浅析中美人工智能监管的差异（双语）

如今，人工智能已成为中国跻身先进经济体的核心支柱。 在中国，人工智能的发展得到了政府的大力支持，亦得益于中国庞大的人口数量和通过各类数字平台收集的数据。 因此，中国在通过国家立法管理人工智能方面一直走在前列，为企业提供指导，并为人工智能的应用创造安全的环境。除了对人工智能的开发和使用施加限制外，这些法规 旨在表明政府对人工智能发展的支持态度 ，并明确各利益相关方在人工智能治理中的角色和责任 。
美国则更多地依靠政府机构扩大其现有职权范围，以涵盖对人工智能的使用和开发的监管，总体上鼓励充满活力的私营经济通过创业发展人工智能。
在本文中，我们试图探讨和比较两国监管和执法活动的不同路径。
Artificial intelligence has become a central pillar of China in its drive to become an advanced economy. The development of AI has enjoyed tremendous government support, benefitting from a large population and data collected in China through various digital platforms. As such, China has been at the forefront of adopting national legislation governing AI to give guidance to companies and to create a safe environment for its adoption. In addition to imposing restrictions on the development and use of AI in China, these regulations aim to present the government’s pro-growth attitude towards AI and delineate the roles and responsibilities of various stakeholders in AI governance.
The United States has relied more on existing agencies expanding their existing scope to cover AI use and development, by and large encouraging its dynamic private sector to develop AI through entrepreneurship.
In this article, we look deeper and compare the approaches to regulation and law enforcement activities in both countries.
算法的透明度和公平竞争
中国在研究算法的设计和部署方面一直具有前瞻性 ，其重点尤其放在算法驱动的推荐技术的应用上，包括人工智能生成内容（ AIGC）、个性化推送、自动决策和其他相关在线服务。
其中一项重要法规是 2021 年颁布的《互联网信息服务算法推荐管理规定》（《算法管理规定》）。据此，中国国家互联网信息办公室有权（CAC）要求平台对其使用的推荐系统保持公开透明，并公布所使用算法的基本原则、目的和主要机制。此外，该规定还允许用户可以选择不使用算法，以避免成为分析和算法应用的对象。如果算法可能对用户的权益产生重大影响，服务提供商必须做出合理解释，并对任何滥用行为负责。如果服务能够影响公众舆论，则必须将算法的详细信息提交给 CAC，以便对潜在的偏见和滥用行为进行审查。
在保护公众利益方面，《算法管理规定》特别强调了保护弱势群体的重要性。特别是，企业应根据未成年人和老年人的特定需求、其心理和身体状况为其定制算法推荐服务，避免利用算法使未成年人沉迷于网络服务（如网络游戏或短视频）或使老年人遭受电信诈骗。在使用算法进行人力管理调度时，必须保护工人获得报酬和休息的权利。
在促进互联网服务企业的正当竞争方面，《算法管理规定》禁止企业利用算法干扰、破坏或不合理地限制其他企业的在线服务。
美国尚没有一部专门规制算法技术的法律。 但相关政府机构会利用现有的职权保护公众利益。 例如，美国联邦贸易委员会（FTC）有权对 "不公平或欺骗行为 "进行监管，涉及对人工智能的误导性宣传或者使用人工智可能导致的歧视性结果。
在美国，也有人试图直接对人工智能进行立法。 例如，一项名为《算法问责法》的法案已经被提交至国会，要求开发者对人工智能系统的歧视和隐私风险进行审计。 但迄今为止这项立法尚未获得通过。此外，相关行业组织也推出了自愿遵守的行动指引，如美国国家标准与技术研究院（NIST）提议的人工智能风险管理系统就要求开发者采用最佳实践来管理风险并提高算法的透明度。
Transparency and Fair Competition when Using Algorithms
China has been proactive in looking at the design and deployment of algorithms. The focus in particular has been on the application of algorithm-powered recommendation technologies, including AIGC, personalized push, automated decision-making, and other relevant online services.
One of the key regulations has been the Administrative Provisions on Algorithm Recommendations for Internet Information Services promulgated in 2021 (Algorithm Regulation). It enables the Cyberspace Administration of China (CAC) to require platforms to be more open and transparent about the recommendation systems they use and publicize the basic principles, purpose and major mechanisms of the algorithm in use. In addition, it allows users to be able to opt-out of the algorithm to avoid profiling. In case the algorithm may have a significant impact on a user’s rights or interests, the service provider must provide reasonable explanations and will be held liable for any misuse. Details of the algorithms have to be filed with the CAC if the services can shape public opinions, so that (in theory), they can be looked at for potential biases and abuse.
On the protection of public interests, the Algorithm Regulation specifically empathizes the importance of protecting vulnerable groups. In particular, businesses shall tailor the algorithm-powered recommendations for minors and elders according to their specific needs as well as their mental and physical conditions, and shall avoid using algorithms to make minors addictive to online services (e.g. video games or video streaming) or expose elders to telecommunication frauds. When using algorithms to assign tasks to contract workers, workers’ rights to obtain compensation and rest must be protected.
On the promotion of fair competition in online services, the Algorithm Regulation prohibits businesses from using algorithms to interfere, undermine or impose unreasonable restrictions on other businesses.
The U.S. does not have a law focused on algorithms. Certain government agencies instead use their existing authority to try to protect the public. For example, the Federal Trade Commission (FTC) can police “unfair or deceptive practices” that might be related to misleading claims about the capabilities of AI, or how use of AI can lead to discriminatory outcomes.
There have also been some attempts to legislate directly on AI. For example, a bill called the Algorithmic Accountability Act has been introduced in Congress which would mandate developers to audit AI systems for bias and privacy risks. But to date, this legislation has not been adopted. Finally, there have also been voluntary guidelines proposed, for example, by the National Institute of Standards and Technology (NIST) and its proposed AI Risk Management System which would have developers adopt best practices for managing risk and promoting transparency in their algorithms.
数据保护和隐私
人工智能与个人数据相辅相成。 中国已经制定了一套全面的数据保护法规，这对于应对人工智能时代的数据监管需求亦大有裨益。数据监管法律框架包括《网络安全法》（要求网络运营商必须在中国境内存储某些数据）、《数据安全法》（规定应如何保护与国家安全相关的数据）以及《个人信息保护法》（与欧洲的 GDPR 相似，规定了获得用户同意和对使用个人数据的各类限制）。在使用数据开发或应用人工智能技术时，企业必须遵守所有适用的数据保护法规。特别是，提供人工智能工具用于编辑个人生物特征信息（如人脸或人声）的企业必须提醒用户通知相关个人并征得其单独的同意。
相比之下，美国并没有一套全面的数据保护法规。只有一些法律涉及特定行业应如何处理不同类型的数据。 例如，某些医疗数据受《健康保险便携性和责任法案》（HIPAA）的监管， 某些财务记录数据受《格拉姆-里奇-比利雷法案》（GLBA）管辖。
除此之外，各州也纷纷介入这一空白领域——《加州消费者隐私法》（CCPA）就人工智能系统如何收集和管理个人数据为加州居民提供了保护，而弗吉尼亚州和科罗拉多州也有自己独特的数据隐私法要求。
Data Protection and Privacy
AI and personal data go hand in hand. China has developed a comprehensive set of regulations governing data, and this has been useful to adapt to the need for oversight of AI as well. The data regulation framework consists of the Cybersecurity Law, which mandates network operators to store certain data within China, the Data Security Law which lays out how data related to national security should be protected, and the Personal Information Protection Law, which is similar to Europe’s GDPR and sets forth the need for obtaining user consent and limiting use of individual’s data. When using data to develop or apply AI technologies, businesses must comply with all applicable data protection regulations. Businesses that offer AI-powered tools to edit biometric information of a person (such as face or voice) must remind users to notify and seek separate consent from that person.
The U.S., by contrast, does not have an overarching set of data regulations. There is merely a mix of laws which govern how different types of data should be handled by industry. For example, certain medical data is regulated by the Health Insurance Portability and Accountability Act (HIPAA). Certain financial records data is governed by the Gramm-Leach-Bliley Act (GLBA).
States have also stepped into the void – the California Consumer Privacy Act (CCPA) provides some protection to California residents on how AI systems can collect and manage personal data. Virgina and Colorado have their own unique data privacy laws and requirements as well.
版权
中国和美国的一个分歧点在于，使用人工智能生成的作品是否可以获得版权保护。 美国的一贯立场是版权是针对原创作品的。 法院和美国版权局对 "作者 "的解释是指人类创作者。 这可以追溯到 Naruto 诉 Slater 案，该案认定非人类（猴子）创作的作品不能获得版权保护。 法院和美国版权局据此认为，没有人类创作者的作品（如人工智能生成的内容）不能受到版权保护。
最近颁布的一个决定是 2023 年的 "黎明的扎里亚"案。美国版权局在该决定中最终认定，在一部包含使用人工智能图像生成器生成图片的图文小说中，应区分与图片一起的文字编排（可受版权保护）和图片本身（不可受版权保护）。值得注意的是，对人工智能最初生成的内容进行实质性的编辑也可能使内容变成可以获得版权保护的作品。
在美国，关于人工智能生成内容的可版权性已基本达成共识。但在中国，对于法院是否可以对人工智能生成的内容给予版权保护仍然存在很大争议 ，尤其是当原告声称自己曾为诉争的人工智能生成内容付出了巨大努力时更是如此。一方面，中国没有像美国版权局那样的权威机构有权责就某些作品的可版权性发表权威意见；另一方面，中国法院在裁定人工智能生成的具体作品是否具有可版权性时，似乎 在个案中采取不同的观点。少数法院的判决认为，当人工智能生成的图片符合中国《著作权法》关于作品的法定标准，其有权获得版权保护，即输入相当复杂的提示词并使用人工智能工具生成、修改和润色图片的过程可被视为人类的智力活动，而人工智能软件相当于协助人类作者创作作品的工具（如照相机）。然而，这些案例对未来审判的影响可能非常有限，因为中国并不是判例法国家；此外，一些判决未经上级法院审查即成为有约束力的判决，而法院有争议的说理也引发大量的批评。
Copyrights
One point where China and the U.S. differ is on whether works generating using AI can be copyrighted. The U.S. has consistently taken the position that copyrights are for original works of authorship. Authorship has been interpreted by courts and the U.S. Copyright Office to mean having a human creator. This traces back to the Naruto v. Slater case in which it was determined that a work made by a non-human (monkey) cannot be copyrighted. This has been extrapolated by courts and the U.S. Copyright Office to mean works without meaningful human authorship, such as AI generated content, are not able to enjoy copyright protection.
One relatively recent decision was “Zarya of the Dawn” in 2023 in which the Copyright Office ultimately determined that a graphic novel which had contained images using an AI image generator distinguished between the text arrangement that went along with the images (which was copyrightable) and the images themselves where were not. Notably though, substantial editing of content originally generated by AI may turn the content into a work that can be copyrighted.
Unlike in the US where consensus has almost been reached on the copyrightability of AI generated content, in China it remains very controversial when it comes to whether the courts may grant copyright protection to content generated by AI. This is especially true when a plaintiff claims it has put substantial efforts to cause the creation of the AI generated content in dispute. On the one hand, there is no authority in China like the U.S. Copyright Office that has power and responsibility to issue its authoritative opinions on the copyrightability of certain works; on the other hand, Chinese courts seem to take different views in different cases when deciding whether a specific work generated by AI is copyrightable. A few court judgements recognized the AI generated images at issue are entitled to copyright protection because they meet the statutory standards of a work under China Copyright Law—the process of inputting quite complicated prompts and using AI tools to generate, modify and polish the images can be regarded as human’s intellectual activities while AI software is equivalent to a tool (e.g. a camera) which assists human authors to create works. However, these cases are likely to have very limited influence on future ones, as China does not follow a case law system; moreover, some judgements became final without review by higher courts, and the courts' controversial reasoning has drawn significant criticism.
执法和处罚
有些人认为，美国采取的较为宽松立法态度对于促进人工智能创新发展非常重要，但也有人可能会说在国家层面制定一套全面的人工智能管理条例，将为从事人工智能开发和使用的企业提供所需的清晰方向。
实际上，在美国和中国，均有多个政府机构参与监管，特别是在执法领域。 就中国而言，执法机构包括CAC和公安部等。《个人信息保护法》规定违法行为最高可被处 5000 万人民币（约合 700 万美元）或企业上一年收入的 5%的罚款。《数据安全法》也规定了处罚措施，违法行为可能导致企业暂时或永久停业。
如预期，在美国执法工作由州和联邦两级共同实施。 FTC可以对从事不公平或欺骗性行为的公司处以罚款；此外，还可以执行同意令，对企业使用数据的方式进行长期限制。此外，美国食品和药物管理局（FDA）对人工智能驱动的医疗设备执行特殊规定，开发者必须通过上市前提交或 De Novo 分类等途径证明其有效性和安全性。 美国证券交易委员会（SEC）和美国商品期货交易委员会（CFTC）也对机器人咨询和人工智能在交易中的应用进行监管。
各州通常都有自己的执法方式，例如伊利诺伊州的《生物识别信息隐私法》（BIPA）允许通过私人诉讼来监督人工智能数据政策。
Enforcement and Penalties
While some might argue the more relaxed approach the U.S. takes is important to foster innovation using AI, one might also argue that having a comprehensive set of regulations governing AI at a national level would provide needed clarity to businesses engaged in developing and using AI.
In reality, in both the U.S. and China, multiple government agencies are involved in regulating the space and in particular, enforcement. In China, the mix of enforcement agencies includes the CAC and the Ministry of Public Security, among others. Under the Personal Information Protection Law, violations can include up to RMB 50 million (approx.. US$7 million) or 5% of the previous year’s revenue. The Data Security Law can also include penalties, and may lead to businesses being forced to temporarily or permanently shut down.
Enforcement in the U.S. is, as expected, a combination of efforts at the state and federal levels. The FTC can impose fines against companies found to be engaging in unfair or deceptive practices. It can also enforce consent orders that place long-term restrictions on how organizations use data. The Food and Drug Administration (FDA) enforces special rules for AI powered medial devices, for example. Developers must demonstrate efficacy and safety through approval pathways such as premarket submissions or De Novo classifications. The Securities and Exchange Commission (SEC) and the Commodity Futuers Trading Commission (CFTC) also watch over robo advisory and AI use in trading.
States often have their own enforcement, for example, the Illinois Biometric Information Privacy Act (BIPA), which allows private lawsuits to police AI data policies.
展望未来
中国和美国的人工智能立法都在不断发展。例如，在中国和美国包括中国各省市和美国各州都致力于通过生物识别隐私法（如人脸识别）来解决与人工智能相关的数据收集问题。
预计还会有更多关于算法公平及消费者保护的规定即将出台。中国监管机构在这方面一直走在前面，美国FTC也对涉嫌歧视的人工智能提出了明确警告。 预计公平性审计和透明报告最终将成为常态。
显然，尽管中国和美国有各自不同的法律框架，但都认识到人工智能要求对个人数据以及算法的使用进行强有力的监督。 随着时间的推移，如果两种立法路径开始出现趋同，亦不足为奇。
Going Forward
Rules on AI for both China and the US continue to evolve. For example, in both China and the U.S., Chinese cities and provinces and U.S. states are adopting biometric privacy laws (for example, on facial recognition) to address AI-related data collection.
More regulation on fairness of algorithms vis-à-vis consumers is also expected. Chinese regulators have been out ahead on this front, and the US FTC has also made prominent warnings against biased AI. Expect fairness audits and transparent reporting to eventually become the norm.
It is becoming clear that even though China and the United States have their own distinctive legal structures, both recognize that AI calls for robust oversight for use of personal data and algorithms in particular. It would not be surprising if in fact, the two approaches even started to converge a bit over time.