If an AI told you it was conscious, how would you know whether to believe it?
My links: https://linktr.ee/frictionphilosophy.
1. Guest
Eric Schwitzgebel is Distinguished Professor of Philosophy and University of California MacArthur Foundation Chair. His work has focused on philosophy of psychology, philosophy of mind, artificial intelligence, moral psychology, classical Chinese philosophy, epistemology, metaphilosophy, metaphysics, and more.
Check out his element with Cambridge Elements, “AI and Consciousness”!https://www.cambridge.org/core/elements/abs/ai-and-consciousness/E77C92088DA3C9F89E7FE7C75CBB1896https://a.co/d/01yvIzYI
2. Book Summary
Schwitzgebel argues that we don’t know whether near-future AI systems will be conscious, and that we won’t know before we’ve built huge numbers of disputably conscious systems, because engineering is outpacing consciousness science. He first rejects the claim that the answer is obvious either way, pointing to serious experts on both sides: Dehaene, Lau, Chalmers and the IIT camp on one, Seth, Godfrey-Smith, Block and Searle on the other. He defines consciousness by example in the standard phenomenal sense, and argues that “AI” is actually the blurrier concept, which matters because sweeping impossibility arguments usually assume something narrow such as digital computation. The core of the book is an argument built around ten possibly essential features of consciousness: luminosity, subjectivity, unity, access, intentionality, flexible integration, determinacy, wonderfulness, specious presence and privacy. If any of these is essential, it constrains which AI systems could be conscious, but we can’t establish which are. Introspection fails because of its poor historical track record, its structural bias toward experiences that are known and accessed, and the narrowness of a human-only evidence base for claims about all possible minds. Conceptual arguments fare no better: Rosenthal’s case for higher-order theory works only by sliding between the epistemic and experiential senses of “conscious,” Brentano’s intentionality thesis rests on enumerating human cases, and felt unimaginability may just reflect imaginative ruts, like a child who denies that imaginary numbers could exist.
Materialism and functionalism are broadly friendly to AI consciousness, and none of the main non-materialist alternatives (dualism, panpsychism, idealism) rules it out. Behavioral tests don’t settle the question either. There is no single Turing test, only a spectrum of bars, and no fair bar reliably tracks consciousness, since high bars are unfair and low bars are already passed. Searle’s Chinese Room and Bender’s underground octopus fail as refutations because their intuitions don’t scale to systems of realistic size, but they point toward the book’s best skeptical argument, the Mimicry Argument. When a system is designed to reproduce the surface features that indicate consciousness in humans, that design undercuts the usual inference from those features to consciousness, though it does not establish that the system is nonconscious. LLMs pretrained on human text currently fall under this argument, although post-training and world-directed training would weaken it. The leading scientific theories (Global Workspace, Higher Order, Integrated Information, Local Recurrence and Unlimited Associative Learning) lack consensus even for the human case. Most also face what he calls the Problem of Minimal Instantiation, since trivially simple systems satisfy their criteria, and all face the Problem of the Narrow Evidence Base, since empirical support from vertebrates doesn’t license claims about radically different architectures. The partial remedies he discusses, the indicator-properties approach and Chang-style iterative natural-kind methods, are long-horizon.
Arguments from biology cut neither way decisively. Autopoiesis lacks a developed defense as a necessary condition, and robots or artificial-life systems could plausibly be autopoietic. The neural replacement argument also fails, because gradual replacement guarantees unchanged self-reports whether or not consciousness persists, and because, following Cao, faithfully replicating a neuron’s function probably requires another neuron. Even so, a Copernican mediocrity argument suggests consciousness shouldn’t depend on fine-grained, Earth-specific biology, which supports some liberalism about substrate. The book closes with the Leapfrog Hypothesis: the first conscious AI will likely be complex and verbally sophisticated rather than frog-like, since sophisticated capacities already exist without consciousness and could be bolted on once consciousness is achieved. Such systems may be strange intelligences, distributed, disunified, non-private, or impossible to count as whole-number subjects. With science unable to decide, Schwitzgebel predicts a Social Semi-Solution in which social and financial motives drive theory choice and produce a consensus that resolves the conflict without being warranted by the evidence. His recommendation is to avoid building debatably conscious systems at all, creating only systems that are clearly nonconscious on every viable theory or clearly conscious on all of them. His larger aim is to make principled uncertainty something we can hold onto under social pressure.
3. Interview Chapters
00:00 - Introduction
01:21 - Importance
03:28 - What is AI, consciousness?
07:59 - Near vs. distant systems
10:07 - Moral consideration
14:20 - Methodology
17:08 - Questioning the features
20:47 - Sampling bias
22:05 - Determinacy
30:02 - Problem for physicalism?
31:49 - Consciousness and moral standing
38:13 - In the dark?
40:51 - General theories of consciousness
44:36 - Turing test
49:00 - Worry
57:14 - Biology
1:03:00 - Neural replacement
1:09:38 - Leapfrog hypothesis
1:12:55 - Worry
1:16:52 - Policy
1:20:52 - Prediction
1:23:42 - Value of philosophy
1:27:47 - Conclusion
