This article discusses the governance of artificial intelligence ethics in the field of international law, and does not involve arguments and explanations in the field of comparative law. Therefore, this article adopts the general definition of international soft law - although it lacks the legal binding force of international law, it can produce practical effects in international practice and contains standards, norms, principles or other behavioral rules in the content. The origins of international soft law in the ethics of artificial intelligence in practice mainly come from international soft law documents issued by UNESCO () (Intergovernmental International Organizations) and formulated by other international organizations (non-intergovernmental International Organizations).

In order to promote all parties to abide by the ethical principles of artificial intelligence, UNESCO has actively formulated international soft law documents since 2017. The documents formulated in just four years have guided artificial intelligence technology toward a "responsible" direction; while the international soft law documents formulated by other international organizations can be traced back to 2005 (Table 1). Overall, concerns about human well-being (“benefiting mankind”), transparency and accountability are essential ethical elements for most international soft law documents at present. These documents focus on human dignity, as an important value of artificial intelligence ethics, and the human rights concept of concern for human well-being (“benefiting mankind”) runs throughout the documents.

Table 1 Specific provisions on the ethical principles of artificial intelligence in international soft law documents

Table 1 on of in soft law

Among them, the "Recommended on the of" clearly states that with the development of technology keeping pace with the times, the definition of artificial intelligence ethics does not make the only explanation; the main reason is that the current definition needs to be improved with the change of technology, and it is difficult to set the precise requirement of artificial intelligence ethics, and the definition of artificial intelligence ethics can only be unclear. Therefore, the traditional "hard method" is not suitable for jurisdiction due to the accuracy and fixedness of the content of the rules. UNESCO has successively issued "soft laws" such as ethical reports and suggestions to curb the ethical issues of artificial intelligence, becoming a sample of other international organizations' formulation of relevant international soft law documents, and none of the documents have precise definitions of artificial intelligence ethics. In addition, the values of Western eschatology, superhuman culture are different from the values of Chinese symbiosis and humanistic culture. Even if all parties expect to embed unified judgment standards for universal values at the international legal level, it is difficult to fully reach an international consensus on the ethical principles of artificial intelligence when building an international hard law system. Therefore, the governance of artificial intelligence ethics in the field of international law urgently needs conceptual innovation and communication and coordination, and the rise of international soft law is almost inevitable.

The advantages of current international soft law governance

The article "The Global Pattern of Artificial Intelligence Ethical Code" points out that "private companies, research institutions and public sector organizations have issued principles and guidelines for artificial intelligence ethics, and have a considerable overlap in the global convergence of five ethical principles (non-malice, transparency, responsibility, fairness and justice and privacy). Although the global consensus is generally reached on the ethical issues of artificial intelligence that can be solved through technology, many ethical issues that cannot be solved by technology (such as gender gaps in the artificial intelligence industry, sustainability issues, dual-purpose issues, and labor displacement, etc.) are often seriously overlooked. Artificial intelligence has some characteristics that are different from other technologies, and the risks caused by it have exceeded the regulatory authorities' judgment of traditional security, environmental and health risks. It is difficult for a government to comprehensively govern ethical problems in the development of artificial intelligence using traditional regulatory methods. In addition, due to the rapid development of artificial intelligence itself, when it involves cross-industry, government agencies, jurisdictions and stakeholder groups, when using hard methods to coordinate and supervise corresponding ethical issues across subjects, it is difficult to have joint law enforcement due to high law enforcement costs and large differences in opinions of various subjects. Soft laws are easy to apply flexibly to coordinate the interests of all parties at this time.

Similarly, compared with international hard law, current international soft law also has advantages in the governance of artificial intelligence ethics.

International soft law has important advantages, is more flexible and efficient, and has low application cost. General soft laws can be revised and adopted relatively quickly without going through traditional government rule-making procedures. Sometimes, because they are not bound by the legislative body, several different soft laws can even be implemented simultaneously, and even if there will be different professional standards or surges in soft laws, cooperation rather than confrontational relationships can be established among stakeholders. Because in an era of fierce international competition, most governments (such as the United States, France, and Japan) are reluctant to hinder emerging technological innovation through preemptive supervision. Governments will pass many domestic soft laws such as suggestions, guidelines, codes of conduct, best practices, professional standards, etc. to try to solve the ethical problems of artificial intelligence at the realistic level. This is especially true at the international level. NGOs (such as the International Organization for Standardization) can flexibly and efficiently issue international soft law documents because they do not involve law enforcement mechanisms. The overall operating costs of the formulation and implementation of these documents (international soft law) are lower than those of international hard law.

International soft law can supplement or improve hard law by filling formal legal gaps as a primary alternative or governance tool. International soft law is not subject to the authorization of limited institutions and therefore can solve any problems arising from technology. And since it is not adopted by a formal legal body, it is not limited to a specific legal jurisdiction, but may have international applicability. For example, the Data Charter (Data) issued by the Institute of International Finance (IIF) will help supplement (rather than replace) national and regional legislation and regulations (domestic level) and international standards (international level). With the use of customer data and algorithms, although the "Data Ethics Charter" does not have a mandatory binding or legal status, it does not recommend that governments and their financial regulatory departments "regulate" data ethics principles be implemented, but the IIF in the document invites all stakeholders to incorporate data ethics into financial practice activities to fill the legal gap in the financial services field.

(3) International soft law is conducive to the distinction and governance of artificial intelligence ethics and layered response. At present, the development of artificial intelligence technology is still in the stage of weak artificial intelligence, and most uses do not pose high risks, and we do not need to face similar ethical problems such as the "tram problem" of autonomous vehicles. For low-risk AI applications, soft laws developed under the opinions of companies, civil society groups, academic experts and governments will be a key approach to promoting innovation in line with ethical frameworks and principles. If soft methods can hierarchically regulate artificial intelligence ethics, they can also become an important supplement to hard methods for high-risk applications of artificial intelligence. For example, the Der Weg der report issued by the German Data Ethics Committee (Der Weg der) divides ethical considerations into concerns about data or algorithm systems, and determines five levels of criticality of algorithm systems: applications with zero potential hazards or negligible potential hazards will not be regulated; as potential hazards increase, regulatory burden will increase until they are completely prohibited. For applications with serious potential hazards, the report recommends ongoing supervision.

Challenges of the International Rule of Soft Law in the Ethics of Artificial Intelligence

Cooperation among international soft law subjects is unstable

Given the widespread application of artificial intelligence and the uncertainty caused by its rapid future development, so far, non-state actors (such as IEEE, FLI, OECD, etc.) have no proposals and coordination of international hard methods. When countries choose to allocate benefits that are beneficial to them without relying on international hard laws with compulsory effect to ensure the continued cooperation, non-state actors cannot guarantee the credibility of commitments made by countries through the help of international soft laws alone. Therefore, cooperation among countries is not stable enough, and this can be explained based on the "prisoner's dilemma" game (collaborative game).

Transnational digital trade for digital products is a "prisoner's dilemma" game: when implementing digital trade measures domestically, countries citing the "security exception" clauses in the World Trade Organization (WTO) or the Free Trade Agreement (FTA) rules are most likely to meet their domestic data ethical requirements; but at the international level, this means of safeguarding the interests of various countries are not conducive to the development of the global digital economic order. For example, the United States has used security exceptions to suppress Huawei's sales in global markets. If all countries abuse the "security exception" clause to prohibit digital products from entering their respective domestic markets (i.e. disguised trade protectionism), it will seriously hinder the development of transnational digital trade, which is the worst collective outcome; but if all countries moderately open their own digital trade markets, they can achieve collective optimal results. In fact, for each country, the optimal strategy is still to open up digital trade markets to other countries, while the country implements disguised trade protectionism for other countries. When individual rationality is contrary to collective rationality, it is easy for countries to violate relevant digital trade policy commitments in WTO or FTA rules, even if this is very likely to meet the domestic data ethics (including artificial intelligence ethics), so international hard methods are urgently needed to suppress the possible betrayal strategies of countries. Although international hard law with compulsory effect was needed in the past, in the current mutual game of digital economy development among countries, how to give full play to the unique advantages of international soft law, and complement the possible international hard law in the future, and promote stable cooperation, is a challenge facing the international soft law governance of artificial intelligence ethics.

International soft law sometimes cannot be effectively implemented

International soft law is sometimes not effectively implemented, which is reflected in the field of artificial intelligence ethics. What is quite controversial is the accountability system involving legislative issues. The accountability system of artificial intelligence refers to the establishment of specific legal systems to explain why and what approaches can make the deployers and designers of artificial intelligence systems accountable. When the artificial intelligence system has certain autonomy in behavior and decision-making, it becomes increasingly difficult and complex to clarify the responsible subjects, and it is difficult to divide responsibilities by establishing a multi-layered responsibility system for various behavioral subjects (such as designers, manufacturers, operators and end users). Even though accountability has been stipulated and described in the transnational ethical normative model, these soft laws lack formal and specific responsibility division systems and may not be effectively implemented in reality.

As the leading body for formulating the strategic guidelines and roadmap of the United Nations system-wide artificial intelligence governance body, the International Telecommunication Union (ITU) coordinates the participation of various United Nations agencies. Although it works closely with all relevant United Nations entities and, as appropriate, with external partners, it can only serve as an institutional coordination center for “promoting” coordinated implementation, with limited implementation capacity. In addition, Article 42 of the "Recommendation on Ethical Issues in Artificial Intelligence" concerns "appropriate supervision, impact assessment, audit and due diligence mechanisms should be established"; among them, the wording of "appropriate" itself does not have compulsory enforcement, and its implementation is mainly guaranteed by the attractiveness of the article itself and the commonality of the member states' orientation to value and interests. Relying solely on the above recommended and recommended soft measures, it is inevitable that soft laws sometimes fail to achieve effective implementation results.

Countermeasures of the international soft law governing the ethics of artificial intelligence

In the field of artificial intelligence ethics where international soft law is developed and hard law is backward, in order to maintain the stability of international cooperation and promote the enforceability of soft law content, it is of practical significance to construct a "center-peripheral" model of international soft and hard law hybrid governance and build an indirect execution mechanism.

Constructing the "center-peripheral" model of international mixed governance of soft and hard methods

Proposed of the "center-peripheral" model

In the transnational legal system, international hard law ranks at the center of the system structure and together with international soft law defined as peripheral, forming a "center-peripheral" relationship at the international level. In the field of artificial intelligence ethics where international soft law is developed and hard law is backward, international soft law has the advantages of governance functions and de facto effect when establishing mechanisms for the expression, compensation and coordination of interests of all parties, but this does not mean that it should be at a central position because of its necessity. Because soft law in the form of international norms and professional standards can exert influence without causing extraterritorial legal issues and is not subject to government jurisdiction for promulgation of laws and regulations. Market-oriented incentives for enterprises to formulate professional standards can enable global market consumers to test the "good or bad" of standards. This kind of multi-subject participation in governance at the international level has a synergy, which can enhance the collaboration awareness of public-private governance subjects, thereby forming a "center-peripheral" model of mixed governance of international soft and hard methods (Figure 1).

Figure 1 "Central-Operipheral" model of international soft and hard-to-hard governance

1 "-" model of soft-hard mixed law

In most cases, the advantageous areas of international soft law are limited to transactional international relations, rather than basic international relations. This is determined by the strong technical and complex and changeable characteristics of global governance in artificial intelligence ethics. Compared with international hard law (traditional international law), international soft law is more flexible and convenient to be implemented. Because traditional international law regulates the "coexistence" relationship between countries such as diplomacy, territoriality, and recognition, the basic part of the "cooperative" relationship between countries related to artificial intelligence ethics that emerged later still relies on strong stability and reliable international hard law support. For example, important international organizations such as the G20 and OECD have promoted all parties to reach agreement on important concepts such as "responsible artificial intelligence" and "trusted artificial intelligence" that include human rights. The ethical consensus involved must accept the implementation of the Universal Declaration of Human Rights as the minimum order.

Therefore, international soft law sometimes only reflects transitional forms. Once the time is ripe, international hard law can still be put on the agenda for the regulation of artificial intelligence ethics. For example, the ethics of autonomous driving cars in artificial intelligence applications need to clarify the responsibility according to the different risk levels of traffic accidents that occur, and finally determine the damage caused by the decision-making and selection of autonomous driving cars, so that vehicle developers, producers, sellers, and owners can abide by it in accordance with the provisions of international hard laws with ethical principles. Because the time for self-driving vehicle products is similar or longer than the time for international hard-process rules, it is suitable for the hard-process fields of specific technologies where environmental changes are not too fast.

The essence of the "center-peripheral" model

In the field of governance of artificial intelligence ethics, the essence of the "center-peripheral" model composed of international hard law and international soft law includes four aspects.

The penetration of international soft law into domestic hard law, especially the transformation of technical standards through domestic law, is binding. For example, the "Private Information Management Requirements and Guidelines" (—and, ISO/) formulated by the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC) are linked to the "Code of Conduct and Certification Part" standard of Chapter 4, Section 5 of the EU General Data Protection Regulation (Data), and are transformed through the EU statutory law General Data Protection Regulation and promote mutual recognition of standards. The written law stipulates the conditions and specific methods for embedding data protection specifications, which can further clearly confirm the legal effect of technical standards.

Usually, governments that are interested in developing artificial intelligence technology will encourage enterprises, user groups, non-governmental organizations, scientific research institutions and other diversified regulatory entities to actively participate in the discussion of artificial intelligence ethics by relevant industry associations, and promote the mutual influence and mutual transformation of domestic soft and hard methods. For example, the EU has successively issued guiding documents "Guidelines for AI" and "A for and" and "A for and", which tend to stipulate mixed clauses of artificial intelligence ethics, that is, to draw on the open coordination procedures of soft law to regularly revise the clauses, and by blurring the boundaries between hard law and soft law, the dual means of soft law and hard law complement each other's advantages in practice and jointly promote the process of building artificial intelligence ethics.

Domestic enterprises, academic circles and intergovernmental international organizations in various countries can issue various documents on specific uses of artificial intelligence, in order to penetrate the domestic technology governance preferences of various countries into international soft law and strive for the right to speak in the formulation of international standards. Professional standards can take years to develop, so they are not always more speed advantage than hard methods. However, compared to hard methods, professional standards are usually formulated specifically to promote innovation within the industry and can be updated relatively quickly as technology develops. For example, IEEE is developing AI standards that include “transparency in autonomous systems” and “certification methods to address transparency, accountability, and algorithmic bias”, as well as the impact of AI on human well-being; or ISO, like IEC, is working with IEC to develop a series of standards that focus on AI. Once these professional standards are created, multi-stakeholders of the digital industry or existing private institutions can be disseminated and implemented globally, gradually affecting and even leading the formulation of international standards.

Even if the influence of international soft law is evenly matched with international hard law, it will not affect the "center-peripheral" model of mixed governance of international soft and hard law from the overall perspective. International hard law is legally binding and usually has a greater influence in practice than international soft law with only factual effect. The international hard laws jointly formulated by various countries in the future will be more stable and have a solid foundation for international cooperation than the international soft laws formulated by other non-state actors. For example, autonomous vehicle technology that is prone to high risks is concerned about human welfare (such as the right to life and health, etc.), even if international soft law has the advantages of low legislative costs, it is still necessary to determine its basic ethical norms in international hard law. The cost of hard law in the implementation stage is likely to be lower than that of soft law.

In summary, when enterprises and industry associations propose artificial intelligence ethical standards, we can refer to the specifications of technical standards on the same topics, especially universal technical issues such as transparency, accountability, and concern about human welfare. For example, the establishment of transparency principles in artificial intelligence ethics can adopt the "center-peripheral" model of international mixed governance of soft and hard methods: using hard method (law jointly formulated by the state) to solve higher-level and universal prominent data/algorithm/computing power/knowledge problems, and implementing stronger bottom-line constraints can improve the reliability of artificial intelligence; at the same time, using soft method (rules formulated by non-state actors) to solve personalized problems in specific application scenarios, allowing experts closer to technology than legislatures to develop a soft law framework, and providing resilience specifications with a certain fault tolerance space to guide the development of technology.

Build an indirect execution mechanism

Establish an international governance coordination committee

And proposed a coordination entity from an international perspective, which can be called the "Governance Coordination Committee" (). If an international governance coordination committee is established in the field of AI ethics, it is “located outside the government but includes participation from government representatives, industries, NGOs, think tanks and other stakeholders” and “the entity does not seek to replicate or replace many organizations dedicated to developing AI governance methods, but provides a coordination function…to ensure that all different participants connect with each other, understand and respond to each other’s proposals while identifying gaps and inconsistencies in existing plans”. Therefore, the participation of multiple subjects in the International Governance Coordination Committee is conducive to its credibility in defining ethical norms and can promote the "hardening" of soft law to enhance the binding force and execution possibility of soft law.

The Inter-Agency Working Group on Artificial Intelligence (IAWG-AI), jointly led by UNESCO and the International Telecommunications Union, seeks to cooperate with relevant inter-agency and multi-stakeholder mechanisms on AI-related issues by supporting the work of the Chief Executives’ Coordination Council and the High-Level Committee on Programmes on AI ethics. The working group focuses on the ethical field and can establish an international governance coordination committee on it, which can not only reduce the cost of establishing supervisory bodies but also increase international influence. When regulating the ethics of artificial intelligence, various types of soft laws (including ethical standards, guidelines, codes of conduct and principles, etc.), the International Governance Coordination Committee may rely on punishment means, social public opinion supervision and technical evaluation of technical experts to ensure implementation, indirectly affect and guide relevant responsible parties to assume their due responsibilities (Figure 2).

Figure 2 United Nations artificial intelligence governance working mechanism and the idea of building an indirect implementation mechanism

2 for AI and of

The establishment of an International Governance Coordination Committee will not only have the expert resources to maintain "technologically neutral" in the inter-Artificial Intelligence Working Group, but also improve the governance efficiency of matters under the CEB and the High-Level Committee on Programmes. As a non-state actor, the international soft law formulated by the International Governance Coordination Committee cannot be supported by the state coercive forces alone (mainly relying on the power of the sovereign state itself) or collective (in some cases the collective power of the international community) like international hard law, but it can be supplemented by informal and decentralized indirect implementation mechanisms to achieve corresponding results.

Informal and decentralized indirect implementation mechanisms

Under the framework of the transnational legal system, international soft law as a governance tool does not have legal binding or enforceable effect, but has the effect of implementation. Although many non-state actors formulated international soft law do not have the enforcement force to ensure implementation, and cannot retaliate against violators like traditional international law, etc., it must not be denied that international soft law has the function of self-regulation and self-discipline, and the role of external forces can still enable it to be followed in fact. According to game theory, usually relying on the "self-implementation" of various soft law subjects does not require a third-party mandatory force guarantee and centralized authority, supplemented by informal and decentralized execution mechanisms, such as punishments of stakeholders, social pressure and market forces, to promote the de facto effect of international soft law.

Punishment from stakeholders increases the cost of violations. Generally, only those who are interested will follow the soft law by deterrent based on the deterrence of hard method tools. However, non-state actors punish those who do not comply with soft laws, increasing the potential cost of violating soft laws, and can also form positive deterrence in advance and achieve the purpose of prevention. For example, a country violates the "Guidelines for Foreign Direct Investment Treatment" formulated by the World Bank and is not only unable to obtain guarantees and loans from the World Bank, but will also be declared by this authoritative international economic organization as a country with a poor investment environment, which seriously affects the country's ability to absorb foreign investment in the future. Suggestion: The International Governance Coordination Committee sets up penalties to increase the cost of violating soft law, and indirectly improves the enforcement of international soft law.

Social pressure promotes reputation mechanisms. Subjects that violate international soft law are extremely vulnerable to pressure from the international community. Condemnation by multiple stakeholders such as states, transnational non-governmental organizations, intergovernmental organizations will discredit the reputation of violators and lose all their faces. Some international soft laws are nominal soft laws and substantial hard laws, which greatly damage the economic or commercial interests of the offenders. Their punishment and sanctions are no less effective than international arbitration awards or judicial judgments that are unfavorable to them. This "humiliation of group growth" indirectly prompts countries to abide by international soft law, safeguard their credibility in the international community, and avoid the adverse impact of losing the trust of international partners due to the failure to enforce international soft law. Suggestion: The International Governance Coordination Committee actively discusses and promotes action initiatives or ethical rules for multilateral joint supervision, and explores consensus and differences in artificial intelligence ethics. The reputation mechanism revolves around the implementation of initiatives or rules by non-state actors, requiring punishment for countries that undermine the initiatives or rules and commends countries that maintain the initiatives or rules, so that countries will assume international social responsibilities under pressure in order to maintain the good reputation of the international community.

Market forces have spawned a "network effect". When countries that accept unified standards or standards aggregation, and countries that use the standards or standards increase, the value of standards to them also increases accordingly, thereby attracting more countries to spontaneously aggregate to form a "snowball effect", continuously expand the group size that complies with soft laws, and enhance the binding force of soft laws. Countries that do not accept will be marginalized, and countries that accept will benefit more and more, thereby promoting countries to accept these international soft laws (unified norms or standards). For example, ISO has formulated international standards for nanotechnology risk management as a quasi-regulatory standard for safe handling of nanomaterials. Although these standards cannot be enforced directly, many international contracts and some insurance companies require compliance with applicable ISO standards (or their equivalents), thereby providing an indirect enforcement mechanism. For example, the International Governance Coordination Committee learns the guidelines formulated by the International Stem Cell Research Society to develop research guidelines to limit certain types of (anti-ethical) research, and provides ethical guarantees for other types of research. Although not directly implemented, these guidelines set professional expectations for stem cell researchers and can be implemented indirectly by research institutions, funding agencies, and scientific journals that require scientists to adhere to. Suggestions: Establish a diversified participation path, promote multi-level international dialogues such as enterprises, industry associations, experts and scholars, and use scientific journals to share best practices to support standardization work; the coordinated interaction, exchange and collision of artificial intelligence ethical rules in different countries under different cultural backgrounds will gradually form unwritten guidelines or standards (i.e., international practices) in international exchanges, thereby prompting countries to abide by the conventional artificial intelligence ethical norms in international practice.

(Authors: Zhu Mingting, Xu Chongli, School of Law, Xiamen University. Contributed by "Proceedings of the Chinese Academy of Sciences")