Contribution
It sets out the computer model as a set of distinct claims, about intelligence, about meaning-respecting processing, and about intentionality, and shows they stand or fall separately. Its new pieces, relative to Block’s earlier papers, are the clean intelligence/intentionality split (future- versus past-oriented), the counterfactual answer to “is a wall a computer?”, and the “Reductionist Cruncher” against Stich.
Key insight
The software metaphor is three metaphors. “Functional analysis” explains intelligence by mechanical decomposition. The “syntactic engine” explains how a symbol-shuffler can respect meaning, by an isomorphism. Neither explains intentionality, which needs a theory of where symbols get their meaning, and a swamp-brain shows intelligence without it.
Assumptions
- Two kinds of definition: of the word versus of the thing, as with water and H₂O; intelligence is a functional kind, like a gene or a mousetrap (§11.1.2).
- Primitive processors are where cognitive science hands over to a “realization” science, and the boundary is as fuzzy as chemistry’s with physics (§11.1.4).
- The orthodoxy: Fodor’s language of thought, with meaning from functional role, including perception and action (§11.3).
- A protoscientific concept of belief: causally active, explicitly represented belief, distinct from the ordinary concept (§11.3.1).
Key results
- Turing test (§11.1.1). Practical and theoretical uses fail; as conceptual clarification it lacks a specification of the judge, and making the judge good at thinking about thinking turns the test into “A machine thinks if our best thinkers … think it thinks”, which is false. The Aunt Bubbles machine refutes supervenience of the mental on the behavioural (n. 1, crediting Block 1978, 1981 and White 1982, and Jackson’s name “Blockhead”). Turing’s test focuses on performance, not competence.
- Functional analysis (§§11.1.3–11.1.5). Multiplication decomposes into adding, decrementing and testing for zero; an adder into AND and XOR gates. Computational equivalence of non-primitive devices is not input–output equivalence: two multipliers with identical behaviour, one converting to binary, are not equivalent.
- Intelligence vs intentionality (§11.2). Intentionality without intelligence (thermostats on McCarthy’s view, tree rings, words on a page, with only derived intentionality); intelligence without intentionality (the swamp-brain).
- Syntactic and semantic functions (§11.2.1). The symbol function maps numeral pairs to numerals; the semantic function maps numbers to numbers; their isomorphism explains how the adder adds. Symbols are found by first mapping rational relations, then identifying states by function.
- Is a wall a computer? (§11.2.2). Searle’s isomorphism is “cheap” only if it need cover the actual run; it must cover what the machine “could have performed”. Whether something is a computer is “in part up to us”, but “not totally”.
- Dennett’s objections to the language of thought (§11.3.1). Infinite beliefs: protoscientific belief is explicit, so most people did not believe trees do not light up like fireflies until they read the sentence. Queen out early: no explicit rule, so no belief in the protoscientific sense, and predicting from it can mislead. Sister in Cleveland: belief is a holistic computational relation, not mere storage.
- Systematicity (§11.4). It is right for conscious thought, where external symbols can themselves be the combinatorial medium (arithmetic), but its application to unconscious processing is “more tenuous”; an imagery system might explain animals’ red/green symmetry.
- Reductionist Cruncher (§11.5). The argument that syntax undermines content would let physicochemistry undermine syntax and quantum theory undermine statistical mechanics; “nature has real joints at many levels”. Whether a vocabulary marks a genuine level is empirical.
- Chinese room (§11.6). The systems reply’s negative half: from “Bill does not understand Chinese” one cannot infer the system does not. Its positive half: two systems can share a body, as in multiple personality. The nine-to-five extension shows which system implements which. Block doubts the Chinese system is phenomenally conscious and suggests Searle’s real argument is that intentionality requires consciousness (Searle 1992).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Passing the Turing test is neither necessary nor sufficient for intelligence | strong for sufficiency | Aunt Bubbles machine; judge problem; First Turing Test |
| C2 | Intelligence can exist without intentionality, and conversely | moderate | swamp-brain; thermostats, words on a page |
| C3 | Computation requires a counterfactual isomorphism, so not everything implements every computation | moderate | X-OR gate example, §11.2.2 |
| C4 | Arguments that a lower level undermines a higher one prove too much | moderate | Reductionist Cruncher, §11.5 |
| C5 | The systematicity argument shows a language of thought for conscious thought, not for encapsulated modules | moderate | §11.4 |
| C6 | The systems reply is right against Searle’s internalisation move | moderate | §11.6 |
| C7 | The Chinese system may lack phenomenal consciousness | expressed doubt | §11.6, “My doubts arise from the idea that perhaps consciousness is more a matter of implementation” |
Concepts
- Primitive processor — a component whose operation is explained by a realization science, not cognitive science (§11.1.4).
- Syntactic engine driving a semantic engine — a device whose symbol transitions are isomorphic to rational relations among their meanings (§11.2.1).
- Original vs derived intentionality — content a system’s states have for it, versus content inherited from such systems (§11.2).
- Protoscientific belief — a causally active, explicitly represented belief (§11.3.1).
- Reductionist Cruncher — the observation that any argument from a lower level’s greater generality and grain to a higher level’s dispensability applies again one level down (§11.5).
Connections
- “Psychologism and Behaviorism” (LIT-642) is the source of the Aunt Bubbles machine; “Advertisement” (LIT-648) is cited for the circularity of Fodor’s ideal-conditions theory of truth conditions (pp. 657–660 there).
- Chalmers (LIT-404, NOTE-342) gives the full implementation theory that §11.2.2 sketches.
- Searle (LIT-400, NOTE-349); Turing (LIT-394, NOTE-355); Dennett’s intentional-system papers (LIT-440, LIT-398).
- Marr’s three levels appear in §11.5 as several content levels above syntax.
Bearing on the record
- THEORY-023 says architectural indicators presuppose computational functionalism. Block, who accepts the computer model for intelligence and intentionality, still doubts it for experience because experience may depend on implementation. That supports THEORY-023’s description of the dispute as one about what counts as evidence.
- NOTE-342’s verdict that the machine-table China brain is the weak target and the neuron-level system the strong one fits Block’s §11.1.4: computational equivalence of non-primitive devices is fixed by their internal organisation, not input–output behaviour.
- No instruction for machine-learning practice.
Limitations
- A survey: Stich’s argument is “a simplified version”, and the reply to Searle is the systems reply without CRS’s theory of meaning.
- The intelligence/intentionality split rests on the swamp-brain intuition, which Block says “arguably” lacks intentional states.
- The consciousness doubt is stated, not argued.
Open questions
- If consciousness depends on implementation, which features of implementation, and how would that be tested? The chapter leaves it at a doubt.