Jan Henrik Wasserziehr

Recent proposals in neuroscience and philosophy suggest that coarse-grained computational functionalism may suffice for artificial consciousness. However, I argue that even if such accounts are right, there is no reason to assume that consciousness so realised would, by default, be valenced. On a naturalistic conception of value, valence—the affective quality of subjective experience—presupposes entities for whom things can be non-derivatively good or bad. In living organisms, valence is primordially grounded in a predisposition toward self-preservation, relative to which states of the world can be objectively better or worse. Silicon-based artificial systems appear to lack functionally equivalent dispositions relative to which certain states of the world would be objectively preferable for them over others. This, I argue, gives rise to a value grounding problem: if artificial systems do not possess non-derivative goals, it becomes unclear what could ground subjective attributions of value on their part, i.e. valenced states. I discuss four potential pathways to artificial valence—designer-independent goals, reinforcement learning, rational evaluation, and hallucinations—arguing that none satisfactorily solves the value grounding problem. If the account offered here is correct, it is unlikely that recent proposals for artificial consciousness will entail valenced states, that is, sentience.

Recent proposals in neuroscience and philosophy suggest that coarse-grained computational functionalism may suffice for artificial consciousness. However, I argue that even if such accounts are right, there is no reason to assume that consciousness so realised would, by default, be valenced. On a naturalistic conception of value, valence—the affective quality of subjective experience—presupposes entities for whom things can be non-derivatively good or bad. In living organisms, valence is primordially grounded in a predisposition toward self-preservation, relative to which states of the world can be objectively better or worse. Silicon-based artificial systems appear to lack functionally equivalent dispositions relative to which certain states of the world would be objectively preferable for them over others. This, I argue, gives rise to a value grounding problem: if artificial systems do not possess non-derivative goals, it becomes unclear what could ground subjective attributions of value on their part, i.e. valenced states. I discuss four potential pathways to artificial valence—designer-independent goals, reinforcement learning, rational evaluation, and hallucinations—arguing that none satisfactorily solves the value grounding problem. If the account offered here is correct, it is unlikely that recent proposals for artificial consciousness will entail valenced states, that is, sentience.

In this paper, I review the recent debate on the prospect of AI consciousness and assess its ethical relevance. Intuitively, when a being is conscious, this is sufficient to ground its non-derivative moral status. However, the semantic content of consciousness is ambiguous, particularly when attributed to non-biotic entities. Standard ethical accounts concerning the moral status of sentient beings presuppose affective or valenced consciousness. Yet, recent speculations about artificial consciousness have largely been driven by considerations of cognitive architecture, often neglecting the affective aspect of subjective experience. Crucially, no persuasive narrative currently exists about how valenced states may emerge in artificial systems. This significantly reduces the ethical relevance of current AI consciousness claims. As ethicists, we are well-advised to treat such claims with care, demanding precision and rigor when consciousness is attributed to artificial systems. This involves developing logically coherent claims about the emergence of artificial valence.