Episode Summary
Executive Summary: Sean Carroll and philosopher Ned Block explore consciousness, especially the split between access consciousness and phenomenal consciousness, and whether AI can truly be conscious. Block argues computational functionalism is insufficient: what matters may be not just computation, but how it is realized—possibly by biological, electrochemical processes. The discussion links philosophy, neuroscience, and AI safety.
Main Topics: Defining consciousness: phenomenal vs. access (Priority: 5/5): Block distinguishes phenomenal consciousness as subjective 'what it’s like' experience, and access consciousness as globally available information used in cognition, decision-making, and problem-solving. Carroll and Block also discuss self-awareness as a related but distinct form. Computational functionalism and its limits (Priority: 5/5): They examine the view that if a system computes the right inputs and outputs, it is conscious. Block argues that computation alone may not be sufficient; the realizer and mechanism of computation may matter for consciousness. Biology, mechanism, and 'meat machines' (Priority: 4/5): Block defends a non-substrate-essential but mechanism-sensitive view: biology may matter not because of 'meat' itself, but because of specific electrochemical processes that may underlie consciousness. Carroll says he is becoming open to this possibility. AI consciousness and moral status (Priority: 5/5): The conversation addresses whether advanced AI systems could be conscious, how to tell, and whether they deserve moral consideration. Block notes current criteria are weak and that companies and ethicists are increasingly engaging the issue. Neuroscience and partial progress on the hard problem (Priority: 4/5): Block suggests neuroscience is making modest progress by explaining how things look and feel in perception, though not solving the hard problem outright. He cites attention-related changes in perceived size/contrast as examples. Philosophical thought experiments and color experience (Priority: 3/5): The pair discuss inverted spectrum and pseudonormal color vision as ways of probing phenomenal experience. Block uses these cases to argue that subjective experience may differ even when behavior and external discrimination look normal. Subconscious processes, repression, and consciousness without access (Priority: 3/5): Block proposes that phenomenal consciousness may depend on subconscious or sub-access processes, including possibly experiences isolated from access consciousness, and that repression could hide access without eliminating phenomenology.
Key Arguments: Phenomenal consciousness is the 'what it’s like' aspect of experience and cannot really be defined, only pointed to, using examples like inverted spectrum and Mary’s room. Access consciousness is global availability of information to cognitive systems such as reasoning, decision-making, and problem-solving; it is not identical to self-awareness. Computational functionalism is likely incomplete: a system may perform the right computations yet still fail to generate consciousness if the implementation/realization is wrong. The relevant distinction for AI is not just roles vs. outputs, but realizers and mechanisms; consciousness may depend on the kind of physical process doing the computing. Block is physicalist, not dualist or panpsychist; he thinks adding nonphysical ontology does not solve the hard problem and often just relocates it. Neuroscience can still help by explaining aspects of phenomenal appearance, such as attention making objects seem larger or higher-contrast. AI systems may eventually become plausible consciousness candidates, but current LLMs are far from clear evidence because they are trained on human first-person expressions and still show brittle reasoning. A purely lookup-table or brute-force Turing-test passer would not obviously count as conscious, suggesting input-output behavior alone is insufficient. The biological details of brains—especially electrochemical signaling—may matter for consciousness in ways that current computers do not replicate. Moral and safety debates about AI consciousness are already active, and the issue matters because companion-style uses of AI create incentives to frame systems as sentient.
Data Points: Publication year of Block's related paper: 1997 - Block cites his earlier BBS reply 'Biology versus Computation in the Study of Consciousness' as an early statement of his view. Attention-related perceptual change: Slight increase in perceived size, contrast, and speed - Carroll and Block discuss Marissa Carrasco’s findings that attention alters visual appearance, with a neurological explanation. Large language model arithmetic accuracy: ~20% accurate on three-digit multiplication - Block cites GPT-3 performance as an example of brittleness in symbolic computation. Timestamp of alternative color-vision paper: About 25 years ago - Block references Martina Nita-Rümelin’s paper on pseudonormal color vision as a possible inverted-spectrum case. AI companion usage share: More than 50% - Block suggests over half of AI uses are as companions, which may affect company incentives around consciousness narratives. Training regime suggestion: Exclude first-person-point-of-view data - Block proposes that an AI trained without first-person narratives and still expressing a first-person perspective would be more convincing evidence of consciousness. Probability estimate for consciousness criteria: 50% / 50% - Block says we have no reason yet to favor computational properties over subcomputational properties in judging alien minds.
Pivotal Quotes: "What matters is the function of various things going on and how they are embodied in some kind of computation." — Sean Carroll: Carroll summarizes the computational functionalist view early in the discussion. "It's not just about what you compute, it's how you compute it that really matters." — Sean Carroll: Carroll describes the alternative view he and Block are exploring regarding consciousness. "I think what's happened in current thinking about artificial intelligence is that a lot of very influential people are computational functionalists." — Ned Block: Block explains why he thinks the AI debate makes his long-standing objections newly urgent.
Implications: The episode argues that AI consciousness cannot be inferred from chat behavior alone; implementation and biology may matter. For listeners, the takeaway is that consciousness research needs better distinctions, better neuroscience, and more serious ethical planning for AI and animals.
About Sean Carroll MindScape
Ever wanted to know how music affects your brain, what quantum mechanics really is, or how black holes work? Do you wonder why you get emotional each time you see a certain movie, or how on earth video games are designed? Then you’ve come to the right place. Each week, Sean Carroll will host conversations with some of the most interesting thinkers in the world. From neuroscientists and engineers to authors and television producers, Sean and his guests talk about the biggest ideas in science, ...