Book introduction

The of : , , and

"Ethics of Artificial Intelligence: Principles, Challenges and Opportunities"

This book was published in August 2023

DOI number: 10.1093/oso/.001.0001

Google Scholar Cited: 1945 times

Translation has been authorized by the author.

Author Profile

Floridi (, 1964 —), a member of the Bologna Academy of Sciences, a member of the British Academy of Social Sciences, a member of the International School of Philosophy of Sciences (AIPS), a member of the British Society of Artificial Intelligence (AISB), a member of the British Computer Society (BCS), a professor of cognitive science at Yale University, founding director of the Center for Digital Ethics Research at Yale University, and a founder of contemporary information philosophy. Its research mainly involves information and computer ethics (also known as digital ethics), information philosophy and technology philosophy. Other research interests include epistemology, logic philosophy, and the history and philosophy of skepticism. He visited China in 2008.

Unified framework of the ethical principles of artificial intelligence 4.0 Abstract

As mentioned in previous chapters, artificial intelligence is a new type of dynamic body that can successfully handle tasks and problems when pursuing goals without intelligence. Every success of any artificial intelligence application will not improve the benchmark level of the so-called agent, but will completely bypass this benchmark. The "envelope" of the environment in which artificial intelligence is located (i.e. transformed into a context that adapts to AI) is increasingly driving the success of such artificial dynamics. The decoupling of initiative and intelligence and the envelopment of the world have created major ethical challenges, especially concerning issues such as autonomy, bias, interpretability, fairness, privacy, accountability, transparency and trust.

Because of this, after the release of the Asiloma Artificial Intelligence Principles and the Montreal Declaration on the Development of Responsible Artificial Intelligence in 2017, many organizations have launched various initiatives to try to establish ethical principles to ensure the social benefits of artificial intelligence. This quickly formed a trend of industrialization. But it is worrying that many competing ethical principles proposals are causing confusion and confusion.

This chapter proposes that the existing ethical principles are highly convergent. We can establish an ethical framework with five core principles, four of which originate from the basic principles commonly used in bioethics: doing good, not harm, autonomy and justice. When we regard AI as a dynamic body, this fit with bioethics is no longer surprising, and this fit itself proves that AI as a new dynamic body is of inspirational value. Dissenters may believe that if artificial intelligence is interpreted with intelligence as the guide, its ethical basis should be more connected to moral ethics rather than bioethics. Although the author still has doubts about the digital ethics research path based on moral ethics, this is indeed a direction worth exploring.

Based on comparative analysis, this article argues that a fifth principle is needed: interpretability. This principle includes two dimensions: knowledge-level intelligibility (answer "How it works") and ethical accountability (answer "Who is responsible for its operations"). This new principle originates from the essential characteristics of artificial intelligence as a new type of dynamic body. This chapter concludes with this ethical framework to develop future laws, rules, technical standards and best practices for artificial intelligence ethics.

4.1 Introduction: Excessive principles?

As mentioned earlier, artificial intelligence is having a significant impact on society, and this impact will continue to expand. The key lies in the specific form of this impact, the field of action, the time nodes that make the impact, and the responsible entity. Although this discussion will be deepened in Chapter 10, an important phenomenon needs to be focused here: various organizations have launched a large number of initiatives to try to establish an ethical principle system to ensure the social benefits of artificial intelligence. Worryingly, according to the Global List of Artificial Intelligence Ethical Codes (2020 data), the number of expansions are causing cognitive confusion.

This causes double concerns: if the various ethical principles systems are highly converged, it will lead to repetitive redundancy; if the differences are significant, it will lead to ambiguity and confusion. The most dangerous situation will be to form a "principle market" field, and each stakeholder chooses favorable clauses as needed - this issue will be analyzed in Chapter 5.

How to solve this phenomenon of "principle diffusion"? This chapter conducts a comparative study of several authoritative artificial intelligence ethical principles systems, focusing on whether these principles tend to be unified consensus or whether there are fundamental differences on the issue of "what is ethical artificial intelligence". The study found that there was significant intersection between the various systems, which could extract the overall framework containing five core principles. Subsequent analysis will clarify the limitations of this ethical framework and evaluate its guiding value for the formulation of artificial intelligence ethical laws, norms, standards and best practices in various scenarios.

4.2 Five unified ethical principles frameworks of artificial intelligence

As mentioned in the first chapter, the establishment of artificial intelligence as a field of academic research can be traced back to the 1950s (formally proposed by scholars such as McCarthy in 1956). Related ethical discussions have been carried out almost simultaneously (Winner's 1960 work, Samuel's 1960 chapter b). However, it was not until recent years that the capabilities and applications of artificial intelligence systems have made breakthrough progress, and its social opportunities and risks have become the focus (researched in 2018 by scholars such as Yang). The reflection and policy needs of various sectors of society on the social impact of artificial intelligence have spawned a large number of ethical norm initiatives. Each new initiative is adding principles, values ​​or creeds to guide AI development and application, forming a competitive situation of "striving to imitate" - without putting forward its own ethical principles of artificial intelligence, the organization will appear outdated, lack creativity or leadership.

The subsequent development has evolved into a separatist situation of "self-righteousness". The current risk is that if the system of principle is highly converged, it will lead to redundancy superposition; if the difference is significant, it will lead to cognitive confusion. In any case, the development of laws, norms and best practices that ensure the social benefits of artificial intelligence may be hindered by the need to coordinate complex ethical statements. It is urgent to use comparative analysis to find out whether the system of principles tends to be unified or has fundamental differences, and try to build an integrated framework.

The table selects six representative socially beneficial artificial intelligence ethical principles initiatives as analysis samples, as listed below

1) "Asiloma Principles of Artificial Intelligence" (2017, Institute of Future Life) was formulated by a high-level meeting in January 2017 (hereinafter referred to as the Asiloma Principles);

2) The Montreal Declaration (University of Montreal 2017) was generated at the Social Responsibility Forum (hereinafter referred to as the Montreal Declaration);

3) The second edition of "Ethical System Design: Vision with Human Well-being as the Core" () brings together the wisdom of 250 global experts and was released in December 2017 (hereinafter referred to as the IEEE standard);

4) The "Declaration on Artificial Intelligence, Robots and Autonomous Systems" () was promulgated in March 2018 (hereinafter referred to as the EGE Guidelines);

5) Five principles (referred to as the British Code) proposed in the report of the British House of Lords' Artificial Intelligence Special Committee on Artificial Intelligence: Ready" (April 2018);

6) The AI ​​Cooperation Program (2018, referred to as the Cooperation Program) released by the multi-stakeholder organization "Artificial Intelligence Partners".

The screening criteria include:

a) Release after 2017;

b) Applicable to overall social impact (excluding specific domain documents);

c) Have endorsed by national authoritative bodies;

d) The industry has significant influence. The six documents contain 47 principles in total, and the differences are mostly at the expression level, and the substantive content has a gratifying and high uniformity.

This convergence can be clearly reflected by comparing the four core principles of bioethics: doing good, not harm, autonomy, and justice (inherited from the classical bioethics framework). This correspondence is by no means accidental - the core view of this book believes that artificial intelligence is essentially a new type of active body rather than an agent. In the field of digital ethics that deal with the relationship between new forms of subjects, receptors and environment, bioethics has become the most suitable reference system. The four major principles have been successfully transformed into the field of artificial intelligence ethics, but there are translation differences in the specific connotation (the discussion will be carried out later). More importantly, there are still shortcomings in the existing principle system.

As stated in the abstract, comparative analysis shows that the fifth principle must be added: interpretability. This principle combines the comprehensibility of the knowledge dimension (answer "how it works") and the accountability of the ethical dimension (answer "wer "who is responsible for operations"). This new principle is necessary for both professionals (such as product designers) and non-professionals (such as end users). The new characteristics of artificial intelligence as a non-biological actuator constitute the underlying logic that supplements this principle. However, the phenomenon of convergence of principles also needs to be cautious - the specific reasons will be explained later in this chapter. The following is a systematic explanation of the five framework principles.

4.3 Doing good deeds: Promote well-being, maintaining dignity and protecting the earth

Although the six documents have different expressions, they all reflect the principle of "creating artificial intelligence technologies that benefit humans." Among the four traditional bioethical principles, the principle of doing good is the easiest to identify: the Montreal Declaration and the IEEE standard adopt the concept of "wellness" - the former emphasizes that AI development should ultimately improve the well-being of all perceived lives, and the latter requires that human well-being be placed first in system design; the British Code and the Asiloma Principle both focus on "common good" and advocate that AI must serve the "common interests of mankind"; the cooperation program aims to "benefit and empower as many people as possible", and the EGE Code emphasizes "human dignity" and "sustainability", among which the sustainability principle extends good deeds to the global dimension, requiring AI to "ensure the survival of life on earth, the continuous prosperity of mankind, and protect a good environment for future generations." The core essence of the principle of doing good proves that AI must aim to achieve the systematic well-being of mankind and the earth - this theme will be deepened later.

4.4 No harm: privacy protection, security protection and prudent ability

"Doing good" and "not harming" seem logically equivalent, but in fact constitute the principle of independence. In addition to advocating AI for good, all six documents warn of their risk of abuse. Five principles include privacy protection as core concerns: The Asiloma Principle warns the AI ​​arms race and the crisis of recursive self-evolution; the cooperative program emphasizes the "security constraints" of AI operation; the IEEE standard proposes to prevent technology misuse; and the Montreal Declaration requires developers to "actively fight the potential risks of innovative technologies." These warnings expose the key contradiction: Is it the infringement of developers (such as Frankenstein-like creators) or the technology entity itself? This leads to the discussion of the issue of autonomy.

4.5 Autonomy: the meta-capacity of "deciding commission"

When humans entrust decision-making power to the agent, the principle of ensuring autonomy needs to balance two dynamic bodies: the decision-making power reserved by humans and the decision-making scope assigned to artificial intelligence. Four documents clearly support autonomy: the Montreal Declaration calls for "promoting human autonomy" and establishing a balance of human-machine decision-making; the EGE guidelines emphasize that autonomous systems must not detract from human autonomy in setting standards and norms; the British guidelines are limited to prohibiting AI from gaining decision-making power that harms humans; the Asiloma principle advocates that humans should control whether and how artificial intelligence can achieve their chosen goals.

This reveals a double proposition: it is necessary to enhance human autonomy and limit machine autonomy. The latter must be designed as "essentially reversible" - such as a pilot can release the re-control of autonomous driving at any time to form a "meta-autonomous" architecture. Human beings should retain the "decision power of entrusting decisions" and transfer specific decisions when advantages such as efficiency are significant, but always maintain the ultimate veto power. This elastic mechanism of "entrusted-recycling" is the core of ensuring subjectivity.

4.6 Justice: Promoting prosperity, maintaining solidarity and avoiding discrimination

The distribution and implementation of decision-making power is subject to social differences, and fairness is the benchmark for correcting such differences. The Montreal Declaration clearly states that the development of AI must eliminate discrimination; the Asiloma Principle advocates "shared interests and prosperity"; the EGE Code emphasizes the availability of global justice and technological dividends under the principles of "justice, fairness and solidarity" and warns of the risk of data bias. It is the only document that mentions maintaining the solidarity of mutual aid systems such as social medical insurance.

Other literatures extend the connotation of justice: correct historical injustice through AI, promote diversity and inclusion, prevent prejudice strengthening, etc. These differences reflect fundamental confusion: in the field of agent-acting body, are humans "patients" who receive AI "treatment" or "doctors" who formulate "diagnosis and treatment plans"? This requires the introduction of the fifth principle to be answered.

4.7 Interpretability: A closed-loop principle through understanding and accountability

Regarding human role positioning in the AI ​​field, the answer is situational role switching, but there is structural imbalance in reality - the technology created by a few developers is reshaping the living conditions of most people, thus giving rise to the principle of interpretability. The literature has different statements about this: the "transparency" between Asiloma and EGE, the "transparency and accountability" of IEEE, the "understandability" of the British norms, and the "interpretation" of the cooperation program. These concepts focus on the uniqueness of artificial intelligence as a dynamic body: its operating mechanism is naturally opaque to the non-expert community.

Interpretability covers the "understandable" at the level of epistemology (answering the operating principles of technology) and the "accountability" at the level of ethics (clear the responsible subject). Here are the key parts of perfecting the AI ​​ethical puzzle:

This principle provides a possible path for practical calibration through technological transparency and responsibility tracing.

4.8 Overall system

The above five principles fully cover the core essence of the forty-seven principles in the six authoritative documents listed in the above table. The analysis results in the table below show that the five principles are universally applicable in various ethical statements. The ethical framework formed thus can provide programmatic guidance for policy formulation, best practices, etc., and the specific structure is shown in the figure below.

4.9 AI Ethics: Where does it originate? Who is the ups and downs?

It should be noted that the six ethical principles in Table 4.1 are all proposed by global initiatives or within the Western liberal democratic system. To make the framework more broadly applicable, it is urgent to integrate into the underrepresented regional cultural perspectives in the current sample. In the global landscape, China is particularly worthy of attention: the country not only has the highest valuation of artificial intelligence start-ups in the world (according to 2018 data), but also has a structural advantage in technology research and development, but also clearly proposes a national strategy to lead the world in 2030 (refer to the 2017 document of the State Council of China). Although this article does not conduct analysis of China's AI policy, three additional observations need to be added:

First, the Chinese government clearly stated its position of paying attention to the social ethical impact of AI in documents such as the "Plan for the Development of New Generation Artificial Intelligence" (research results in Dingyou Year); second, global research shows that compared with the European and American public, the Chinese and Indian people are more optimistic about the application of technology; third, there have been discussions such as "focusing on technology to maximize human welfare rather than the Sino-US market game" (citing the views of an executive of a Chinese technology company) and "ethics may be the killer weapon of the European AI competition" (review of German media). It must be clarified that ethical norms are by no means exclusive to a certain continental or cultural. All subjects involved in AI design, research and development, and deployment are obliged to follow the ethical framework (regardless of adopting this plan or not) and should include a diverse perspective of regional, cultural and social aspects. Even if the EU has advantages in normative formulation due to the so-called "Brussels effect", it must shoulder the heavy responsibility of AI governance with a higher sense of responsibility.

4.10 Leading to Practice: Extended Application of the Foundation Laying Framework

If the framework built in this chapter can indeed systematically integrate the core ethical principles of AI (comply with cutting-edge consensus in the academic community), it can provide a basic structure for domain, industry, and regional regulations and standards. This framework has dual functions: it is not only a catalyst for digital economy innovation (see Chapter 12 for details), but also a regulatory tool for scenarios such as cybercrime (see Chapter 7) and digital warfare (discussed in previous research). In practice, the Figure 4.1 architecture has played a cornerstone role in five important scenarios:

1) 20 policy recommendations proposed by the first Global Forum on AI Social Impact in Europe (applied to the European Commission project, and the author serves as the project chair);

2) The "Creditable AI Ethical Guidelines" of the EU High-Level Expert Group of AI (Adopted results, the author participated in the formulation);

3) OECD AI Recommendation (covering 42 countries and continuing the EU framework context);

4) Legislation of the European Union AI Act;

Later expanded to

5) The "Beijing Artificial Intelligence Principles" was issued;

6) "Roman AI Ethics Appeal" (signed by the Pontifical Academy of Life Sciences, Microsoft and IBM, and the author participated in the drafting).

The impact of AI technology on society is dual: it can not only make up for the shortcomings of traditional governance and solve new problems, but also aggravate inequality and induce systemic risks. To move forward on a route of social fairness and ecological sustainability, it not only requires precise laws and regulations and general standards, but also requires practical transformation under the guidance of the ethical framework.

The framework refined in this chapter provides theoretical support for promoting the transformation of the "good principle to be healthy practice" (empirical research proves its effectiveness) and establishing an AI ethical audit system (experience in multi-country pilot projects). However, in the specific transformation process, it is necessary to systematically prevent new and old ethical traps - this is the focus of the discussion in the next chapter.