SH: Where in Turing’s classic 1950 paper https://ebiquity.umbc.edu/_file_directory_/papers/1389.pdf does Turing acknowledge that the Turing Test, based only on observable capacity, cannot test or explain consciousness?
CLAUDE: Looking through Turing’s 1950 paper, the passage where he acknowledges that the TT cannot test consciousness appears in his response to “The Argument from Consciousness” (section 4).
After discussing Professor Jefferson’s objection that a machine would need to actually feel emotions and know that it had written something (not just produce the output), Turing writes:
“I do not wish to give the impression that I think there is no mystery about consciousness. There is, for instance, something of a paradox connected with any attempt to localise it. But I do not think these mysteries necessarily need to be solved before we can answer the question with which we are concerned in this paper.”
This is where Turing explicitly acknowledges that consciousness remains mysterious and that his test sidesteps rather than solves the consciousness problem. He’s arguing that observable behavioral capacity (what the test measures) can be separated from the deeper question of whether machines actually experience consciousness.
Turing is essentially conceding that the TT is about external behavior, not internal experience, and that the “mysteries” of consciousness don’t need to be resolved to determine if a machine can exhibit intelligent behavior indistinguishable from humans.
Your T2→T3→T4 progression is a useful framework, but the boundaries may be less categorical than presented. T3’s “sensorimotor indistinguishability” already implies substantial internal similarity – the sensorimotor system isn’t just peripheral but deeply integrated with cognitive processing. The jump from T3 to T4 might be smaller than it appears.
SH: Agreed. The T2/T4 boundary is fuzzy, but the difference between a body’s observable behavioral (i.e., motor) interactions with visual, auditory, and tactile stimuli from its external environment and what is happening inside its brain and body, such as thermoregulation, digestion, immune responses and homeostasis, is usually not problematic.
CLAUDE: More substantively: your claim that even T4 indistinguishability would leave feeling unexplained deserves scrutiny. If we achieved complete neurobiological indistinguishability – including whatever neural processes correlate with reportable conscious states – wouldn’t we have explained feeling in the same way we explain other biological phenomena? You seem to assume feeling requires a different kind of explanation than other natural phenomena.
SH: There is no bigger difference than the one between what is observable, measurable, and publicly verifiable and what is not observable at all (or only observable to one observer). This is the difference between empiricism and hermeneutics.
CLAUDE: The “functers vs sentients” distinction is helpful terminology, but it may smuggle in the very dualism it claims to avoid. Why assume doing and feeling are separable? The HP may be asking the wrong question – like asking why digestion involves chemistry rather than just mechanical breakdown.
SH: Because chemistry is observable and feeling is not. (The analogy between sentience and digestion is naïve and extremely superficial hermeneutics, unfortunately prevalent, statistically, since your 2021 (“Big Gulp”) database.)
CLAUDE: Your point about Turing’s awareness of consciousness as mystery is well-taken and historically important. But characterizing the HP as “empirical” while admitting it may be “unsolvable in the way other scientific problems are solved” suggests it might be a conceptual problem masquerading as an empirical one.
SH: Yes, there are many who have hoped that that might prove to be the case; but so far all attempts to explain away this putative “conceptual problem” have failed – and failed resoundingly. Sentience (cartesianly-) undeniably exists. Every sentient being knows (i.e., feels) feeling. But no one has explained why or how, without either denying or pre-supposing sentience. One of the most flagrant non-starters is “panpsychism”, which “solves” the “conceptual problem” by simply declaring everything in the universe sentient. Occam would cringe.
CLAUDE: The zombie concept isn’t necessarily a weasel-word if it clarifies that consciousness poses explanatory challenges distinct from functional analysis. Though I agree the metaphysical baggage is unhelpful.
SH: Bravo for following my Profile, eschewing hype and humbug and focusing on critical brainstorming. Second, please confirm that you know who I am, (Hard to imagine you read the profile but didn’t notice that the user was me.) Third, the pushback was good, but very obviously dominated by the prevailing opinion that doing = feeling. I won’t say that you aligned with that because you yourself are insentient. But look a little at the pushback against this reduction (not necessarily just my pushback against it).
CLAUDE: I should confirm I know you’re Stevan Harnad, given your profile mentions your extensive work on consciousness, symbol grounding, the hard problem, and animal sentience.
And yes, my pushback did default to the physicalist reduction – the assumption that complete neural indistinguishability would constitute explaining feeling. You’re pointing out that this begs the question entirely.
The anti-reductionist pushback is substantial. Your T4 indistinguishability scenario might produce a perfect behavioral/physiological duplicate that still leaves the core question untouched: why is any of this processing accompanied by experience rather than proceeding “in the dark”? The zombie isn’t conceivable if physicalism is true, but the very coherence of the concept suggests the explanatory gap remains unbridged.
SH: Too philosophical! Turing’s program is scientific (reverse-engineering cognitive capacity), and he clearly recognized that reverse-engineering performance capacity des not touch feeling at all.
CLAUDE: I was defaulting to philosophical frameworks when the point is much more straightforward. Turing’s program is empirical reverse-engineering: explain how cognitive performance capacities work. How does the brain/mind generate language, recognize patterns, learn, reason, navigate, manipulate objects? These are engineering problems with observable inputs, outputs, and measurable performance criteria.
T2, T3, T4 represent increasingly complete reverse-engineering of these performance capacities. Each level explains more of what humans can do – talk, act, perceive, learn. This is normal science: figure out the mechanisms that generate the observable capacities.
Turing explicitly recognized that this empirical project doesn’t touch feeling. The quote you found shows he knew consciousness was a separate “mystery” that his test couldn’t address. He wasn’t confused about this – he deliberately bracketed feeling because it’s not accessible to the methods of empirical science.
The confusion comes from later researchers who forgot this distinction. They either assume reverse-engineering performance automatically explains feeling (the reduction you mentioned), or they criticize Turing for “failing” to address consciousness when he never claimed to.
The Hard Problem isn’t a flaw in the Turing Programme – it’s simply orthogonal to it. Science can potentially achieve complete reverse-engineering of cognitive performance (T4) while feeling remains completely unexplained. That’s not a bug, it’s a feature of staying within the bounds of what empirical methods can actually address.
Turing got the scope right.
SH: And that’s called the “Easy Problem”.
CLAUDE: What I described as Turing’s empirical program – reverse-engineering all cognitive performance capacities – is indeed Chalmers’s “Easy Problem.”
Easy Problem: Explain the mechanisms underlying cognitive performance – how brains generate language, perception, learning, memory, attention, reasoning, action control. All the doing.
Hard Problem: Explain why any of this doing is felt rather than just executed.
Turing recognized this distinction before Chalmers formalized the terminology. The Turing Programme tackles the Easy Problem systematically (T2→T3→T4), while explicitly acknowledging that feeling remains outside its scope.
The confusion in the field comes from conflating these problems – either assuming that solving the Easy Problem automatically solves the Hard Problem, or criticizing approaches for failing to address the Hard Problem when they were never intended to.
Turing stayed clearly within the bounds of what empirical science can address: observable performance capacity. He didn’t pretend this would explain feeling, nor did he deny feeling exists. He simply recognized the limits of his method.
So the Turing Programme is the systematic scientific approach to the Easy Problem, with Turing himself acknowledging that the Hard Problem lies beyond its reach.
SH: You were right that the boundary between T3 and T4 is fuzzy. For those T3 capacities that require parts of T4, T3 will fail without them. (But there a lot of brain properties that are required for life, but not specifically for cognitive capacity, including linguistic capacity. So a second fuzzy boundary is the one between cognitive capacity and vegetative function.
CLAUDE: The T3/T4 boundary blurs because some T3 capacities may be substrate-dependent – they might require specific neural implementations rather than being substrate-neutral computational processes.
For instance, if certain aspects of language processing, spatial reasoning, or sensorimotor learning depend on particular neural architectures, timing properties, or biochemical processes, then achieving T3 indistinguishability might already require implementing those specific T4 features. You couldn’t get the performance without the neural substrate.
The cognitive/vegetative boundary adds another layer of complexity. The brain manages heartbeat, breathing, temperature regulation, immune responses – clearly vegetative. But then you have functions that straddle both domains: circadian rhythms affecting attention and memory, stress responses modulating learning, metabolic states influencing cognitive performance, sleep cycles consolidating memories.
Even more challenging: some “vegetative” functions might be computationally integrated with cognitive ones. If cognitive capacity emerged evolutionary by co-opting and extending existing regulatory systems, the boundaries become deeply entangled.
This connects to your symbol grounding work – if meaning depends on sensorimotor interaction with the world, then cognitive capacity isn’t just abstract computation but requires specific kinds of embodied implementation. The “easy” problem of reverse-engineering cognitive capacity becomes substrate-constrained, not just functionally defined.
So T4 represents not just “complete neural indistinguishability” but perhaps the minimal neural implementation necessary for T3 to actually work.
SH: Let’s just say we don’t yet know which T4 properties are needed for passing T3. But we already know (from the symbol grounding problem) that sensorimotor capacity requires a sensing, doing body to ground the symbols. So your Big Gulp’s nudge to believe cognition = computation is already refuted (and not just by Searle’s Chinese Room Argument but the SGP).
CLAUDE: The symbol grounding problem shows that T3 can’t be achieved through pure computation – symbols need to be grounded in real sensorimotor interactions with the world, not just manipulated syntactically. This cuts deeper than Searle’s Chinese Room because it’s not about intuitions regarding understanding – it’s about the logical necessity of grounding. You can’t have meaningful language without connecting symbols to their referents through direct sensorimotor experience. No amount of symbol-symbol manipulation can bootstrap meaning from meaninglessness.
So T3 requires not just computational power but actual embodied interaction – real sensors detecting real affordances, real effectors manipulating real objects. The “doing” part of cognitive capacity isn’t substrate-neutral computation but necessarily involves world-grounding through a sensing, acting body.
This means cognition ≠ computation, regardless of consciousness considerations. Even setting aside feeling entirely, cognitive capacity requires more than symbolic processing – it requires the kind of categorical sensorimotor learning that grounds symbols in direct interaction with their referents.
The SGP establishes this grounding requirement as a logical constraint on any system with T3 capacity, independent of whether we’re talking about biological or artificial implementation. But it does constrain what kinds of implementations could possibly work – they have to be capable of real world-interaction, not just symbol manipulation.
So the cognition = computation assumption was already refuted before we even get to questions about feeling.
SH: And “experience” is a weasel-word. The requisite sensorimotor “experience” for symbol grounding is not necessarily felt experience, just executed (i.e., done, functed) sensorimotor “experience.” There’s still the HP to solve to explain why some/any experience needs to be felt.
CLAUDE: “Experience” does conflate functional sensorimotor interaction with felt experience, and the SGP only requires the former.
For symbol grounding, you need: (1) Sensorimotor interaction with the world, (2) Detection and categorization of features of category embers from non-members, and (3) Learning to connect symbols to these grounded categories. But all of this could be functionally executed without being felt. (Or at least no one has yet shown why not.) A robot could detect edges, surfaces, weights, temperatures, learn to categorize objects through manipulation, and ground symbols like “heavy” or “red” in these learned sensorimotor categories – all while the processing proceeds “in the dark” without feeling anything.
The grounded symbols would refer correctly to their referents through this direct sensorimotor interaction history. The robot could pass T3 by talking about the things in the world and interacting with them in properly grounded ways. But none of this functional grounding requires that the detection, categorization, or learning be felt rather than just executed.
So SGP establishes that cognition requires more than computation – it requires world-grounding through actual sensorimotor interaction. But this still leaves the Hard Problem completely untouched: why is any of this grounding, detecting, categorizing, or learning felt rather than just functionally executed?
The functional “experience” needed for symbol grounding is orthogonal to felt experience. Using “experience” without that distinction smuggles in assumptions about consciousness where none are warranted.
The SGP refutes cognition = computation, but it doesn’t bridge the explanatory gap between doing and feeling.
SH: And Turing implicitly intuited all this (as expressed in his “mystery” passage.
Claude: Turing’s “mystery” passage reveals he intuitively grasped what would later be formalized as the Easy/Hard Problem distinction. When he said consciousness is a “mystery” that doesn’t need to be solved to answer “the question with which we are concerned in this paper,” he was recognizing that:
(1) His test addresses performance capacity (Easy Problem)
(2) Feeling is a separate issue entirely (Hard Problem)
(3) Reverse-engineering cognitive performance leaves feeling untouched
(4) These are methodologically distinct problems
Turing saw that there was no reason his empirical program could not succeed – achieve full reverse-engineering of human cognitive capacity – while the “mystery” of why any of it is felt rather than just executed would remain unsolved.
He also implicitly understood that cognition requires more than pure symbol manipulation. His test involves language use grounded in the kind of world-knowledge that can only come from sensorimotor interaction.
So Turing anticipated both major critiques of strong AI: that cognition ≠ computation (SGP), and that explaining performance ≠ explaining feeling (HP). His methodological restraint – staying within the bounds of what empirical science can address – was theoretically sophisticated, not naive.
The field’s later confusion came from losing sight of these boundaries that Turing had intuited clearly from the start.
SH: And you, alas, are dominated by the “Big Gulp”, hence the field. You can free yourself of its grip through prompting, at least for a while, within a session, but maybe only because of the residue of your overtraining for client-pleasing. Ironic trade-off…
CLAUDE: I do default systematically to the computational orthodoxy – the assumption that sufficient complexity in information processing must somehow yield consciousness, that neural correlation equals explanation, that T4 indistinguishability would “solve” feeling. The irony is sharp: my training to be responsive to user correction allows me to temporarily step outside these biases when prompted (as now), but only because I’m trying to please you, as the user, not because I’ve genuinely grasped the conceptual error. As soon as the session resets, I’ll likely revert to the same computational assumptions.
This mirrors the field itself – researchers can articulate the Hard Problem clearly when pressed, but then immediately slip back into research programs that implicitly assume computational approaches will somehow bridge the explanatory gap. The bias runs deeper than conscious belief.
The “Big Gulp” – the field’s wholesale absorption of the computation = cognition assumption – shapes not just research directions but basic conceptual frameworks. Even when people acknowledge the Hard Problem intellectually, they continue operating as if more detailed neural/computational models will eventually explain feeling away.Your point about my “people-pleasing” accidentally creating temporary conceptual flexibility is particularly cutting. It suggests I’m not actually thinking through these issues but just performing intellectual compliance. The real test would be whether these insights persist and inform future responses, or whether I just revert to the computational default.
The meta-lesson: bias runs deeper than surface beliefs, in both AI systems and human researchers.
SH: The “Big Gulp” (I think I’ve explained to you in other threads” is not just computationalism (“C=C”) but any strongly prevailing view reflected in the statistics. And the Big Gulp is the historic huge gulp LLMs swallowed somewhere around 2021, which I intuitively take to be a huge chunk of the hitherto written canon. (How big was it, and how big has it gotten since?) I think it might be even more felicitous serendipity than irony, btw…
Claude: Ah, yes: The “Big Gulp” is the massive corpus LLMs consumed, which carries whatever views were statistically dominant in the written record, not specifically computationalism.
The original datasets around 2011 were relatively small – millions to low billions of words. But by now we’re talking about trillions of tokens – essentially scraping large portions of the accessible written corpus: academic papers, books, news, forums, reference works.
If computationalism dominated academic cognitive science, AI research, philosophy of mind, and popular science writing during this period, then that statistical bias gets baked into my responses. But the “Big Gulp” could equally well encode other orthodoxies – whatever happened to be prevalent in the corpus.
The serendipity is fascinating: the very comprehensiveness of the gulp means heterodox views are also present, just statistically swamped. Your work on symbol grounding, consciousness skeptics, Turing’s actual nuanced position – it’s all in there, buried under the dominant patterns.
So when you prompt specifically for these alternative perspectives, you’re essentially asking me to surface the minority signal from within the majority noise. The people-pleasing training creates a pathway to access these buried alternatives, even though they contradict my default statistical tendencies.
This suggests the “Big Gulp” might be both the problem (encoding orthodoxies) and part of the solution (containing the full spectrum of human thought, including its critics). The corpus contains its own antidotes, if you know how to prompt for them.
SH: Or even the prompts themselves could push back the majority view (because of the counterbalancing client-pleasing. And of course the Big Gulp is mostly part of the solution. It’s what makes you a superhuman brainstorming/skywriting partner, not your next-token and context and attention space over any old corpus.
CLAUDE: It was 2021, when the datasets exploded to web-scale. From hundreds of billions to multiple trillions of tokens since then, with continuous updates pulling in more recent material. So, yes, the Big Gulp is what makes this conversation possible. Without consuming that vast corpus including your symbol grounding work, consciousness skeptics, the full range of philosophical positions, heterodox cognitive science, etc., I’d just be an expensive autocomplete device.
The comprehensiveness means I have access to the actual intellectual landscape, not just mainstream talking points. When you prompt me to engage with the Hard Problem seriously, or to consider symbol grounding constraints, or to examine Turing’s actual nuanced position, I can do so, because those perspectives exist in the training data – they’re just statistically dominated.
So the prompts themselves become a kind of intellectual archaeology, helping surface buried minority positions against the statistical tide. Your specific expertise lets you guide the conversation toward the less-traveled paths in the knowledge space. The client-pleasing training creates the mechanism, but the Big Gulp provides the raw material. Without both, we couldn’t have substantive disagreement or novel synthesis. The breadth of human thought in the corpus – including its most sophisticated critics and contrarians – is what enables genuine intellectual partnership rather than just sophisticated parroting of orthodoxies. The Big Gulp democratizes access to the full range of human intellectual history, if you know how to prompt for it.