Fei Song

In this paper, we present an overview of the main conceptual framework for trustworthy AI and argue that the network account offers a superior alternative. We identify and illustrate the possible nodes within an AI network, specify the key attributes each node should possess, and clarify what constitutes a sufficient threshold for each attribute. Building on this foundation, we introduce a novel framework for the network account of trustworthy AI: Lexical Threshold-Based Multidimensional Utility Theory (LTMU). This framework lexically ranks trust-relevant attributes, establishes minimum thresholds for each and assesses each attribute with a distinct utility scale. We then demonstrate how the LTMU framework can be applied to analyse two exceptional cases in which trust-relevant attributes come into conflict in different ways, arguing that its application yields reasonable and defensible outcomes in both scenarios.

Existing approaches to machine ethics harbor an unquestioned commitment to the development of ethical machines and an unreflective optimism that ethical principles can be executable by machines. The first part of this paper raises two challenges to such dogmas: the hypocrisy challenge and the disagreement challenge. The first challenge is that, aside from finding the right machine ethics program, machine ethicists must consider whether their development of such machines is consistent with the precepts of their adopted ethical theory. The second challenge concerns whether the focus on particular ethical theories (especially rationalist and codifiable ones) in the field of machine ethics involves a biased representation of ethical traditions that contradicts commitments to inclusion and pluralism. In the second part of the paper, we offer a possible candidate for machine ethics we call the “skillful-expert model.” This model is based on the assumption that, for machines that have well-defined functions, the requirement of ethical goodness is equivalent to the requirement of skillful performance. We argue that this model has strong support from the neglected, virtue-ethical and phenomenological traditions of ethics, and sidesteps the major challenges raised in the first part.

While most contemporary proposals of ethics for machines draw upon principle-based ethics, a number of recent studies attempt to build machines capable of acting virtuously. This paper discusses the promises and limitations of building virtue-ethical machines. Taking inspiration from various philosophical traditions—including Greek philosophy (Aristotle), Chinese philosophy (Zhuangzi), phenomenology (Hubert and Stuart Dreyfus) and contemporary virtue theory (Julia Annas)—we argue for a novel model of machine ethics we call the “skillful-expert model.” This model sharply distinguishes human virtues and their machine analogues and rejects the direct transposition of human virtues to machines, as suggested by most proposals for machine virtues. Instead, we propose that human skills are better candidates for machine virtues.

Simion and Kelp develop the obligation-based account of trustworthiness as a compelling general account of trustworthiness and then apply this account to various instances of AI. By doing so, they explain in what way any AI can be considered trustworthy, as per the general account. Simion and Kelp identify that any account of trustworthiness that relies on assumptions of agency that are too anthropocentric, such as that being trustworthy, must involve goodwill. I argue that goodwill is a necessary condition for being trustworthy and further suggest a network account of trustworthy AI, which retains the goodwill as the essential requirement for the concept of trustworthiness and meanwhile predicts that current AI can be trustworthy. The alternative account suggests that the focus of trustworthy AI is not merely AI technology, but the whole network of AI involving AI technology, AI designers, AI companies, and other social and legal institutions. A trustworthy AI requires that AI technology is reliable and that other involved agents are trustworthy.

Previous research on moral dilemmas has mainly focused on decisions made under conditions of probabilistic certainty. We investigated the impact of uncertainty on the preference for action (killing one individual to save five people) and inaction (saving one but allowing five people to die) in moral dilemmas. We reported two experimental studies that varied the framing (gain vs loss), levels of risk (probability of gain and loss) and levels of ambiguity (imprecise probability information) in the choice to save five individuals by sacrificing one. We found that participants preferred actions with uncertainty (risk/ambiguity) over inaction. Specifically, we found that participants preferred actions with precise probability information (risk) over inaction, and they preferred actions with modest or high levels of ambiguity over actions with precise probabilities, especially when moral dilemmas had a loss frame. We also observed commission bias in Study 2. We discussed the implications for research in moral decision-making.

With the increasing degrees A.I.s and machines are applied across different social contexts, the need for implementing ethics in A.I.s is pressing. In this paper, we argue for a pluralist hybrid model for the implementation of moral A.I.s. We first survey current approaches to moral A.I.s and their inherent limitations. Then we propose the pluralist hybrid approach and show how these limitations of moral A.I.s can be partly alleviated by the pluralist hybrid approach. The core ethical decision-making capacity of an A.I. based on the pluralist hybrid approach consists of two systems. The first is a deterministic algorithm system that incorporates different moral rules for making explicit moral decisions. The second is a machine learning system that calculates the value of the variables required by the application of moral principles. The pluralist hybrid system is better than the existing proposals as (i) it better addresses the moral disagreement problem of the top-down approach by including different moral principles, while (ii) it reduces the opacity of ethical decision-making compared with a bottom-up system by implementing explicit moral principles for moral decision making.

This paper reveals a serious flaw in the consequentialist solution to the inefficacy problem in moral philosophy. The consequentialist solution is based on expected utility theory. In current philosophical literature, the debate focuses on the empirical plausibility of the solution. Most philosophers consider the cases of collective actions as of the same type as a horse-racing game, where expected utility theory is adequate to solve the choice problem. However, these cases should be considered as of the same type as a coordination game, where the assumption of common knowledge is also required. However, the assumption is implausible—it is impossible to obtain common knowledge of rationality in cases such as voting.

It has been 46 years since the publication of “Famine, Affluence, and Morality” by Peter Singer. In the paper Singer famously challenges readers to radically change their lives to save the lives of others, often in distant lands. With this paper, Peter Singer, perhaps the most famous living philosopher today, made his name and spawned the field of global justice. Although there have been improvements and successes, easily preventable deaths from poverty still occur in large numbers today. Philosophically the paper has been very influential and continues to generate considerable debate. We return to Singer’s argument in “Famine, Affluence, and Morality” and make the case that the argument is incomplete in that while Singer’s position that we are morally obliged to help relieve famine, hunger, and so on is supported by his argument, his position that we act all things considered is not.

The literature has shown that different types of moral dilemmas elicit discrepant decision patterns. The present research investigated the role of uncertainty in contributing to these decision patterns. Two studies were conducted to examine participants' choices in commonly used dilemmas. Study 1 showed that participants’ perceived outcome probabilities were significantly associated with their moral choices, and that these associations were independent from the dilemma type. Study 2 revealed that participants had significantly less preference for killing the individual when the outcome probabilities were stated using the modal verb ‘will’ than when they were stated using the numerical phrasing of ‘100%’. Our findings illustrate a discord between experimenter and participant in the interpretation of task instructions.

Previous studies have found that the proportions of people who endorsed utilitarian decisions varied across different variants of the trolley dilemma. In this paper, we explored whether moral choices were associated with beliefs about outcome probabilities in different moral dilemmas. Results of two experiments showed that participants’ perceptions of outcome probabilities were different between two dilemmas that were similar to the classical switch case and footbridge case. Participants’ judgments of the outcome probabilities were associated with their moral choices. The results suggested that participants might not accept task instructions and thus did not perceive the outcomes in the dilemmas as certain. We argued that researchers who endorse descriptive tasks in moral reasoning research should be cautious about the findings and should take participants’ beliefs in the outcomes into account.

Galen Strawson defends his pessimist position with his famous “Basic Argument”. He attempts to prove that no agent can meet the demands for the ultimate moral responsibility. I argue that the Basic Argument is not impressive because it commits to a linear justification framework under which not only the notion of free will and moral responsibility but every notion would inevitably involve a vicious infinite regress. Surprisingly, this point has not been significantly addressed in the literature of Strawson’s Basic Argument. I scaffold my argument against “Basic Argument” by critically reviewing and comparing Frankfurt, Christman and Ekstrom’s approach to autonomy. I show that any approach, which commits to a linear justification framework, would inevitably involve an infinite regress problem. To make his argument more impressive, Strawson will have to show why he commits to a hierarchical justification framework in the Basic Argument in the first place.

In this article, I argue for a distinct and novel right-based account of risks and I call it the Sophisticated High-risk Thesis. I argue that there is a distinction between rights-infringing risk impositions and no-rights-infringing risk impositions. An action imposing a high risk of harm infringes rights, whereas an act imposing a low risk of harm does not. I also suggest three principles that govern the permissibility of highly risky actions. If a highly risky action meets the conditions specified by any of these three principles, it can be justified. These principles are the consent principle*, the prevent disaster principle* and the reciprocity principle**. I show that the Sophisticated High-risk Thesis is, in general, better than the alternative Risk Thesis defended by McCarthy.

In this paper, I propose and defend a distinct and novel approach to compensation for risk impositions. I call it the Risk-Pooling System of compensation. This system suggests that when X performs an action that imposes a risk of harm to Y, then X is liable to Y, and is therefore obliged to make an ex ante compensation that is roughly equivalent to the expected cost of potential harm to a social- risk pool. If and when Y suffers harm as a result of the risk imposed by X, they then receive an ex post compensation roughly equivalent to the cost of actual harm suffered. This system of compensation creates a social buffer between the risk imposer – the one who has the duty to pay compensation into the pool – and the victim – the one who has the right to receive compensation from the social pool. I contend that the Risk-Pooling System is an improvement over its alternatives due to its capacity to produce the best social utilities, particularly, in terms of reducing information costs, obtaining optimal deterrence in the society and creating incentives for people to be engaged in social activities.