Contribution
It gives behaviourism about intelligence its best form, a capacity to respond sensibly to any input sequence, with “sensible” granted, and refutes it by a single counterexample. That shows the standard objections to behaviourism had left the sufficient-condition claim untouched, and it establishes that the concept of intelligence is partly etiological.
Key insight
Behaviour is evidence of intelligence only because we assume it was produced by something that thinks of the responses. Store the responses in advance and the same behaviour, actual and counterfactual, comes from a process “of the most elementary sort” (n. 21). Echoes of intelligence are not intelligence, whether the echo comes from programmers or from a cosmic accident.
Assumptions
- Sense of “intelligent”: “the possession of thought or reason”, not the comparative sense (p. 8).
- Concession: “sensible” can be defined non-question-beggingly (pp. 10–11), “an enormous concession to behaviorism”.
- Finiteness (pp. 19, 23–24): a test of bounded length has finitely many typable strings; quantised senses make the same true of a whole robot’s stimuli, so the argument covers intelligence simpliciter. The empirical version (analogue-to-digital conversion makes no mental difference) is only mentioned; the conceptual version is that our concept allows an intelligent being with “dot”-picture eyes (p. 24).
- Modality (Objection 6): behaviourist analyses are conceptual, so logical possibility suffices; if the view is a law of psychology instead, a world with infinitely divisible matter and the same psychological laws still contains the machine (pp. 30–33, n. 25).
Key results
- Dispositions vs capacities (pp. 14–19). Perfect actors and super-super-spartans show no disposition is necessary for pain or intelligence; but “how could an unintelligent system perfectly pretend to be intelligent?”, so they leave sufficiency standing. Recast as a capacity, specified by conditionals with internal states in their antecedents, the conception escapes Chisholm–Geach, perfect actors and paralytics altogether (p. 18).
- The string-searcher (pp. 19–21). Interrogator types A; the machine picks a stored sensible string beginning with A and types its next sentence B; and so on. Tree-searching variant: one programmed reply per branch. Programmers may simulate “my Aunt Bertha”. Conclusion: the capacity to emit sensible responses “is not sufficient for intelligence”, and the neo-Turing Test conception, “along with the older and cruder Turing Test conceptions”, is refuted (p. 21).
- Conduit, not radio (p. 22). The programmers’ causal efficacy ends before the test begins; no signal from the interrogator reaches them. A memoriser of Chinese responses, or Luria’s mnemonist fed strings by philosophers, has the capacity without the intelligence.
- Objections and replies (pp. 25–43).
- O1, every machine’s intelligence is its programmers’: no; learning machines’ intelligence “is its own”.
- O2, current events: irrelevant to intelligence.
- O3, a stipulated new meaning: the two-grandmaster chess trick shows our concept already treats echoed performance as misleading; natural-kind terms allow stereotypical Xs that are not Xs (Putnam).
- O3a, chauvinism: the claim is about “richness” of processing, not likeness to us; refutation by counterexample needs no definition, as a paper clip refutes a definition of life.
- O4, “what if you were one?”: like asking an atheist “what if you turned out to be God?” (pp. 28–29).
- O6, combinatorial explosion: about 10^1500 strings (Miller’s estimate, n. 24); answered by logical possibility, a law-of-psychology argument, Basic English, tree-search and hybrid tricks that “postpone, but do not avoid” explosion.
- O7, unbounded tests: humans cannot pass them either; idealising to unbounded competence requires mentalistic talk of memory and motivation no behaviourist can use; loop-searchers and ever-extended tapes answer the rest.
- O8, Dennett’s “averts exponential explosion” amendment: psychologistic already, ad hoc, and caught between postponing and avoiding.
- Simulators (pp. 40–42). Machines computing your particle trajectories, neurons or psychological mechanisms may avoid explosion and are “arguably” unintelligent: manipulating descriptions of cogitation “is hardly obvious” to be cogitation. A Chinese-speaking homunculus with a manual of your psychology makes the point. Block calls this part “incomplete”.
- Note 30. Lycan (“Form, Function, and Feel”, LIT-419) and Rosenthal objected to Block’s homunculi-head intuition as a forest-for-the-trees illusion; Block answers with a single homunculus attending to a column of the machine table. Some symbol-manipulating homunculi-heads are intelligent; this one is not because its states’ causal relations “do not mirror” ours.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | The capacity to respond sensibly to any input sequence of bounded length is not sufficient for intelligence | strong, given logical possibility suffices | the string-searcher, pp. 19–21; replies to O1–O8 |
| C2 | Psychologism is true: whether behaviour is intelligent depends partly on its etiology | strong as a consequence of C1 | p. 21 |
| C3 | The standard objections to behaviourism refute only necessary conditions | moderate | pp. 14–15, argued case by case |
| C4 | Psychologism is not chauvinist | moderate | Martian example p. 6; p. 21: excludes a kind of processing, requires none |
| C5 | The machine may be nomologically possible | weak | infinitely divisible matter, pp. 31–32; “Perhaps” |
| C6 | No interesting type of processing is common to all unintelligent test-passers | weak, sketch | simulators, pp. 40–43; “will have to remain incomplete” |
| C7 | Some symbol-manipulating homunculi-heads are intelligent | assertion | n. 30, “While I cannot defend it here” |
Concepts
- Psychologism — whether behaviour is intelligent depends on the character of the information processing that produces it (p. 5).
- Neo-Turing Test conception — intelligence is “the capacity to produce a sensible sequence of verbal responses to a sequence of verbal stimuli, whatever they may be” (p. 18).
- Disposition vs capacity — input–output conditionals versus conditionals with internal states in the antecedent (pp. 16–17).
- Sensible string — a typable string interpretable as a conversation in which at least one party is sensible (p. 19).
Connections
- Turing (LIT-394, NOTE-355) is quoted (n. 13) accepting a sufficient condition; that is the condition refuted.
- “Troubles with Functionalism” (LIT-407, NOTE-366): the list-searcher was sketched there (pp. 294–295); this paper develops it. The homunculi-head intuition is defended in n. 30.
- Searle (LIT-400, NOTE-349): n. 30 anticipates the Chinese room and rejects its generalisation to all symbol manipulation.
- Dennett: named as near the neo-Turing view (n. 12) and as the source of Objection 8 (n. 29). LIT-440 and LIT-398 are his intentional-system papers in the record.
- Chalmers (LIT-404, NOTE-342) cites this paper for the giant lookup table his recorder-equipped automata resemble.
Bearing on the record
- THEORY-023 and Butlin et al. (LIT-056, NOTE-052) discount behavioural evidence for consciousness in systems trained on human output. This paper is the canonical argument that behaviour, even counterfactual behaviour, underdetermines the mental property when the system could be reproducing stored output. It concerns intelligence, not experience, and a lookup table, not a trained model, so it supports their premise by analogy.
- The
anthology-candidateflag on LIT-642 asks whether the anthology wants this as a source on what behavioural evaluation cannot show.
Limitations
- “Richness” of processing is named, not defined (p. 28).
- The extension from conversation to all behaviour rests on a quantisation claim argued only briefly.
- The simulator cases at the end are explicitly incomplete, and the intuition they rely on is the one Block admits is “too easily manipulable” (n. 30).
- The machine is a lookup over complete histories; whether its lesson transfers to systems that compress or generalise from stored text is not addressed.
Open questions
- What is the minimal processing condition that separates the string-searcher from an intelligent system, given that Dennett’s exponential-explosion condition fails?
- Does the argument transfer to systems that interpolate rather than look up, which are neither Block’s machine nor the “learning” machines of his reply to Objection 1?