A. Comparing U.S., China, and EU Regulation

Picture a scene from spring 2024 at the 'AI for Good' Global Summit in Geneva, Switzerland. On stage sat a technology evangelist from Silicon Valley, a European Union bureaucrat who had flown in from Brussels, and a Chinese policy maker dispatched from Beijing. They all used the same phrase: 'safe AI.' But what the phrase meant to each of them was completely different.

(1) America's Fair Use vs. China's Strict Protection

To understand the American approach, you first need to grasp the concept of 'Fair Use.' Put simply, it means there are situations where you can use someone else's copyrighted work without permission. A student quoting a book in a report. A critic analyzing a movie scene. These are acceptable. The United States is trying to apply this long-standing legal doctrine to the AI era as well.

In American courts, the lawsuit between The New York Times and OpenAI is ongoing. Class-action suits from authors continue. But the basic posture of the U.S. judiciary is clear: don't block innovation; if problems arise, deal with them then. In May 2025, the U.S. Copyright Office (USCO) released its third report on generative AI training. The report left open the possibility of fair use while also leaving room for licensing negotiations. It reached no conclusion. In American terms, the attitude is: "Let the courts decide."

China is different. China created the world's first binding regulation on generative AI. That is the 'Interim Measures for the Management of Generative AI Services,' which took effect on August 15, 2023. The core of this regulation is the legality of training data. To train an AI model, data sources must be lawful. Intellectual property rights must not be infringed. And here a uniquely Chinese condition applies: the data must conform to "core socialist values."

The Beijing Internet Court's 2024 'AI text-to-image' ruling left an interesting precedent. The court recognized copyright in AI-generated images where a human had adjusted prompts and parameters. This contrasts with the U.S. ruling in Thaler v. Perlmutter, which denied copyright for AI-created works without human involvement. China is more progressive in granting copyright to AI-generated outputs. But in return, its control over data use is far stricter. This looks like a contradiction. But within the larger picture of 'state-led technology control,' it is a coherent strategy: protect the fruits of creation, but let the state manage the raw materials.

The Guangzhou Internet Court's 'Ultraman' ruling shows this philosophy even more clearly. In that decision, the court imposed a high duty of care on AI service providers. Beyond merely providing a tool, providers must take active measures such as keyword filtering and data-source verification to ensure generated content does not infringe on others' copyrights. This is a level of platform responsibility difficult to imagine in the United States.

(2) EU's Rights-Centered vs. China's Control-Centered Approach

Both the EU and China have introduced strong regulations. On the surface, they look similar. But when you ask whom the regulation is meant to protect, a fundamental difference emerges.

On August 1, 2024, the EU AI Act entered into force. It is the world's first comprehensive AI framework law. Its core is the 'risk-based approach.' It classifies AI into four tiers by risk level: unacceptable, high-risk, limited risk, and minimal risk. For example, government-operated social credit scoring systems are prohibited in principle because of the significant risk they pose to civil liberties. The case where Italy's data protection authority (Garante) temporarily blocked ChatGPT and ordered improvements illustrates the EU's posture well. In the EU, 'safe AI' means AI that does not violate individuals' fundamental rights.

The EU AI Act's Code of Practice, published in July 2025, contains specific guidelines on copyright compliance. General-purpose AI model providers must transparently disclose the sources of their training data. They must respect copyright holders' 'opt-out' declarations refusing to allow training use. EU regulation focuses on protecting copyright holders' rights, individual privacy, and democratic values.

China's regulation looks in a different direction. The Cyberspace Administration of China (CAC) requires AI service providers to register their algorithms. AI services that influence public opinion or possess social mobilization capabilities must undergo security assessments. The 'AI-Generated Content Labeling Regulations,' effective September 1, 2025, mandate that all AI-generated content carry explicit or implicit labels. This appears to be about transparency. But the purpose of that transparency differs from the EU's. Where the EU aims to inform consumers, China aims to track and control the flow of content.

In China, 'safe AI' means politically correct AI. The 'AI Plus' initiative implementation guidelines released in August 2025 set targets for the penetration rate of a new generation of intelligent terminals and AI agents to exceed 70% by 2027. China wants AI to spread widely. But it wants that spread to happen within party lines.

(3) Concentration in the Top 5% Among 1,500 AI Companies

China has approximately 1,944 AI companies. The number alone suggests a thriving ecosystem. But the reality is different. A handful of giant corporations dominate most of the market.

In 2024, China's AI public cloud market reached 19.6 billion yuan (roughly $2.7 billion), growing 55% year over year. Baidu and Alibaba each held about 25% of this market. Tencent and Huawei followed. In the first half of 2025, Alibaba Cloud commanded 35.8% of the AI cloud services market, ByteDance's Volcano Engine held 14.8%, Huawei Cloud 13.1%, Tencent Cloud 7%, and Baidu Cloud 6.1%.

The top five companies take more than 75% of the market. The remaining 1,900 companies split the other 25%.

This concentration is no accident. China's regulations raise barriers to entry. To launch a generative AI service, you must register your algorithm with the CAC. You must pass a security assessment. You must prove the legality of your training data. You must have a system capable of filtering tens of thousands of sensitive keywords. You need a monitoring framework that can respond immediately when problems arise. All of this demands enormous manpower and capital. It is a cost that startups can hardly bear.

The Chinese government also prefers a small number of large companies for the efficiency of oversight. Supervising 1,900 small and mid-sized firms one by one is harder than managing the AI ecosystem through a few 'national champion' companies under firm control. The State Council has designated 15 companies, including Baidu, Tencent, Alibaba, SenseTime, and iFlytek, as the 'National AI Team.' Each company has been assigned a role leading AI development in a specific field such as facial recognition or speech recognition.

In the United States, open-source communities and numerous startups compete or collaborate with Big Tech, diversifying the ecosystem. In China, high regulatory barriers create a structure that favors the top players. Newcomers like Baichuan AI, Moonshot AI, and Stepfun were founded in 2023 and jumped into large language model development, but whether they can threaten Baidu or Alibaba remains uncertain. They need the capital and political connections to clear the regulatory hurdles.

B. The Self-Limiting Nature of China's Regulation

Imagine an AI startup office in Beijing's Haidian District. Developer Li sits in front of his monitor, head in his hands. His goal is to build a large language model that surpasses GPT-4. Technically, it seems possible. He is a brilliant engineer. His company has sufficient funding. But there is an invisible ceiling above him.

(1) How the Data Blacklist Stunts Growth

An AI model's intelligence comes from the volume and diversity of its data. This is known as the 'Scaling Law.' The more data you train on, and the wider the range of perspectives it absorbs, the smarter the model becomes. The problem is that China's regulations restrict this diversity of data.

Article 4 of China's 'Interim Measures for the Management of Generative AI Services' stipulates that data used for AI training must come from lawful sources, must not infringe intellectual property rights, and must be free of 'content problems.' The phrase 'content problems' is vague. But in practice, it is clear. Politically sensitive topics, certain historical events, content critical of the government, and data containing Western values must be excluded from training.

This is called the 'data blacklist.' It is not an official term. But it exists in practice. The CAC strictly prohibits blacklisted content from seeping into training data.

The problem is that for AI to understand the world, it must learn from diverse viewpoints and conflicting information. A model trained only on datasets with predetermined 'correct answers' performs poorly at solving complex problems. Its creative reasoning weakens. A significant portion of global internet data, including English-language sources and communities with free-flowing debate, either cannot be used or must undergo extensive censorship before use. This leads to degraded model performance.

The DeepSeek case that emerged in early 2025 illustrates this dilemma well. DeepSeek-R1 demonstrated performance comparable to global frontier models. China's AI industry celebrated. But at the same time, one fact became apparent: the more the model relied on China-centric data and domestic developers, the stronger the 'Chinese characteristics' in its AI responses became. For domestic users, this may not be a problem. But in global markets, it becomes a factor limiting universal acceptance.

(2) Concerns Over Weakened International Competitiveness of Domestic AI Firms

China's strict regulatory environment acts as a structural obstacle for domestic companies competing in international markets.

First, algorithm pre-approval requirements and ideological compliance mandates slow the pace of innovation. The CAC and standardization body TC260 require pre-approval for high-risk AI systems, mandatory registration of foundation models, and safety audits. Three national standards to take effect in November 2025 specify requirements for training data security, pre-training and fine-tuning data security, and basic security requirements for generative AI services. This regulatory burden becomes a barrier to entry for startups and small firms.

Second, isolation from global supply chains deepens the technology gap. In 2024, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) expanded semiconductor export controls. Controls on high-bandwidth memory (HBM) and dynamic random-access memory (DRAM) were added, affecting Korean companies doing business in China as well. Huawei's latest AI chips depend on technology from memory chip manufacturers like Samsung Electronics and TSMC. But these companies, subject to U.S. export regulations, cannot supply cutting-edge technology to Chinese customers.

Capable Chinese AI chip design firms like Cambricon exist. But they have been unable to match the performance of Nvidia or AMD due to a lack of economies of scale. Shortly after TechInsights published its analysis of Huawei's latest AI chip in January 2025, China blacklisted the research firm, prohibiting all Chinese companies from cooperating with it. Such measures further reduce transparency in China's chip manufacturing sector and deepen the disconnect from the global technology ecosystem.

Third, a lack of depth in basic research undermines long-term competitiveness. Chinese AI companies' patents cite academic literature at rates exceeding those in the United States. But the average number of citations per patent is low. This reflects a deficit in basic research capacity and industry-academia collaboration.

(3) The Long-Term Cost of Market Closure

China blocks access to advanced Western AI services such as OpenAI's ChatGPT, Google's Gemini, and Anthropic's Claude. In the short term, this provides a shield for domestic companies like Baidu and Alibaba, letting them capture the domestic market.

But in the long term, it exacts the cost of 'technological isolation.'

Developers around the world share the latest models and data through open-source communities like Hugging Face and GitHub, exercising collective intelligence. Chinese developers risk being marginalized from this current due to restricted network and data access. In 2025, The Washington Post reported that China had surpassed the United States in releasing publicly available 'open' AI models. But if these models are used only within a closed ecosystem, they may lose compatibility with global standards.

China's complex regulatory environment also suppresses foreign direct investment (FDI). Data requirements and trade investigations create uncertainty for Western investors. In 2024, the U.S. Treasury Department issued rules regulating American outbound investment in China. Many Chinese AI companies, cut off from U.S. investment, are struggling to raise capital.

The most damaging cost is the brain drain. Top-tier AI researchers who prize creativity and free inquiry avoid environments where their research topics are censored or their technology is used only as a political tool. Many Chinese-born AI talents have already left for Silicon Valley, Canada, or Europe.

The Chinese government wants the most powerful AI possible. At the same time, it fears AI becoming too intelligent and slipping beyond control. Caught in this dilemma, China is building walls to block external threats. But those walls may also become a prison, preventing internal innovative energy from flowing out and fresh air from coming in.

C. Implications for Korean Companies

The lights in Pangyo, south of Seoul, stay on late into the night. In the strategy meeting room of an AI company, CEO Park stands before a whiteboard. His company possesses outstanding image-generation AI technology. The Korean market is narrow. The American market is fiercely competitive. His eyes turn toward China, with its enormous data pools and consumers. But he knows: China is a land of opportunity and a minefield of regulation at the same time.

(1) Mandatory Regulatory Compliance When Entering the Chinese Market

Any company entering the Chinese market must know three laws: the Cybersecurity Law (CSL), the Data Security Law (DSL), and the Personal Information Protection Law (PIPL). These three statutes form the backbone of China's data regulation framework. For AI companies, they are scripture.

The first requirement is data localization. Under the PIPL, personal information collected within China must, in principle, be stored within Chinese borders. Data processed by Critical Information Infrastructure Operators (CIIOs) or classified as 'important data' must remain onshore. This means AI models, training data, and service outputs all need to reside on servers inside China. Korean companies must either build dedicated data centers in China or partner with local providers such as Alibaba Cloud or Tencent Cloud.

The second requirement is a security assessment for cross-border data transfers. To send data collected in China to a Korean headquarters or a third country, companies must undergo a security assessment by the CAC. They must follow one of three mechanisms: signing Standard Contractual Clauses (SCCs), obtaining personal information protection certification, or completing an individual security assessment. New regulations in 2024 introduced some exemptions and flexible thresholds, but the approval process remains complex.

The third requirement is algorithm registration and prior approval. If an AI service can shape public opinion or has the capacity for social mobilization, its operator must report the algorithm's mechanism and data to the CAC and register it within ten days of launch. Labeling regulations effective September 2025 mandate that all AI-generated content carry explicit or implicit labels. Korean companies must understand and prepare for the CAC's approval process before releasing AI products for the Chinese market.

The fourth requirement is content control and ideological compliance. Chinese regulations explicitly require that AI-generated content align with 'socialist core values.' Content filtering is necessary for politically sensitive topics, criticism of the Chinese government, and certain historical events. Korean companies must reflect China's content policies in AI model training and build real-time monitoring and filtering mechanisms. In 2024, China's CAC took enforcement actions against companies that introduced overseas generative AI services without proper procedures or used customer personal information to train algorithms without authorization. In the first half of 2024 alone, the Chongqing CAC shut down 142 websites, conducted corrective interviews with 101 platforms, removed 21 mobile applications from app stores, and issued 11 administrative penalties.

(2) Data Purification Strategy

A dataset that works perfectly well in Korea can cause serious trouble if brought into China without modification. Chinese law stipulates that AI training data must not only be free of intellectual property infringement but must also not harm social order or threaten national security.

The first principle of data purification is data minimization and anonymization. Before entering an AI workflow, only essential data should be collected and processed. Where possible, personal information can be anonymized or pseudonymized to fall outside the scope of PIPL.

The second principle is the screening and verification of AI training data. When using training data derived from Chinese legal documents, companies must ensure such data is processed exclusively within Chinese borders. A filtering process is needed to trace data sources and remove elements that conflict with China's content policies in advance. On topics such as Taiwanese independence, the Hong Kong protests, or criticism of leaders, models must either refuse to respond entirely or be fine-tuned to provide only the Chinese government's official position.

The third principle is diversification and separation of data sources. Korean companies should consider developing separate AI models for the Chinese market and the global market. A China-specific AI instance is trained on Chinese data and deployed only within China. The global model either excludes Chinese data or uses only sufficiently anonymized data. This is the 'two-track' strategy. It carries the burden of duplicate investment, but it minimizes regulatory risk and effectively meets each market's requirements.

The fourth principle is the use of automated data governance tools. AI-based data purification tools provide functions such as audit trails, automated data masking, and encryption. Korean companies can reduce human error and respond quickly to regulatory changes through automation and real-time monitoring of the data purification process.

(3) The Importance of a Localization Strategy

Success in business, beyond just regulatory compliance, requires thorough localization. This means 'regulatory and cultural localization' that goes well beyond language translation.

First is cultural adaptation of products and services. The experiences of KFC and McDonald's in the Chinese market are instructive. KFC adapted its products and advertising to align more closely with Chinese traditions and cultural expectations. The result was greater success. Korean AI companies must also adjust their AI models and interfaces to reflect Chinese consumer preferences, linguistic nuances, and cultural context. Strategies optimized for Chinese holiday themes, collaborations with Chinese celebrities and influencers, and Chinese social media platforms like Weibo and WeChat are all necessary.

Second is building local partnerships and joint ventures. In China's complex regulatory environment, strategic partnerships with local firms play a critical role in regulatory compliance, market access, and brand trust. Korean AI companies can tap into local knowledge, data access, and regulatory networks through collaboration with major Chinese technology companies like Baidu, Alibaba, and Tencent or with regional AI startups. It is extremely difficult for a foreign company to navigate China's complex regulations and opaque administrative procedures on its own.

Third is a deep understanding of the Chinese AI ecosystem. China has made AI education mandatory in primary and secondary schools by 2025. Roughly 4,500 AI companies are currently operating in the country. Korean firms must continuously monitor China's AI talent pool, research trends, and government policy direction.

Fourth is a sub-market and regional expansion strategy. China's AI market is concentrated in tier-one cities like Beijing, Shanghai, and Shenzhen. But untapped opportunities exist in tier-two and tier-three cities. Korean companies need to study consumer behavior and the competitive landscape by region and develop tailored product positioning and pricing strategies.

Fifth is digital marketing and e-commerce integration. China has the most advanced e-commerce ecosystem in the world. Companies must actively use China-specific digital channels such as TikTok (Douyin) live commerce, 24-hour AI streaming, and virtual streamers.

On January 21, 2025, Korea enacted its own 'Framework Act on Artificial Intelligence,' set to take effect on January 22, 2026. Korean companies now face the challenge of complying with both their own country's AI regulations and China's simultaneously. The last word Director Park wrote on the whiteboard was 'survival.' If a company cannot clear the waves of Chinese regulation, even the most brilliant AI is nothing more than a pretty ornament. But clear those waves, and a market of 1.4 billion people opens up.