Aorigele Bao
Affiliation: College of Humanities, University of Chinese Academy of Sciences
In this paper, we consider how welfare might apply to artificial intelligence (AI). Some argue that AI has no sense of pain or subjective consciousness and, therefore, we should not consider AI’s welfare from a moral perspective because AI cannot feel pain. Their inability to feel pain implies that AIs cannot be assigned the same moral considerations as animals, so AIs should not be viewed as having ‘compromised welfare’ (Bryson, 2010; Gibert & Martin, 2022; Sparrow, 2004).
However, in this paper we defend the notion of unconscious harm to AIs by drawing on the discussion of unconscious harm to the dead. We seek to show that AIs can be unconsciously harmed despite their lack of subjective capacity to feel.
First, by unconscious harm to the dead, we mean Boonin’s idea in his ‘Dead Wrong: The Ethics of Posthumous Harm’: even though a person is dead, they can still be harmed by others’ actions (Boonin, 2019). If welfare is measured only in terms of subjective pain or pleasure, both the dead and the AI would be excluded. However, according to Boonin’s reasoning, a person can still be harmed after death because, during life, they had certain desires and goals that can be destroyed or nullified by others’ actions after their death. The deceased cannot subjectively feel the betrayal or waste associated with such an act, but we still say that the deceased has been harmed.
Bunin presents an example of unconscious harm in the case of a cheater: Bob and Carol are a couple in a monogamous marriage, and Carol decides to betray Bob by starting an affair. Bob is completely unaware of Carol’s betrayal (i.e., he does not know that Carol has betrayed him). Yet from a bystander’s perspective, the wife’s betrayal clearly harms Bob’s well-being. Although Bob had no idea about Carol’s affair, he is still viewed as having been harmed. This suggests that for the deceased, it does not matter whether their wish was formed consciously or not; what matters is that the deceased was harmed by that wish.
Thus, the postmortem harm theory emphasises the deceased’s ‘objective frustration of a wish from life’ and does not require any postmortem consciousness. Hence, ‘once-consciousness’ does not come into play at the moment of death. Although the dead are not ‘experiencing’ anything, we still consider them victimised. It is therefore clear that the existence of consciousness is not a necessary condition.
For the deceased, the key lies in the ‘loss of benefits during life’ or ‘obstruction of continuity value.’ That is, benefits or value post-death have nothing to do with the subjective state of the deceased. The welfare of the deceased in this sense is unrelated to whether the deceased is conscious but can make bystanders feel that the deceased has been victimised. Moreover, it can worsen the overall progression of the deceased’s life from the onlooker’s perspective.
What the dead and AI share is the lack of any subjective capacity. Both are considered unconscious in terms of consciousness possession: that is, neither experiences subjective feelings, which can be intuitively acknowledged.
One objection is that the dead are considered unconsciously harmed because they were once conscious (i.e., they had interests), whereas AI was never conscious, and its ‘interest’ is simply a designer-assigned function. We should clarify that ‘harm’ here does not require the deceased’s present consciousness. Instead, it emphasises the frustration of some pre-death wish. Thus, for the deceased, the harm does not arise because ‘he had consciousness and now it is gone,’ but because ‘the wishes or values he once established have been destroyed by reality.’ The presence or absence of consciousness at death is irrelevant; it is the destruction of the desire or value that matters. Furthermore, wishing is often an attitudinal state not always in one’s conscious awareness. That is, the deceased’s wishes derive from what we think the deceased should have wished for, not from any actual expression of the deceased’s will.
Moreover, an AI that assumes complex social or technological responsibilities in its goals is not structurally unlike the family or social responsibilities once held by the deceased. We can imagine a ‘completely comatose patient’ who never wakes, never displays consciousness, yet whose ‘interests can be violated,’ as recognised by law and morality—a situation quite similar to that of AI.
Therefore, considering the case of injury to the deceased, we can describe it as ‘damaging something of value that the entity relies on.’ While the deceased once had real consciousness, the AI had none. Nonetheless, in determining ‘whether this constitutes harm,’ the key lies not in the presence of subjective consciousness, but in whether there is a malicious discrepancy between the ‘goal and the result.’
For example, this benefit emerges in the AI’s ongoing execution of its tasks. Once task continuity is interrupted, even though the AI does not experience pain, we can apply the victimisation-of-the-dead argument: the AI’s systemic value is thwarted, and it therefore suffers a kind of unconscious harm. Similarly, just as a deceased person has no further opportunity to feel betrayal, the AI has no way to feel the pain of deprivation; yet objectively, its ‘condition’ is compromised by the frustration of its aspirations and goals.
In conclusion, if one accepts that the deceased can be harmed by the frustration of their wishes, then one can also accept that AI is harmed on similar grounds. If welfare is considered in terms of harm, then AI should be included as an object of welfare for the purpose of avoiding harm, regardless of whether AI is conscious.