Contribution
The paper names a predictive strategy that everyone already used, the Intentional stance, and separates it from two others, the design and physical stances. It then puts the strategy to three uses. It gives a criterion for having beliefs that needs no answer about consciousness or composition. It gives an accounting rule, the loan of intelligence, for finding the hidden homunculi in a theory of the mind. And it locates the normativity of belief in the preconditions for ascribing it.
Key insight
To treat something as a believer is to bet that it is well designed for its ends and then to compute what a rational agent in its position would do. The bet can be good without settling what the thing is made of or whether it feels anything. But every time a theory uses that kind of description for a part of the mind, it borrows intelligence it has not yet explained, and the debt is repaid only when the part’s “reader” is analysed into something that needs no Intentional description.
Assumptions
- Intentionality is linguistic (fn. 1, p. 87). An idiom is Intentional if co-designative substitution fails to preserve truth or its objects resist ordinary quantification. Dennett calls this “apparently slapdash” and relies on consensus about cases.
- Optimal design is the anchor. Rationality “so far means nothing more than optimal design relative to a goal or optimally weighted hierarchy of goals” and constraints (p. 89).
- Natural selection warrants the assumption for organisms. With it, one can ascribe beliefs and desires to aliens from their needs and environment, “conceptually innocent anthropomorphizing” (pp. 92–93).
- Rationality includes logic. An Intentional system must be supposed to follow modus ponens or believe its corresponding conditional; there is a trade-off between rules and truths (pp. 94–95, fn. 5 on Carroll’s tortoise).
Key results
The paper argues; it proves nothing and reports no data.
- The three stances (pp. 87–90). Design-stance predictions assume no malfunction; Intentional predictions additionally assume optimal design, so they fail on “design weaknesses and lapses”. For the best chess programs only the Intentional stance is practical.
- Observer-relativity and pragmatism (pp. 87, 91). Being an Intentional system is relative to “the strategies of someone who is trying to explain and predict its behavior”; whether one calls what is ascribed “beliefs or belief-analogues or information complexes or Intentional whatnots makes no difference” to the calculation; one may “switch stances at will”.
- Possession of information is already Intentional (pp. 90–91). It is “epistemic possession”, not storage: the sheik who owns the Britannica but reads no English shows the difference.
- Common sense is Intentional for people and animals (pp. 93–94). Consistently failed predictions push us to a design stance, as with the insane.
- Imperfection degrades prediction (pp. 95–96). As logical truths are banished from a creature’s beliefs, Intentional prediction becomes “cumbersome and undecidable” and ends in the design stance. Moving to design-stance explanation is “the proper direction for theory builders”, since stopping at the Intentional level leaves intelligence unexplained.
- Loans of intelligence (p. 96). Endowing a state with content posits a reader; the loan must be repaid by “finding and analyzing away these readers or comprehenders”. Intentionality is the “medium of exchange”: “wherever a theory relies on a formulation bearing the logical marks of Intentionality, there a little man is concealed.”
- Behaviourism masks Intentional prediction (pp. 97–99). The mathematician in a Skinner box presses “35” before training; the mouse trained to take four steps would, on Skinner’s laws, jab at air if the bar moved; a man trained to pick the longer display would be immune to Müller-Lyer. The box leaves room for one motion and no false belief.
- The research programme (pp. 99–100). “Intentional theory is vacuous as psychology because it presupposes and does not explain rationality”, yet “our most promising tactic is to take out intelligence-loans … and then look for mechanisms” that pay them back. AI does exactly this. Game theory and economics are Intentional but normative, and economics succeeds “because individual men are in general good approximations of the optimal operator in the marketplace”.
- Unmetaphysical (p. 100). The concept abstracts from “composition, constitution, consciousness, morality, or divinity”; it is “much easier” to decide whether a machine is an Intentional system than whether it thinks or is conscious. “Are there mental treasures that cannot be purchased with Intentional coin?”
- Belief’s two norms (pp. 101–104). (1) Normally, if x believes p, p is true; (2) normally, if x avows p, x believes p. Hence “a false belief system is a conceptual impossibility” (p. 101), and a capacity for false communication would be “just an emission proclivity”. The norms break the belief–desire circle “in the long run and in general” (p. 103).
- Incorrigibility (pp. 103–106). When avowals are irrational, leaning on rationality overrides the speaker and leaning on avowal grants him irrational beliefs; “there is no coherent Intentional description of such an impasse”. Neither phenomenology nor neurophysiology breaks the tie, since both presuppose a rationality standard. Fixing a tolerated error rate would be an “Official Tolerance Rule”, confusing “the norm they follow with what gets by in their world” (p. 106).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Whether something is an Intentional system is relative to an observer’s predictive strategy, and adopting the stance is pragmatic | moderate | definitional (pp. 87, 91); the chess example shows usefulness, not that nothing more is at stake |
| C2 | The stance is independent of what a system is made of and of its consciousness | strong as a statement of the definition | p. 91, p. 100; it follows from how the criterion is framed |
| C3 | For the best chess programs of 1971, only Intentional prediction is practical | moderate | informal (pp. 89–91); no programs are named |
| C4 | Any content-ascribing theory of the mind takes out a loan of intelligence that must be repaid by analysing away its readers | moderate | argument from the economic metaphor (p. 96); it is a methodological rule, not a result |
| C5 | Skinnerian regularities are Intentional predictions disguised by a restrictive environment | moderate | three thought experiments (pp. 97–99) |
| C6 | Intentional theory is vacuous as psychology but not as normative theory (game theory, economics) | moderate | argument (pp. 99–100) |
| C7 | A false belief system is a conceptual impossibility; belief requires mostly true belief and mostly accurate avowal | moderate | evolutionary argument plus conceptual claim (pp. 101–103) |
| C8 | There is no principled empirical tolerance level for irrational belief; such a rule confuses norm with practice | weak–moderate | the chess analogy (pp. 105–106) |
Concepts
- Intentional system: a system whose behaviour can be (at least sometimes) explained and predicted by ascribing beliefs and desires (p. 87), relative to someone’s strategy.
- Design stance: prediction from functional design assuming no malfunction, “irrespective of the physical constitution” (p. 88).
- Physical stance: prediction from physical state and laws of nature; the only stance that predicts malfunction (p. 88).
- Intentional stance: prediction assuming optimal design, ascribing the information and goals the system ought to have (p. 90).
- Loan of intelligence: the debt a theory incurs by endowing a state with content, which posits an unexplained reader (p. 96).
- Epistemic possession: having information in the sense belief requires, not mere storage (p. 90).
Connections
It cites Brentano for Intentionality, Quine (Word and Object §13) on translating logical connectives, Carroll’s tortoise, Skinner and Quine as the behaviourists, Hintikka’s epistemic logic (fn. 7, fn. 12), Phillips Griffiths and Mayo on belief’s normativity (fn. 9), and Dennett’s own Content and Consciousness (1969) and “Mechanism and Responsibility” (fn. 2, “forthcoming”). It is reprinted as chapter 1 of Brainstorms (1978).
Bearing on the record
- THEORY-043. The bridge account says the downward and upward readings agree on agency and divide on experience. This paper supplies the agency side with a criterion explicitly blind to consciousness (C2), which is why agency and experience can come apart there at all. It does not apply the stance to groups: economics is Intentional because individuals approximate optimal agents, not because a market is one (p. 100). The loan rule (C4) is the constraint on the downward direction: Minsky’s agents (LIT-399) explain a mind only if their content is repaid in parts that need none. “Artificial Intelligence as Philosophy and as Psychology” (LIT-393), unread here, is where Dennett says how.
- THEORY-023 (AI consciousness). C2 is a reason to keep belief ascriptions to AI systems out of the evidence for their consciousness. The paper itself makes the separation, at the definition.
- NOTE-131 and LIT-159. Schwitzgebel’s §6 answers a pragmatic objection Dennett sent him in correspondence. This 1971 paper is pragmatic about belief and silent on ascribing consciousness, so it is not the source of that objection. True Believers’ requirement that a believer be “reliably and voluminously” predictable is the nearer ancestor.
- LIT-206 (Goldstein & Lederman). Their “objective” interpretationism is not relativised to an interpreter; this paper’s is (C1). The relativity is what True Believers (LIT-398) later withdraws.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- The criterion is stated, not defended against over-generation. The paper never asks whether a lectern or a thermostat qualifies; True Believers does.
- “Rationality” is optimal design relative to goals the predictor ascribes, so the criterion’s content depends on goal ascriptions it does not constrain.
- The evolutionary arguments for belief’s norms (pp. 101–102) are assertions about what selection would favour, with no evidence offered.
- Its picture of 1971 chess programs as inaccessible to design-stance prediction is asserted.
Open questions
- Can a loan of intelligence always be repaid by decomposition, or only sometimes? The paper states the debt; Brainstorms ch. 7 claims the repayment schedule.
- If the stance is pragmatic and observer-relative, is there any fact about whether a system “really” believes? The paper says the question is “misplaced” (p. 91); True Believers answers it differently.
Corrections
- The paper is the first printed statement of the stance. Its fn. 2 points to “Mechanism and Responsibility”, then forthcoming, and the 2017 Brainstorms publisher’s description calls that essay the first articulation. “Introduces” in the earlier draft of the LIT was too strong and has been changed.