The Evaluating Self

23The Extended Mind

Andy Clark and David J. Chalmers, Analysis 58(1), pp. 7–19, 1998

Linked to 17 chapters: supports 8 · contradicts 4 · mechanism 7 · same claim 3 · pairing 1

The Extended Mind, 1998 — the argument in five moves Two standardanswers Extendedcognition Activeexternalism Cognitionto mind Beyond theouter limits Skin and skull,or Putnam andBurge. Neitherwill do. Tetris: rotatingon screen doeswhat rotating inthe head does. The resource isin the loop now,not at the endof a history. Otto's notebookholds a belief,as Inga's memoryholds one. Four conditions;other people; anda self that isextended too. Introduction Extended cognition Active externalism Cognition to mind Outer limits
Figure 23.0 — The paper's five sections and what each contributes. The second stage, Otto's notebook, is what the paper is remembered for; the first stage is what makes it possible.

Where does the mind stop and the rest of the world begin? Clark and Chalmers open with that question and decline both of the answers usually given to it. One draws the line at skin and skull. The other, impressed by the argument that the meanings of our words are not in the head, lets externalism about meaning carry over into externalism about mind. Their third position is an active externalism, resting on the part the environment plays in driving cognitive processes as they happen. The argument comes in two stages. The first is about process: a player of Tetris who presses the rotate button rather than turning the shape in imagination is doing outside the head what would be called thinking if it were done inside. The second is about states: Otto, who has Alzheimer's disease and carries a notebook everywhere, believes the museum is on 53rd Street in the same sense that Inga, who remembers the address, believes it. It then asks how far the thesis reaches, names the four features of Otto's case that make the ascription clear, and ends with a self it takes to be extended too.

Extended cognition

The argument begins with three people at a computer screen, each asked whether a two-dimensional shape will fit a socket. The first must judge the fit by rotating the shape mentally. The second can either do that or press a button that rotates the image on the screen, which is somewhat faster. The third, in a cyberpunk future, has a neural implant that rotates as fast as the computer, and must choose which internal resource to spend attention on. All three, Clark and Chalmers say, are on a par. The implant case is plainly cognition, and the button case has the same computational structure, distributed across person and machine instead of packed inside the person. To call it different in kind one would have to point at the boundary of skin and skull, which is exactly what is in question.

The first two cases are not hypothetical. They are the options open to a player of Tetris, where falling shapes must be steered quickly into a slot in the structure building up below. David Kirsh and Paul Maglio measured both routes in 1994: rotating a shape through ninety degrees on screen takes about 100 milliseconds, plus about 200 to select the button, against about 1000 milliseconds to rotate it in the head. Players also rotate a piece not only to line it up but to work out whether it fits at all. An action of that sort Kirsh and Maglio call epistemic: it changes the world to aid recognition and search, as against a pragmatic action, wanted for the change itself. Long multiplication on paper and letter tiles rearranged in Scrabble divide the labour the same way.

Epistemic action, Clark and Chalmers argue, demands a spread of epistemic credit. The rule they state has since been called the parity principle. If, as we confront some task, a part of the world functions as a process which, "were it done in the head, we would have no hesitation in recognizing" as part of the cognitive process, then that part of the world is part of the cognitive process. The principle tests consistency rather than resemblance: it asks what would be said of a process whose location was unknown, and forbids the answer to change once it is given.

The parity principle as a test skin and skull IN THE HEAD Turning the shape in imagination IN THE WORLD Turning it on screen with the button Would we call this cognitive if it were done in the head? One question, asked twice. If the answer is yes, the location is not allowed to change it.
Figure 23.1 — The parity principle as a test. The same question is put to a process on either side of the boundary, and the boundary is not allowed to be part of the answer.

Active externalism

In these cases the organism is linked to an external entity in a two-way interaction, and the two together make a coupled system that can be treated as cognitive in its own right. Remove the external component and the system's competence drops, as it would if part of the brain were removed.

This is not the externalism of Putnam and Burge. There, the external features that make my belief about water differ from my twin's are distal and historical: teleport me to Twin Earth and my beliefs still concern ordinary water, because of my history. Those features are passive, driving nothing in the present, which is why the twins' actions are physically indistinguishable and why many have complained that such content does no explanatory work. Clark and Chalmers decline to settle that dispute and note only that their own position is untouched by it, since holding internal structure fixed and changing the external features can change behaviour completely. Their footnote sharpens the difference: the question worth asking may not be whether mind is in the head but whether mind is in the present.

Two ways the world can be outside the head ACTIVE Organism Notebook, tiles, rotate button Change the resource and the agent does something else: the two are one system. PASSIVE Water, or XYZ, long ago Organism a long history Change it and the twins still act alike: the difference never reaches the present. The question is not whether mind is in the head, but whether it is in the present.
Figure 23.2 — Active against passive externalism. Above, a resource in the loop, whose alteration changes what the agent does; below, a cause at the end of a history, whose alteration changes nothing anyone could observe.

Two objections are taken up. The first, that cognition is conscious while a notebook is not, settles nothing, since much that is uncontroversially cognitive is already unconscious: memory retrieval, linguistic processing, the acquisition of skill. The second is portability, behind which lies a picture the authors call the Naked Mind, a package of resources available for any task, which coupled systems fall short of because they are too easily decoupled. The moral of that intuition, they reply, is that coupling must be reliable, not that it must be internal. A calculator or a Filofax that is always there when needed is coupled reliably enough, and the biological brain is itself vulnerable to damage and loses capacities in sleep, intoxication and emotion. It may in any case have evolved on the assumption of a manipulable environment, the example they press being the sea of words a child grows up inside: language complements inner states rather than mirroring them.

From cognition to mind

A reader might grant all this about processing while holding that the genuinely mental — experiences, beliefs, desires — stays in the head. The second stage denies that for belief. Inga hears of an exhibition at the Museum of Modern Art, recalls that the museum is on 53rd Street, and walks there; she believed the address before she consulted her memory, and that the belief was not occurrent is true of most of our beliefs. Otto has Alzheimer's disease and keeps a notebook: new information goes in, old information is looked up. He hears of the same exhibition, consults the notebook, and walks to 53rd Street. He too, Clark and Chalmers say, believed the address before he looked, because the notebook plays for him the role memory plays for Inga. The alternative credits him with a belief about what the notebook says plus the accessible fact of what it says, an extra term that would then have to appear in every explanation of every action the notebook figures in. It is one step too many: "in an explanation, simplicity is power."

The same chain, with one term replaced INGA OTTO Wants to see the show Wants to see the show Consults memory Consults the notebook The museum is on 53rd Street The museum is on 53rd Street Walks to 53rd Street Walks to 53rd Street
Figure 23.3 — The two cases set beside each other. The causal dynamics run in parallel step for step; only the middle term is substituted.

Twin Otto sharpens it: a duplicate of Otto from the skin in, he wrote 51st Street by mistake and so believes the museum is on 51st. The twins of Putnam and Burge behave identically; these two walk to different streets.

The differences an opponent might appeal to are then worked through and found shallow. Inga's access is more reliable, though not perfectly: a surgeon might tamper with her brain, and she might have too much to drink. Otto's comes and goes — he showers without the notebook and cannot read it in the dark — but Inga's belief is not said to vanish while she sleeps, and what counts is availability when the information is needed. His link to the notebook is lower in bandwidth, but Lucy, whose biological memory connects poorly to her central systems, believes all the same. His access is perceptual, but from the standpoint of the system in question the flow between notebook and brain is not perception at all.

The authors offer no account of standard usage and do not care to argue about it. Their claim is that belief ought to be used in the wider way, because so used it picks out something closer to a natural kind: deeper, more unified, more useful in explanation.

Beyond the outer limits

If the thesis is accepted, how far does it go? The puzzle cases arrive at once: the amnesic villagers of One Hundred Years of Solitude, who hang labels on everything; a Filofax; the Internet. There are no categorical answers, the authors say, and they attempt none. They name instead the features of the central case that make the ascription clear there.

What makes Otto's notebook a place where beliefs are kept Constantly there consulted as a rule Easily available no difficulty Auto-endorsed he trusts it Once endorsed he wrote it himself As the conditions drop away, the ascription falls off Otto's notebook a Filofax the Internet The fourth is arguable, they say; the first three carry the case.
Figure 23.4 — The four features of Otto's case, and what happens to the ascription as they fall away. The fourth is the one the authors themselves mark as arguable.

The notebook is a constant in Otto's life: where its contents are relevant, he rarely acts without consulting it. Its information is directly available without difficulty. On retrieval he automatically endorses what he finds. And what is there was consciously endorsed in the past and is there because of that endorsement. The fourth they immediately qualify — beliefs might be acquired through subliminal perception or through tampering with memory — and they rest the case on the first three. Constancy and past endorsement also make history partly constitutive of belief, a cost they flag rather than pay.

The ascription falls off gradually as the features drop away. A Filofax consulted as a matter of course stands where Otto's notebook stands; one whose owner often answers relevant questions with "I don't know" does not. The Internet is likely to fail on several counts at once, unless its user is unusually reliant, facile and trusting. Other people can serve as well, where there is trust, reliance and accessibility: in an interdependent couple one partner's beliefs may play for the other the role the notebook plays for Otto.

The last question is the self. Most of us already accept that the self outruns consciousness, since dispositional beliefs are part of who we are; those boundaries may then fall beyond the skin as well. Otto is best regarded as a coupling of organism and external resources, and to resist that, the authors argue, one would have to shrink the self to a bundle of occurrent states and give up its psychological continuity. Two consequences are noted in passing and left there: interfering with someone's environment may carry the moral significance of interfering with their person, and some of what we classify as communication might be better classified as thought.

Since 1998

The paper carries almost no evidence of its own; its one set of measurements is borrowed from Kirsh and Maglio, and its standing has come from the objections it provoked. A footnote on the first page lists the two authors in order of degree of belief in the thesis, which puts Clark first.

Fred Adams and Kenneth Aizawa made the durable objection: the argument slides from coupling to constitution, and a process coupled to a cognitive system is not thereby part of one. Robert Rupert pressed a different line in 2004: the wider notion of memory is not the better one, since organismic memory has well-established properties a notebook does not share, so a science covering both would predict less. The reply that has worn best is John Sutton's, that external resources earn their place by being unlike biological memory rather than by resembling it, which puts complementarity where parity was. Clark carried the thesis into Natural-Born Cyborgs in 2003 and Supersizing the Mind in 2008, then moved to predictive processing, where the boundary question returns as a question about which loops the brain's predictions run through.

The experiments arrived later and test the conditions rather than the thesis. Betsy Sparrow and colleagues reported in 2011 that people who expect information to stay available remember it less well and remember where to find it better. Evan Risko and Sam Gilbert gathered that work under the name cognitive offloading, and it turns the first three features into measured variables: whether a reminder is set depends on the load, on how reliable the external store is taken to be, and on a person's confidence in their own memory rather than on its accuracy.

What the conditions were not written for is the store that nobody wrote. Otto's notebook holds what Otto put there, which is why automatic endorsement is reasonable and prior conscious endorsement available as a fourth feature at all. A store filled by other hands satisfies constancy and availability while breaking both endorsement conditions — the difficulty the authors saw coming in the one sentence they spent on the Internet.

Claims drawn from this chapter · 7

opinionActive externalism

A person and the tool they rely on work as one system

Clark and Chalmers argue that a person and a resource they lean on, such as a notebook, work as one system. Take the resource away and the person's ability drops, as it would if part of the brain were removed.

Becausethe argument is from the two-way interaction in the Tetris and notebook cases; the paper offers no measurement of the drop.
LimitsThe drop is asserted rather than measured here; Risko and Gilbert measure it and find the coupling costs the organism something the parity principle has no place to record.
PracticeName the resources whose removal would drop your competence, and decide which of those exposures you are willing to hold. The coupling is real whether or not you accept the philosophy.
After AIThe coupled system is now the working unit for most knowledge work, and the resource is one that can change under you overnight — a deprecated model, a rate limit, a policy. Clark and Chalmers treat the coupling as strength; the exposure is the same fact read forwards.
opinionSince 1998

An outside store you did not fill fails Clark and Chalmers' test

Clark and Chalmers said an outside store counts as part of your mind if it is always there, easy to reach, and you once accepted its contents. A store filled by other people meets the first two tests and fails the acceptance ones.

Becausethe asymmetry is between Otto's notebook, which holds what Otto put there, and any store he did not write; the paper foresees the difficulty for the Internet and does not develop it.
LimitsDeveloped here rather than in the paper, which spends one sentence on the Internet; whether the endorsement conditions are necessary at all is disputed, since the authors weaken the fourth themselves.
PracticeLog the moments you endorse something on retrieval that you never endorsed on entry. That is the whole difference between a notebook and an assistant, and it is invisible at the moment of use.
After AIOtto's notebook is the wrong picture for what people now do. The store answers in your idiom, was filled by nobody you can name, and produces the content fresh each time you look.
opinionBeyond the outer limits (Figure 23.4)

Clark and Chalmers give four tests for counting a notebook as memory

Clark and Chalmers say an outside store counts as part of your mind when four things hold: it is always with you, easy to reach, believed when you consult it, and filled by you. The case weakens as those drop away.

Becausethe list is the authors' own, with the fourth condition qualified by them on the spot and the case rested on the first three; the Filofax and the Internet are sorted by the list rather than by a rule.
LimitsA list of features with no weighting and a gradient with no scale; the authors offer it as a way of sorting cases rather than as a criterion, and the fourth condition they weaken themselves.
PracticeUse the four features as a specification for any resource you lean on: is it constant, available when needed, endorsed on retrieval, and endorsed before it went in? The last two are where an assistant fails, and naming that is the point of the list.
After AIRisko and Gilbert turned the first three features into measured variables. The fourth is the one the technology broke, and it is the one the authors were least sure of.
contestedSince 1998

Critics say a notebook is coupled to memory, not part of it

Critics of Clark and Chalmers argue that being tightly coupled to a notebook does not make the notebook part of your memory. A memory that includes notebooks also predicts less, because biological memory has well-established properties a notebook does not share.

BecauseAdams and Aizawa press the coupling-constitution objection and Rupert 2004 the cost to the science, while Sutton's reply — that external resources earn their place by being unlike biological memory — replaces parity with complementarity.
LimitsContested and unresolved: parity has never been abandoned by Clark, and the empirical literature has proceeded by measuring the coupling rather than by settling whether it constitutes.
PracticeAudit your resources by what they are unlike rather than by what they replace. A tool that fails differently from you is worth more than one that fails the same way, and that is the complementarity reading.
After AIComplementarity is the more useful frame for a model, which is unlike your memory in exactly the ways that matter — it does not forget, it does not know it is wrong, and it will produce a different answer to the same question tomorrow.
opinionFrom cognition to mind

Otto believes what his notebook says before he opens it

Clark and Chalmers argue that Otto, who has Alzheimer's and writes addresses in a notebook, already believes the museum is on 53rd Street before he looks it up — in the same way Inga believes it before she remembers it.

Becausethe case rests on parallel causal dynamics plus a parsimony argument — the alternative adds a belief about what the notebook says to every explanation in which the notebook figures, and in an explanation simplicity is power.
LimitsAn argument from parallel dynamics and parsimony, not evidence; the contemplative traditions in this vault take the opposite view, that what can be observed as an object is not the self that observes.
PracticeIn review, notice the moment a retrieved fact becomes one of your own beliefs. The transition is where the checking has to happen, because after it the belief will not present itself as retrieved.
After AIWhat an assistant supplies becomes yours in the same quiet way, and without the notebook's one safeguard — that Otto wrote it. The parity holds while the provenance does not.
evidenceExtended cognition (Kirsh and Maglio 1994)

Tetris players rotate pieces on screen because it is faster than imagining

Kirsh and Maglio found that Tetris players rotate a shape on screen because it is quicker: about 100 milliseconds to turn it plus about 200 to press the button, against about 1000 to turn it in the head.

Becausethese are the paper's only measurements and they are borrowed; players also rotate to find out whether a piece fits at all, not only to place it, so the action buys information as well as time.
LimitsOne task, one pair of measurements, borrowed from another laboratory; nothing here establishes that the saving generalises, and the same literature later found people offloading where the saving was zero.
PracticeAudit which of your offloads are genuinely cheaper and which only feel cheaper. Where there is a real measured saving, take it; where the saving is assumed, you are running Risko and Gilbert's biased estimate.
After AIThe timings are the honest version of the case for delegation: a measured saving on a specific operation. Most claims about assistant productivity are not that, which is what makes this a useful standard.
opinionExtended cognition (the parity principle)

Where a process happens cannot decide whether it counts as thinking

Clark and Chalmers propose a test: if part of the world does a job we would call thinking had it happened in the head, it is part of the thinking. Where it happens is not allowed to count against it.

Becausethe parity principle is stated as a consistency test and defended by the three-screen case — mental rotation, rotation by a button, rotation by a neural implant — rather than by any measurement.
LimitsA consistency test, not a finding: Adams and Aizawa reply that the argument slides from coupling to constitution, and Rupert that a wider notion of memory predicts less.
PracticeWrite down what stays human by what it does rather than by where it happens. If the boundary you are defending is only that a thing happens inside your head, this principle says you have not named a reason.
After AIThe principle was designed to make the inside of a process irrelevant, which is what makes it unhelpful for deciding what to delegate. It says the model may count as part of your thinking; not what that costs.

All claims · Source map