Hyungrae Noh

Traditional moral frameworks attribute responsibility only when an agent’s actions align with psychological capacities—such as intention and controllability. These human-centered requirements produce the ‘responsibility gap’ in AI ethics: AI systems operate through opaque, complex, and semi-autonomous processes, so human stakeholders involved in AI-caused harm are deemed not responsible because they neither intend nor can predict such harm, and AI systems cannot be held responsible because they lack the requisite mental capacities. Drawing on experimental philosophy research, this paper shows that laypeople’s conceptions of agency and responsibility in AI contexts systematically diverge from these requirements. Empirical studies reveal that people routinely attribute moral agency to AI systems while explicitly denying them consciousness and free will. Moreover, laypeople distribute moral responsibility for AI-caused harm across networks of human stakeholders (programmers, manufacturers, users) and the AI systems themselves. Crucially, responsibility is allocated even when human stakeholders are deemed innocent, and its purpose is to prevent future harm through corrective measures rather than to punish or deter. These findings provide empirical support for philosophers advocating shared moral responsibility, such as Floridi, Latour, Verbeek, and Gunkel. By treating everyday intuitions as evidence for theory assessment, the paper argues that the responsibility gap reflects outdated human-centered assumptions rather than genuine conceptual problems. It concludes that philosophers should abandon the responsibility gap in favor of addressing the ‘problem of many hands’—the practical and conceptual challenge of implementing distributed moral responsibility in AI governance—while developing new moral concepts suited to the AI era through experimental philosophy and conceptual engineering.

The functionalist conception of artificial moral agency holds that certain real-world AI systems should be considered moral agents because doing so benefits the recipients of AI actions. According to this view, human agents who are causally accountable for the morally significant actions of these AIs are deemed blameworthy or praiseworthy and may face sanctions or rewards, regardless of whether they intended the AI actions to occur. By meta-analyzing psychological experiments, this paper reveals a close alignment between the functionalist conception and the folk understanding of artificial moral agency. People treat certain real-world AI systems as moral agents even when they do not attribute consciousness or free will to them. When ordinary people attribute moral responsibility to these systems, these attributions are also distributed among the users, programmers, and manufacturers of the systems. This distribution holds even when people do not view the causal contributions of these human agents to these systems’ actions as wrongful.

Behaviorally assessing residual consciousness of acutely comatose patients involves a high rate of false-negatives. That is, long-term behavioral assessment shows that 41% of vegetative state patients in fact have residual consciousness. Nonetheless, surrogates need to remove ventilation before the acute-phase passes away if they want to induce medico-legal death due to pragmatic factors, such as financial costs. So, surrogate decision-making regarding behaviorally nonresponsive acutely comatose patients involves a moral dilemma: should we ignore the chance that patients have residual consciousness for the sake of pragmatic factors? This paper examines whether neural methods can resolve the moral dilemma. Neural methods are used to assess residual consciousness of behaviorally nonresponsive postcomatose patients. For instance, by instructing a vegetative state patient to imagine wiggling all of her toes, conscious- ness is ascribed if brain activities are localized in the supplementary motor area. Since the most extensive application of neural methods has been in chronic population, it is unclear whether such methods can resolve the moral dilemma. I argue that neural methods also involve a high rate of false-negatives because current tasks of neural methods are structurally misguided. Given the argument, there is no significant difference between behavioral and neural methods regarding the moral dilemma.

Intuitively, proper referential extensions of psychological and moral terms exclude artifacts. Yet ordinary speakers commonly treat AI robots as moral patients and use psychological terms to explain their behavior. This paper examines whether this referential shift from the human domain to the AI domain entails semantic changes: do ordinary speakers literally consider AI robots to be psychological or moral beings? Three non-literalist accounts for semantic changes concerning psychological and moral terms used in the AI domain will be discussed: the technical view (ordinary speakers express technical senses), the habit view (ordinary speakers subconsciously express ingrained social habits), and the emotion view (ordinary speakers express their own affective empathetic emotional states). I discuss whether these non-literalist accounts accommodate the results of relevant empirical experiments. The non-literalist accounts are shown to be implausible with respect to the ordinary use of agency-terms (e.g., “believe,” “know,” “decide,” etc.), and therefore I conclude that the concepts ordinary speakers express by agency-terms in reference to AI robots are similar to the concepts they express when applying the same terms to humans. When ordinary speakers extend emotion-terms and/or moral-patiency-terms to AI robots, however, I argue that semantic changes have taken place because ordinary speakers are in fact referring to their own affective empathetic emotional states rather than AI robots. This argument suggests that the judgments made by ordinary speakers regarding the proper referential extensions of emotion-terms and moral-patiency-terms are fallacious.

Post-comatose patients are classified as being in a minimally conscious state when they have executive functions. Because traditional behavioral assessments may not capture signs of executive functions in post-comatose patients, clinicians look to localized brain activities in response to task instructions, such as imagining wiggling toes, to diagnose minimal consciousness. This paper critically assesses the assumption underlying such alternative methods: that brain activities are neural signals conveying information about minimal consciousness. Based on a Shannon-inspired idea of information, I distinguish between informational and engineering aspects of clinical tasks. The informational aspect concerns the conditional probability that, for example, given activity in the motor areas of the brain in response to task instructions, a patient is imagining wiggling toes. The engineering aspect concerns efficient activation of the relevant brain areas in a patient under the task conditions. This distinction shows that the current alternative methods are not informationally problematic, but are structurally “ill-formed.” For instance, the toe-imagery task requires the capacity to comprehend syntactically complex sentences, which can be dissociated from minimal consciousness. I propose a misrepresentation task, which tests the capacity to misconceptualize lukewarm water as melting wax, as a supplement to the current alternative methods. This task is as informationally reliable as these methods, but is structurally “well-formed,” as it does not rely methodologically on prerequisites such as language comprehension.

The minimally conscious sta te (MCS) is usually ascribed when a patientwith brain damage exhibits obser vable volitional behaviors that predict recovery ofcognitive funct ions. Nevertheless, a patient with brain damage who lacks motorcapacit y might nonetheless be in MCS. For this reason, some clinicians use neuralsignals as a communicative means for MCS ascription. For instance, a vegetativestate patient is diagnosed with MCS if activity in the motor area is observed whenthe instruction to imagine wiggling toes is given. The validi ty of using neuralsignals in ascribing MCS requires a special sort of inference. That is, no-reportparadigmatic assessments must have inductively strong ways of inferring a pur-ported informationa l content from the observed neural signal that grounds the factthat the patient has top-down cognitive control (or residual volition). Shannon’smathemat ical theory of communication and Bayes’ theorem reveals the formalstructure of neural communication. On the basis of relevant data from the neuro-science literature, I conclude that the formal structure combined with the data showsthat neural signals can be used as a communicative means for operational diagnosticcriteria for MCS ascription.