend with heightened Chinese competitiveness. However, they must be cautious when implementing global GAI approaches to stay China-compliant. One suggestion: develop "two brains".
‘We are going to have two different AI ecosystems. As a business, you will have started to adopt a global brain (plugged into AI), but you will have to have a China brain, too.’
This Insight is based on key points of discussion and comments among members of our China CEO Forum and General Managers Network at briefings in May 2023. It is augmented by further research by our editorial team between June and September 2023.
Members may access the full report via login. Contact us to learn about membership.
Key Takeaways
+ China's AI ecosystem will evolve differently from the West due to distinct regulatory conditions
prioritising security, accountability and industry-specific growth to boost the economy.
+ MNCs will use AI with multiple ‘brains’ – one for public information, another for global
proprietary information, and a third to train on local data in China.
+ China’s AI focus is not limited to security and control, but also includes industrial and economic
applications. Chinese competitors will exploit generative artificial intelligence (GAI) to boost
productivity, lower costs, and appeal to consumers.
Comparable to other general-purpose technologies like steam and electricity, generative artificial intelligence (GAI) will transform the world. How the two leading AI nations – the US and China – approach it will markedly differ, putting MNCs in China in the middle.
China will seek to deploy generative AI to enhance its national and domestic security, but also, as is often overlooked by the press, to supercharge its economic productivity. Adopting AI is mission-critical so China can clear the middle-income trap.
Chinese firms view AI adoption as an opportunity, while many in MNCs’ home markets may see it as a threat. Indeed, China faces a declining demographic dividend but has a robust data dividend to exploit – although much of that data will remain in China.
A quick catch-up on recent developments
It should be no surprise that China is blazing a trail in GAI on its own terms, driven by the Party’s priorities.
New rules announced: China launched interim GAI rules on August 15th. China’s AI rules for data security, training data, traceability, watermarking, algorithms, and accountability are more advanced than Western countries, which are still in the discussion phase.
‘China’s AI regulations create new bureaucratic and technical tools: disclosure requirements, model auditing mechanisms, and technical performance standards.’
China approves GAI models: OpenAI’s launch of Chat GPT took Chinese AI experts by surprise. Chinese models went public nine months after ChatGPT’s launch in November 2022. The state approved Baidu’s Ernie Bot and other large language models (LLMs) for general use on Aug. 31, 2023.
Pent-up demand will mean Chinese language AI tools will soon flood the market as developers experiment with the models. Ernie Bot was downloaded one million times within 19 hours of its release, a milestone that took Chat GPT five days to reach.
‘The US already has many AI start-ups that are building a new generation of applications on top of the LLMs. There have been few in China because they were waiting for Baidu and Ali to provide their LLMs officially.'
Chinese chip capabilities improved: The 7 nm semiconductor chips (Kirin 9000s) found in the Huawei Mate 20 Pro phone means China caught up faster to the leading edge than expected, exploiting loopholes in the US CHIPS Act. Better chips allow China to secure the considerable computational power needed for GAI more cheaply.
Building your China brain alongside your global brain
Much like the internet and social platforms developed differently in China, so will AI models and the tools built upon them. The ubiquitous ‘ecosystem’ metaphor remains apt; as species evolve according to a unique set of conditions, so will AI.
‘We are going to have two different AI ecosystems. As a business, you will have started to adopt a global brain (plugged into AI), but you will have to have a China brain, too.’
Two unique conditions in China that will spur adaptation are:
1. rules for a ‘digital border’ that hems in many data types; and
2. leading AI models designed for industry-specific use cases.
A ‘China AI brain’ can serve as a booster or co-pilot to service the China business. Firms can rely on third-party providers or attempt to develop their AI brains in-house.
‘Firms can use the services of internet giants to build applications quickly. Or, for firms with enough data and long- term aspirations, they can build their own models leveraging open sources or proprietary services.’
Many employees outside China have been experimenting with existing public AI models, such as Chat GPT. Chinese employees have likely done the same with Baidu's Ernie and others in the past few weeks. China CEOs will have to decide whether their firms can trust their data to these public LLMs and chatbots.
'Many MNCs are setting up islands within their IT organisation to test Chat GPT. But, when proprietary business information is uploaded, it becomes public. If you have uploaded an instruction manual, it could mean that you gave up a revenue stream for your service organisation.’
The big guys know they have a problem being enterprise safe. OpenAI, Salesforce and Azure recently launched business-grade models with GAI to allay concerns about data security, proprietary information, and hallucination (or lack of accuracy). Adobe has trained its model on its proprietary image database, not the internet, avoiding copyright tussles plaguing other foundational models.
Action point: Firms need guidelines for how to use LLMs securely.
When two brains are better than one
One consultant predicts that firms will adopt a two ‘brain’ system that will leverage foundational public data and internal models trained on domain-specific, proprietary data.
‘The AI companies, like OpenAI, Google and Microsoft' can provide a massive brain with most of the world's knowledge. But there are smaller brains within enterprises, what we call “privately deployed AI services.”
‘When domain knowledge is required, it gets routed to an internal data centre, while common knowledge questions will get routed to the big brains, like Google or OpenAi. Firms will use the big brain to enhance the private brain.’
This two-brain (or more) system will allow firms to develop a local China system to stay within China’s data residency rules, while avoiding disintermediation by stronger AI players.
‘When Microsoft launched Windows and made it easier to access information directly through the internet, many MNCs feared they would lose control of their consumers. It's going to happen again; what's happening with AI is a major shift in power and ability to control information.’
China’s pragmatic, industrial focus
Written and unwritten rules governing political sensitivity hem in China's LLM developers, making it a great motivator to stay on safer ground, like developing AI for industrial applications.
'Huawei, Baidu, Tencent, and Alibaba are trying to figure out how to leverage their underlying datasets and AI expertise in developing LLMs to service clients in specific industry verticals.’
Huawei launched its enterprise model in July. The Pangu system via Huawei Cloud offers customers large-scale industry development kits. Through secondary training on customer-owned data, Huawei customers can have exclusive large models by industry, such as shipping, one area of focus for the firm.
Huawei CEO Zhang Pingan said Pangu, has no time to compose poetry, making an undisguised dig at Chat GPT. During Pangu’s July launch, Zhang remarked, ‘It was designed for industry and will be dedicated to industry.’ This is no surprise since China’s approach to AI is entrenched in the practical.
‘China has focused on automating drudge work: inspecting parts on a factory conveyor belt, checking the bins near the coal face for foreign objects, detecting anomalies in machines, picking containers out of ships and placing them on autonomous trucks, and so forth.’
A race is on to see which GAI model is the best. China will benefit significantly from open-source models, which are catching up fast.
‘The technology gap between OpenAI and Baidu is relatively small. Open-source models have almost broken the advantage of advanced Western models. In terms of language capability, open-source models are on par with Open AI and Google.’
AI regulations prepare for the 4th Industrial Revolution
Although China approved the use of foundational models well after the US, Chinese firms were quick to develop their models. Chinese researchers have been galvanised by open-access research and made significant contributions to the field of natural language processing.
China’s New Generation Artificial Intelligence Development Plan of China (2015–2030) set itself up for success to exploit this GAI era. The plan prioritises funding for core AI algorithms, hardware infrastructure, AI talent, and open-source models. It also emphasises integrating AI into industry verticals, such as healthcare and manufacturing. Practical and commercial applications are a recurring theme in China's AI trajectory and GAI will be no different.
Since the plan’s release, the Cyberspace Administration of China (CAC)
continues to issue detailed rules for AI to address real-world concerns
as they arise such as deepfakes, manipulative algorithms, and fake
news. The rules and regulations:
+ Bar excessive price discrimination and protect workers' rights.
+ Require conspicuous labels on synthetically generated
content.
+ Insist that training data and model outputs be ‘true and
accurate.’ This may be an insurmountable hurdle for AI
chatbots to clear.
No other country regulates algorithms like China. Developers must file with China’s algorithm registry and pass a security self-assessment.
Action point: Legal experts recommend that companies conduct an AI audit and develop an AI governance structure to ensure compliance.
The interim GAI rules attempt to balance security with innovation.
Some key points:
+ Public access restrictions. The measures only apply to GAI
public service providers and do not apply to businesses using
GAI tools for internal purposes.
+ Industry-led. Regulators will be responsible for further
industry-specific rules, so more detailed rules are to be
expected (e.g., for mining, and healthcare).
+ Clear corporate accountability. Companies will have to ensure
technologies and services do not infringe on the legal rights of
individual users and are in line with existing privacy laws.
+ Data residency. Foreign businesses using generative AI will
become reliant on local partners to ensure compliance.
China will become a testing ground for advanced AI controls and will have sway over international rules. What happens in China regarding AI is unlikely to stay in China.
‘The requirements imposed on China's AI products matter. They will reshape how the technology is built and deployed but their effects will not stop at its borders. China is the world's largest producer of AI research. Its regulations will drive new research as companies seek to meet regulatory demands.’
Members may access the full report via login. Contact us to learn about membership.
