Christian R. de Weerd

Can AI be conscious? A high-profile debate is currently unfolding between proponents of non-biological views and proponents of biological views. Whereas the latter argues that certain small-scale biological properties are necessary to instantiate consciousness, the former denies this fiercely, instead suggesting that the relevant properties are exclusively found at larger non-biological scales. Much attention is currently devoted to resolving this dispute. In this paper, I don’t address the dispute directly. Instead, I will pose a more fundamental problem that remains even if this dispute is resolved. Specifically, despite their disagreements, most researchers on both sides of the dispute adopt functionalism as their metaphysical scaffold to describe which properties are required to instantiate consciousness. However, I argue that if functionalism on either side is used as a metaphysical scaffold, then agnosticism about AI consciousness follows because of what I call the Functional Grain Problem: There are a plethora of functional individuations available for any candidate property, whether biological or non-biological, some of which imply that AI can be conscious, whereas other’s don’t, and neither scientific insights nor the central tenets of functionalism tells us anything about which individuations to go for. Because of this, neither large-scale functionalism (assumed by most non-biological views) nor small-scale functionalism (assumed by most biological views) have any obvious implications about the possibility of AI consciousness, despite a prevailing sentiment suggesting that they do. The result is that even if we figure out whether consciousness requires biology or not agnosticism about AI consciousness is still rationally mandated. I briefly analyze potential ways out of the agnostic predicament, suggesting that they are either problematic or undesirable, and I end the paper by comparing my arguments with other agnostic proposals.

Until very recently the idea that some AI systems may soon be conscious was widely regarded as a far-fetched and purely hypothetical scenario. Recent developments in AI capabilities, however, have caused a surge of interest in the question whether conscious AI may actually be realized soon. A growing number of researchers are starting to advocate for versions of “biological naturalism” about consciousness, arguing that biology is important in one way or another to bring consciousness about in artificial systems. Biological naturalism is often portrayed as an alternative research program to “computational functionalism”, a widely endorsed view according to which consciousness requires implementing the right computations and only contingently relies on biological properties in the human and animal case. Despite being positioned as one of the more formidable alternatives to computational functionalism, biological naturalism as a research program arguably still remains in its infancy. By analysing recent work on the topic, this article aims to outline the contours of biological naturalism as a research program, and identify a number of research directions that can be pursued to further develop it. What can be gleaned from the existing literature is that there are a number of distinct and interesting ways in which biological naturalism can be defended and developed. At the same time, there also exist a number of open questions and challenges that biological naturalists have to contend with moving forward.

On a traditional top-down experimental (TE) approach to consciousness science, researchers start by investigating consciousness in humans, or closely related living animals, and generate hypotheses about the material basis of consciousness based on evidence from experimental paradigms which aim to disentangle conscious from unconscious processing. Only afterwards do they deem it appropriate to trace back the distribution of consciousness across the animal kingdom, including consciousness’ origin. Some scholars have started pushing back against the TE-paradigm, instead arguing that an alternative bottom-up evolutionary (BE) approach to consciousness science is more promising. On this approach, researchers start by investigating the origins of consciousness and heavily rely on evolutionary reasoning to support their hypotheses. Only afterwards do they deem it appropriate to explain how primitive forms of conscious experiences gradually complexified to ultimately give rise to human consciousness. Curiously, there is a glaring difference in the way TE-proponents and BE-proponents treat the explanatory gap. Whereas TE-proponents typically bracket the explanatory gap, BE-proponents suggest that their approach can directly address it. My aim in this paper is two-fold. First, I argue that the way in which two notable BE-proponents, Peter Godfrey-Smith and Walter Veit, address the explanatory gap by trying to close (or substantially narrow) it is ultimately unconvincing. Second, and more generally, I argue that BE-proponents do not have the luxury of bracketing the explanatory gap. Specifically, if the ideas BE-proponents put forward fail to convincingly address the explanatory gap, as I argue is the case for Godfrey-Smith and Veit, then their hypotheses about the material basis and origins of consciousness lose credibility. I end the paper by tentatively suggesting that even though the gap-closing version of the BE-approach seems unconvincing, alternative versions of the BE-approach may be more promising.

On a traditional top-down experimental approach to consciousness science, researchers start by investigating consciousness in humans, or closely related living animals, based on evidence from experimental paradigms that aim to directly disentangle conscious from unconscious processing. Only afterward are these insights (iteratively) extended beyond the human case to investigate and understand how consciousness is distributed more broadly. In A Philosophy for the Science of Animal Consciousness, Walter Veit radically departs from this approach by taking a Copernican turn in suggesting that we should first understand consciousness’ simple beginnings, and only afterward make sense of how these humble beginnings further complexified. Central to Veit’s approach is abruptly removing the reliance on insights about human consciousness by suggesting that evolutionary considerations can provide a credible source of evidence that can independently and directly support hypotheses about consciousness. My aim here is two-fold. Firstly, I will argue that using evolutionary considerations to support hypotheses about consciousness this way is problematic for principled reasons. The consequence is that evolutionary considerations fail to convincingly support Veit’s central hypothesis: That consciousness’ function is to enable organisms to deal with high pathological complexity. However, secondly, I will suggest that these considerations can nevertheless support an adjacent hypothesis: That an evaluative mode of being is required in general to deal with high pathological complexity irrespective of how consciousness is involved. Taken together, I conclude that consciousness science should not be turned upside down, and that Veit’s central hypotheses are better understood as hypotheses not directly concerned with consciousness.

A central question in discussions about artificial consciousness is whether biological properties are necessary for consciousness. In this context, biological properties are often divided between two types: biological substrates as opposed to biological functions. In this paper, I argue that the prospects of convincingly ruling out consciousness in (conventional) AI by appealing to a biological substrate view are unpromising. Specifically, I argue that the biological substrate view faces a dilemma: either the view can be interpreted in a way that makes it empirically respectable in principle, but at the cost of collapsing into a biological function view. Or it can be interpreted as really distinct from a biological function view, but at the cost of being empirically intractable and relying on theoretically arbitrary assumptions. On neither horn does the view amount to a distinct and empirically or theoretically convincing, or even slightly plausible, view. Because of this, I wager that the possibility of AI consciousness will not hinge on what lies dormant beneath, and that the biological substrate view is best be bracketed in discussions about AI consciousness. I conclude the paper on a positive note, suggesting that once the biological substrate view is bracketed, certain dialectical stalemates may dissolve and progress can be made. Specifically, what should be pushed to the forefront is that (1) the dispute between biological naturalism and computational functionalism is arguably less stark than it is often taken to be, and that (2) biological function views have various potential explanatory tales to tale about the relevance of biology to consciousness. Because of this, the focus should be on identifying which doings, whether biological or non-biological, are necessary (and sufficient) for consciousness and why.

Can machines ever become conscious? A central debate in this context concerns the question whether consciousness requires biological states. Within this debate, there exists a fundamental dispute between two widely endorsed views: biological naturalism and computational functionalism about consciousness. Specifically, whereas biological naturalists hold that consciousness requires biological states, computational functionalists fiercly deny this. This fundamental dispute has hitherto remained unresolved, and we discuss several reasons why it is unlikely that it will be resolved anytime soon. Accordingly, we suggest that both sides will likely remain in a dialectical stalemate regarding the fundamental dispute and, by extension, are unlikely to reach a consensus on positive verdicts about consciousness in AI. However, we suggest that progress can be made by investigating potential non-biological constraints on implementation. Such constraints can help reach a consensus between both sides to the effect that certain types of artificial systems are not conscious, if they fail to satisfy relevant constraints. If correct, a consensus on negative verdicts about consciousness in certain challenging types of AI is possible, even if both sides remain in conflict about positive verdicts about AI consciousness.

There is much interest in investigating the evolution question: How did consciousness evolve? In this paper, we evaluate the role that evolutionary considerations can play in justifying (i.e., confirming or falsifying) hypotheses about the origin, nature, and function of consciousness. Specifically, we argue against what we call evolution-first approaches to consciousness, according to which evolutionary considerations provide the primary and foundational lens through which we should assess hypotheses about the nature, function, or distribution of consciousness. Based on the example of Walter Veit's account and additional reasoning, we contend that evolution-first approaches struggle to provide compelling empirical evidence for their key claims about consciousness. In contrast with these approaches, we argue that consciousness science needs to foundationally rely on experimental and observational evidence from humans and other present-day animals. If our arguments succeed, then researchers, when investigating consciousness, are better advised to take as their primary source of evidence consciousness’ present, not its past. Having said this, we acknowledge that evolutionary thinking plays an important role in consciousness science. We delineate this role by stressing several ways in which evolutionary considerations can substantially help advance consciousness research, although in a manner that avoids the evolution-first approach. Since our argument only concerns the assessment of hypotheses (the “context of justification”), it leaves it open which role evolutionary considerations play in generating hypotheses (the “context of discovery”). That is, evolutionary considerations may nevertheless play a foundational role in hypothesis generation in consciousness science.

Many different methodological approaches have been proposed to infer the presence of consciousness in non-human systems. In this paper, a version of the theory-heavy approach is defended. Theory-heavy approaches rely heavily on considerations from theories of consciousness to make inferences about non-human consciousness. Recently, the theory-heavy approach has been critiqued in the form of Birch's (Noûs, 56(1): 133-153, 2022) dilemma of demandingness and Shevlin's (Mind & Language, 36(2): 297-314, 2021) specificity problem. However, both challenges implicitly assume an inapt characterization of the theory-heavy approach. I argue that an alternative characterization of the approach, what I call a credence-based theory-heavy approach, avoids these challenges. Theorists can generate interpretations of their theory, at different levels of generality, and operationalize these into theory-informed markers. These theory-informed are assigned a likelihood and are used to assess the probability that a target system is conscious. In providing this characterization, and mapping out the possible ways in which a credence-based theory-heavy approach can be fleshed out, the aim is to situate the theory-heavy approach as a more compelling approach than it is currently being perceived as. Our attention, then, needs to towards remaining challenges such as the consensus problem and the problem of calibrating the likelihoods associated with theory-informed markers. I also explore methodological pluralism and assess how the credence-based theory-heavy approach can benefit from other methodological approaches.