Contribution
The paper gave the opposition to LIT-097 its standard form. It names the alternative (the hypothesis of embedded cognition, HEMC). It makes the burden explicit: HEC must earn its place against HEMC by explanatory power, because conservatism otherwise favours HEMC. And it tests the claim against the experimental psychology of memory rather than against intuitions.
Key insight
Everything the extended-mind cases show about how heavily thinking leans on the world can be said by an organism-centred theory that treats the world as the thing being leaned on. So the extended view must show that a category spanning brain and notebook does real explanatory work. Memory research suggests it does the opposite: the differences between remembering and looking things up are what memory science studies.
Assumptions
- HEC is a substantive empirical hypothesis. The bare possibility of extended cognition is granted and set aside (§I). HEC is construed as literal mereological extension: cognitive states “literally comprise—as wholes do their proper parts—elements in that environment” (p. 1).
- Individual subjects only. Hutchins’s socially distributed cognition is excluded (§I).
- Conservatism. Other things equal, prefer the hypothesis that needs no new ontology (J. pp. 395, 405–06, by the author’s outline).
- Explanatory kinds are those that support induction and explanation in science (note 43). Rupert avoids “natural kind”, so that microstructural essentialism cannot settle the matter trivially.
- Memory science is a fair test. Memory is “a fundamental cognitive process, subserving virtually all other important cognitive functions” (§V), including Otto’s kind of standing belief.
Key results
- HEMC versus HEC (§II). HEMC: cognitive processes “depend very heavily, in hitherto unexpected ways, on organismically external props”. HEC: the unit of analysis is organism plus environment. Inferring HEC from HEMC by “epistemological dependence” is invalid: understanding Nazi Germany’s invasion of Poland needs 1920s German economics, which is not part of the invasion.
- Content externalism is a different thesis (§III). Neither entails the other. Russellian content can seem to put the horse in the belief, but McGinn’s view shows one can hold worldly constituents of content without external mereological parts of the mind. HEC is about “nuts-and-bolts” processing, not content.
- The past-endorsement dilemma (§IV). Directory assistance satisfies LIT-097’s first three criteria, so without the fourth, Otto believes every listed number: “It is absurd”. Keep past endorsement, and it is arbitrary to require it for extended but not internal belief (Clark and Chalmers’s own worry), and, worse, it makes internal consciousness “the ultimate source of cognitive authority”, so the notebook is “best accommodated by HEMC”. Note 34 adds that the constancy criterion is too strong even for internal beliefs. Note 36 answers Lycan’s suggested fix (the subject caused the marks) and Clark’s “personal tailoring”.
- Conversation (§V). Holding a model of a long conversation needs internal working memory (Ericsson and Kintsch’s long-term working memory). An interlocutor keeping it on paper would “destroy the dynamics of normal conversation”. Rowlands’s view that working memory is mostly external fits the data poorly.
- Negative transfer (§V). In A-B/A-C paired-associate learning, internal memory shows interference. A written list shows none, and no learning curve: lists are “learned” at once. “List-losing curves” will not match forgetting curves. The differences are in learning time and access time, “at the very heart” of memory research.
- Not malfunction (§VI). The generation effect (Bobrow and Bower 1969), a success of memory, has no external analogue: a subject consulting a notebook gets the same answer whoever wrote the context sentence.
- Generic memory (§VII). A kind like “has some sort of access to the information that P” would cover notebooks. But internal systems share power laws of learning and forgetting and Rescorla-Wagner learning that external stores lack. Psychologists who split memory into systems tend to drop the superordinate kind, not widen it. The man who lives in the library does not “know” its contents in any sense that explains his behaviour better than an embedded account.
- Functionalism (§VIII). A three-premise functionalist argument for HEC is valid, but premise 2 (some realisers are external) fails both ways. Commonsense functionalism forbids seeing one’s own memory (“Yesterday I saw my memory of last week’s trip to the beach”). Psychofunctionalism takes its roles from memory science, whose roles external stores do not fill.
- Cognitive systems (§IX). Taking extended systems as the objects of cognitive science sits badly with AI, which wants self-contained systems that work across environments, and with developmental psychology, which explains the skills of “a single coherent system, one with historical integrity”. The conclusion: “HEMC provides the best interpretation” for now, and “without the strength of cognitive science’s successes supporting it, the hypotheses of extended mind and self seem weak indeed.”
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | HEMC is a distinct competitor to HEC, and conservatism favours it unless HEC adds explanatory power | moderate | §II; the conservatism principle is assumed |
| C2 | Epistemological dependence does not imply constitution | strong | counterexample (Poland) |
| C3 | HEC is independent of content externalism | strong | §III, McGinn and the Russellian case |
| C4 | Without past endorsement, LIT-097’s criteria make phone directories into beliefs | strong | directory-assistance case |
| C5 | With past endorsement, the organism is privileged and HEMC fits better | moderate | the endorsement is internal; the inference to HEMC rests on conservatism |
| C6 | External storage behaves differently from internal memory in conversation, negative transfer and the generation effect | moderate | cited memory research, with thought experiments for the external side; no experiment on external stores is run or cited |
| C7 | A generic memory kind would do no causal-explanatory work | weak to moderate | challenge plus the library case; the coherence of internal systems (power laws) |
| C8 | Functionalism does not support HEC | moderate | §VIII dilemma between commonsense and psychofunctionalism |
| C9 | AI and developmental psychology favour organism-bounded systems | weak | §IX, illustrations of research aims |
Concepts
- HEC — the hypothesis of extended cognition, “human cognitive processing literally extends into the environment”.
- HEMC — the hypothesis of embedded cognition: heavy, unexpected dependence on external props, with the organism as the cognitive system.
- conservatism — prefer the theory with fewer new commitments, other things equal.
- past-endorsement criterion — LIT-097’s fourth feature, which the dilemma turns on.
- generic memory — a coarse kind, “a subject remembers P if and only if the subject has some sort of access to the information that P”.
- internal privilege — the organism’s consciousness as the final authority on what counts as belief.
Connections
- LIT-097 (NOTE-180). The four criteria are quoted from LIT-097. C4 is the overgeneration case NOTE-180 asked about (a contacts list), done with directory assistance. Rupert reads LIT-097 as “a clear functionalist strain” and attacks that strain in §VIII, as NOTE-180 found role functionalism carrying the conclusion.
- Adams and Aizawa (LIT-438). Note 87 says this critique was developed independently, rejects their derived-content argument as question-begging, and agrees with their worry that internal and external processes have different cognitively relevant properties.
- Clark (LIT-403, LIT-412). Clark’s 2005 reply does not mention Rupert. His later book takes Rupert’s labels for a chapter title. Clark’s 2005 face-memory case, aimed at the active-memory critics (Weiskopf and others), answers the kind of difference argument made here.
- Chalmers. The acknowledgements thank Chalmers for comments on drafts and for feedback at the 2002 Pacific APA. The 2019 chapter (LIT-415) answers the “different psychological profile” objection, citing Rupert’s 2009 book, as chauvinism, and says Rupert accepts extended-circuit cases.
- The extended self. Note 6 quotes LIT-097’s §5 and argues that criticising HEC “strike[s] a blow” to the extended self. LIT-136 (Floridi) and LIT-138 (Register) are this record’s other readings on that claim.
Bearing on the record
- THEORY-043. Two side remarks touch it. Note 38 says corporate bodies have no plausible theory of consciousness and no paradigm conscious acts, so a socially distributed belief cannot rest on conscious endorsement. And §I’s exclusion of group cognition, because there “many of us … lose our bearings”, is the intuition LIT-159 challenges. Neither is an argument about phenomenal consciousness.
- AI. §IX’s claim about AI’s aims is a 2004 claim; it is not ML practice. Nothing for the anthology.
Limitations
- The empirical case concerns fine-grained kinds. A defender can grant every difference and claim a coarse kind. Rupert’s reply to that (§VII) is a challenge plus conservatism, not a demonstration.
- No experiment on external memory is cited or run. The external side of the comparisons is argued by thought experiment (“there is no reason to think…”).
- The dilemma’s second horn needs conservatism to get from “the organism is privileged” to “prefer HEMC”.
- Only the author’s version was read. Journal page numbers are given only where the author’s outline supplies them.
Open questions
- Would a cognitive science built on generic kinds predict anything an embedded account cannot? Rupert’s challenge is to name one result.
- Does any external store show interference, a generation effect or power-law forgetting? A digital store with fuzzy, reconstructive retrieval might. That is the experiment the paper’s argument invites.
Corrections
- none to a seeded skim (there was no seed)
- Page references. The author’s outline cites journal pages (395–97, 401–07, 402–05, 407, 418–19, 424, 428); the section-to-page mapping above is only as good as that outline.