Summary of Key Points
This article discusses the risks and governance challenges associated with the globalization of AI, emphasizing that the AI industry chain has transcended national boundaries, making it impossible for any single country to control it on its own. Data is not merely a "resource" but the "soil" upon which AI relies; its quality and dynamic flow directly impact the security of AI systems. Generative AI technologies can use poor-quality data to produce seemingly credible content, leading to the spread of misinformation from generation to generation. The focus of governance should shift from static datasets to dynamic data flows. The international community should establish a minimum common ground in governance rather than attempting to create uniform rules to ensure the safe and trustworthy development of AI.
1. The Globalization of AI: Cross-Border Risks That No Single Country Can Overcome
The AI race has transformed into a "global relay race"—an AI system may be developed in the United States, run on servers in Europe, utilize data from India, and ultimately affect users in China. This distributed model means that no country can oversee the entire industry chain independently. For example, if an AI is trained using biased data in Japan and then launches services in Australia, local users could be misled. Since Japan cannot regulate Australian usage, nor can Australia control the training data in Japan, coordination among countries becomes necessary. However, such coordination is only possible if there is a clear understanding of what needs to be regulated: the developers of AI, the providers of data, or the companies that use it? Without a defined target, any effort at coordination is futile.
2. Data Is Not Oil; It's the "Soil" for AI
The saying "data is oil" is misleading because oil can be depleted with use and requires extraction costs, while data can be replicated indefinitely, often being freely provided without us even realizing it (such as browsing history left behind on mobile devices). A more accurate analogy would be "soil"—without proper data quality and appropriate sources, AI cannot function effectively. If the "soil" (data) is contaminated (e.g., contains errors or biases), not only will AI performance suffer, but the contamination can also spread; erroneous data can be used by other systems as "fertilizer," exacerbating problems.
3. The Deceptive Power of Generative AI: Misinformation That Spreads
In the past, computer science held the belief that "garbage in, garbage out"—bad input would result in bad output. However, generative AI technologies like ChatGPT have broken this rule. They can produce content that sounds authoritative but is actually incorrect, using flawed data as a basis for training. For instance, an AI could generate a "credible" article based on inaccurate historical information, which users might believe and share, allowing the error to be further propagated by other systems. Such seemingly credible misinformation is more dangerous because it is easier to accept.
4. The Shift in Governance Focus: From Data Sets to Dynamic Data Flows
Early AI governance focused on training data—whether it was biased, legally obtained, and of sufficient quality. But today's AI systems are capable of accessing information in real time, remembering user interactions, and even assisting with tasks like document editing or shopping. Static data set checks are no longer sufficient; we need to manage the "dynamic data flows"—which information AI systems have access to, what personal information they can gather (e.g., emotions, weaknesses), and what actions they can perform (e.g., altering bank records). Users should be aware of these risks and may need to establish "digital prenuptial agreements" with AI systems, specifying which data should be deleted when the relationship ends.
5. International Governance: Minimum Common Ground Rather than Uniform Rules
Given the varying laws and values across countries, it is nearly impossible to establish a global unified regulatory body or set of rules for AI. However, a minimum consensus can be reached on key aspects such as the legality of data sources, mandatory security testing for AI systems, accountability in case of issues, and clear roles for developers and users. Similar to the aviation industry, where each country has its own regulations but adheres to international safety standards, AI governance should ensure interoperability so that systems can operate safely across borders. If national rules are not compatible, companies will face compliance challenges, and users may suffer due to gaps in jurisdiction.
In conclusion, the article uses a metaphor to illustrate that governance is not about creating barriers to innovation but about providing a "safety belt" that makes innovation possible within a secure framework. While risks associated with AI cannot be completely eliminated, they can be managed through proper data management, clear accountability, and international cooperation to ensure that AI develops in a safe and responsible manner. Only a healthy "soil" of data can nurture truly valuable AI applications.