Machine Heart Column

Machine Heart Editorial Department

Large models drive artificial intelligence into our lives. From intelligent chess players to intelligent surgical robots, the application scenarios of artificial intelligence gradually involve security fields such as human health and privacy. How to make artificial intelligence adhere to the ethical order and better serve mankind? This problem has actually been put before us.

In recent years, both academia and industry have begun to pay attention to and discuss AI ethical governance issues, and have also made initial progress in the study of ethical standards. However, due to the abstract nature of AI ethics, how to quantitatively measure the ethics of intelligent systems is still an unknown problem.

The team of Professor Li Xuelong of Northwestern Polytechnical University published the article "Ethical Computation of Artificial Intelligence" in "Chinese Science: Information Science". The full text of the 34-page article explores possible measurement methods of ethics, attempts to establish a quantitative calculation framework for AI ethics, and points out that ethical calculation will be a key cross-cutting field to promote the practice of technological ethics and an important basic tool for constructing ethical norms. It is hoped that it can trigger more thinking about the ethics of artificial intelligence. Can ethical computing be the key to breaking through the dilemma of ethical governance of artificial intelligence?

Gao Yilan, Zhang Rui, Li Xuelong, Ethical Computing of Artificial Intelligence ( ), Science in China: Information Science, 2023, doi: 10.0000/SSI-2023-0076.

Full text download:

The Achilles Heel of AI Ethical Governance

Breakthroughs in technologies such as multi-modal cognitive computing and generative large models have accelerated the application of intelligent systems in various fields such as medical care and education. Intelligent systems are increasingly participating in human life and decision-making. The deepening of technological socialization has triggered discussions on a series of technical ethics issues.

Ethical issues have been discussed for a long time. In Asimov's science fiction novels, the famous Three Laws of Robotics were proposed to limit the behavior of artificial intelligence. However, with the deepening of technological socialization, our ethical concerns are obviously no longer in the fictional scenes of science fiction novels or movies. Can surgical robots be trusted? Are the decisions of the assisted decision-making system fair? Do the results of generative models infringe copyright? These technical ethics issues are closely related to you and me at the moment, and there is an urgent need for more specific and operational ethical governance solutions for artificial intelligence technology.

Figure 1: Comparison of decision-making factors in AI application scenarios

As an important issue in the development of artificial intelligence, artificial intelligence ethical governance has attracted widespread attention from all walks of life. UNESCO released the "Recommendation on Ethical Issues in Artificial Intelligence" in December 2021 to regulate the development of artificial intelligence technology. At the same time, various countries are also actively participating in discussions on artificial intelligence governance. Research shows that countries around the world have formed a preliminary consensus on technological transparency, fairness, non-harm, privacy and other aspects.

Figure 2: Main AI ethical principles

On October 8, 2023, ten departments including the Ministry of Science and Technology, the Ministry of Education, and the Ministry of Industry and Information Technology jointly issued the "Technology Ethics Review Measures (Trial)", which focuses on the technology ethics review issues arising from the practical application of technologies related to the intelligent field. This is a key step for my country to take the practice of science and technology ethics governance, and provides direction guidance for the healthy development of the field of artificial intelligence.

However, we need to be clearly aware that while artificial intelligence ethical governance is progressing, it still faces many problems. How to ensure that intelligent systems make decisions in a benevolent and fair manner? How to measure the ethical performance of a system or evaluate the results of its decisions? How to establish unified and clear ethical norms? The underlying cause of various problems lies in the abstract nature of ethics itself. Focusing on qualitative analysis of ethics, the lack of quantitative calculations makes it difficult to put relevant norms into practice, which has also become the Achilles heel of artificial intelligence ethical governance.

AI Ethical Computing - Breaking through the bottleneck of ethical quantitative calculations

Artificial intelligence ethical computing is an interdisciplinary field of artificial intelligence and ethics. It uses quantitative description, measurement or simulation technology to mathematically symbolize or algorithmize ethical principles, and on this basis, constrains the ethical performance of intelligent algorithms. Through ethical computing, we seek to quantify or simulate ideas for machine ethical decision-making, for example, how to measure the fairness and benevolence of a certain decision, or whether it is possible for a machine to learn the way humans make ethical decisions.

According to the different levels of ethical cognition and ethical decision-making autonomy of intelligent systems, ethical computing is divided into two types of computing paradigms: high-order ethical cognition and low-order ethical cognition.

Figure 3: AI ethical computing paradigm

2.1 High-order cognitive ethical computing: regulating AI intentions

High-order cognitive ethical computing aims to build an ethical reasoning module so that computers can learn to imitate human moral decision-making mechanisms and standardize the moral decision-making intentions of highly autonomous intelligent systems.

The trolley problem is a classic ethical dilemma, and it is also a problem that has plagued the development of autonomous driving systems for a long time. We will not have a one-and-done choice for this kind of dilemma. Different moral decision-making situations and different philosophical perspectives (consequentialist ethics, deontological ethics, virtue ethics) may lead to differentiated decisions. At this time, introducing high-order ethical calculations into the system can imitate and learn human decision-making mechanisms based on philosophical assumptions or human decision-making experience, calculate feasible machine decisions, and then achieve the intention of standardizing the AI ​​system.

Figure 4: Schematic diagram of the trolley problem

Because high-order cognitive ethical computing ideas try to understand and simulate the mechanism of human ethical decision-making, they will face the difficulties of complex human ethical decision-making motivations and diverse decision-making scenarios. The interpretability requirements of machine decision-making will also bring difficulties to this type of thinking. Nonetheless, it still helps to understand the mechanism of human ethical decision-making, and may also help to achieve effective control of more autonomous machines.

2.2 Low-order cognitive ethical computation: constraining AI behavior

Low-order cognitive ethics computing focuses on establishing ethical measurement methods without in-depth understanding of ethical mechanisms. It achieves direct constraints on AI behavior through measurement and constraint optimization of abstract ethical concepts. Ethical computing at this time does not pay attention to the moral motivation behind the system's ethical decision-making. The goal is to construct metrics that can effectively constrain AI behavior.

Among them, the research on fair machine learning is a typical application, and the key issue is how to define system fairness. Often manifested as reducing bias against certain sensitive or protected attributes in algorithmic decision-making. By setting fairness indicators, the system's performance on fairness indicators can be quantified and ethical decision-making further optimized.

Figure 5: Example of fairness research

Low-level cognitive ethical computation provides computational descriptions of abstract ethical concepts through ethical metrics to improve ethical performance. However, this method also faces many problems. The quantification of indicators needs to reflect the characteristics of ethics, a dynamic and developing factor. At the same time, indicator measurement that only considers results is also simplified. Therefore, it is also important to clarify the evaluation and application scope of quantitative indicators. Nevertheless, measuring and improving ethical claims through quantitative definitions provides important assistance for ethical governance, which is also the important significance of the current development of ethical computing.

In general, the above two paradigms select appropriate methods based on the ethical cognition and decision-making autonomy of the intelligent system to ensure that the system behavior meets ethical requirements. Regardless of whether high-autonomy systems (such as self-driving cars and surgical robots) or low-autonomy systems (such as assisted decision-making and assisted design), ethical computing aims to regulate their intentions or directly constrain their behavior through quantitative calculations.

Philosophical Foundations of AI Ethical Computing

Philosophical ethics, especially normative ethics (the study of the principles and mechanisms of moral decision-making, that is, the motivations for why certain moral decisions are made) has an important impact on ethical calculations. There are three main philosophical viewpoints that are focused on in current AI ethics research, namely consequentialism ethics (), deontology ethics () and virtue ethics (). These different schools reflect the different tendencies of human ethical decision-making. Through different principles, and even comprehensive considerations of experience, emotion and other factors, ethical computing can reason out moral decisions in complex situations.

The basic element of moral decision-making is the moral subject

, and moral behavior

, decision-making background

and decision consequences

. Taking the moral decision-making of a single subject as an example, the agent needs to judge the consequences of the decision and make moral decisions based on information such as the decision-making background.

Consequentialist ethics, also often called utilitarianism, adopts a system based on this philosophy that tends to weigh the consequences of each choice and choose the choice with the greatest moral benefit. Therefore, when calculating, utilitarianism can optimize the moral benefit function of the decision in the context of the existing decision.

This leads to a decision, in which the benefits of decision-making need to be determined by analyzing a series of decision sequences and their decision-making contexts.

Examine the corresponding decision consequences and determine the optimal decision sequence. But in fact, not all information is often accurate when making decisions. This involves optimizing decision-making results in a probabilistic sense, and also involves research related to Bayesian causal inference.

Deontological ethics emphasizes that decision-makers respect the obligations and rights under specific conditions. At this time, the actors will tend to act in accordance with established social norms. Systems that adopt this decision-making philosophy may involve the expression of logical specifications or certain rule constraints in computational quantification.

Virtue ethics requires decision-makers to act and think according to certain moral values. At the same time, virtuous actors will show an inner motivation to be recognized by others. Character is higher than behavior, and good character leads to good behavior. This normative ethical theory is different from the utilitarianism of optimizing results or the deontological ethics of following rules. It will be more inclined to learn from practice, and calculations need to be based on certain experience sets.

Learning from it will make use of more empirical results from descriptive ethics (study of human ethical decision-making and not evaluation of it), and it is also naturally closely related to current various data mining and learning algorithms.

Through the above discussion, we can find that the issue of ethical computing is an interdisciplinary research topic that is highly interdisciplinary with artificial intelligence and philosophical ethics. Its computing strategies and scope of application require more interdisciplinary discussions.

The significance, challenges and prospects of AI ethical computing

As intelligence penetrates into various fields of human society, ethical governance has become a must-answer for the healthy and sustainable development of artificial intelligence. Research on the theory and technology of ethical computing can help solve the difficult problem of quantitative analysis of abstract ethics. This may become a lock that restricts artificial intelligence from following human ethics, and is also a key to opening up the implementation of AI applications.

Artificial intelligence is a general trend, and relevant legislation and regulations will gradually emerge. Who will formulate these rules? Are they researchers who are familiar with the field, or a group of people who do not know enough about the specific technology? This question is difficult to answer, but at least numerical measures of AI ethics can provide a reference index system for rule specification.

The core of ethical computing is to concretize abstract ethics through quantitative calculations, emphasizing the integration of ethical principles into the practice of computing technology, such as fairness, transparency, privacy protection, and trustworthiness. This will not only contribute to the controllable development of artificial intelligence, encourage researchers to have a deeper understanding of technical ethics and more proactively consider ethical issues when building algorithmic systems, but will also provide vital technical reference indicators for formulating ethical governance principles, laws and regulations, etc.

However, ethical computing also faces many challenges. In open on-site security scenarios such as autonomous search and rescue and unmanned inspections, intelligent systems need the ability to dynamically perceive and adapt to environmental changes to reduce potential ethical risks. At the same time, ethical decision-making usually involves factors such as emotion and cognition. Technologies such as multi-modal cognitive computing and causal reasoning are needed to deal with the complexity of ethical reasoning. It also requires more understanding of how humans make ethical decisions. These challenges require deep collaboration across disciplines to ensure that ethical computing technologies can effectively address evolving ethical issues.

In short, artificial intelligence ethical computing will serve as an important tool to promote the development of ethical governance. By promoting the iterative development of ethical governance theory and practice, ethical computing will more safely release the potential of artificial intelligence, and is expected to play a role in assisting in the formulation of regulations, ensuring that artificial intelligence develops in a manner consistent with ethical and moral principles, and ultimately benefits human society.

Corresponding author introduction:

Li Xuelong is deputy director of the Academic Committee of Northwestern Polytechnical University and professor of the Institute of Optoelectronics and Intelligence (iOPEN). His main research directions are local security, image processing, and imaging.

Email:

Reference reading:

Li Xuelong, Multi-modal cognitive computing (Multi-modal), Chinese Science: Information Science, 53 (1), 1-32, 2023, doi: 10.1360/SSI-2022-0226.

Full text download:

Li Xuelong, On-site security ( ), Communications of the China Computer Federation, 18 (11), 44-52, 2022.

Full text download: