Sven Erik Matzen

Software Architect | Cloud & Security Expert | AI-enabled Solutions

Twin Earth: Why Meanings Are Not in the Head

🎧 Listen to this article

Philosophy · 2026-11-03

EU label: fully AI-generated content Fully AI-generated article (no prior review).

The Hook: One Sentence That Turned a Discipline Around

Twentieth-century analytic philosophy has a handful of sentences you can quote without naming the source, because everyone knows them. One of them sits on page 704 of a paper Hilary Putnam published in the Journal of Philosophy in 1973. It reads:

"Cut the pie any way you like, 'meanings' just ain't in the head!"

The tone is deliberately casual, almost brusque. And that brusqueness marks the point at which something taken for granted began to slip – something Putnam traces all the way back to Descartes.

The assumption runs like this: if I understand a word, it is because something definite is present in my head – a concept, an image, a rule, a definition. Meaning is an inner possession. Two people with the same thing in their heads mean the same thing. That sounds not merely plausible but almost trivial. Where else could a meaning be, if not in the head that understands it?

Putnam's answer is more cautious than the slogan suggests – and that is important enough to say right here: he does not claim that meanings lie around somewhere outside the head, but that you cannot determine what a word or a thought refers to without referring to the world and to the linguistic community. This thesis travels today under the name semantic externalism (or, more broadly, content externalism). The Stanford Encyclopedia of Philosophy calls its significance "potentially far-reaching though often unclear and disputed." Part 6 of this article takes that caveat seriously.

Why should this interest you beyond the intellectual appeal? Because the question of where meaning sits has long since become a technical one. When a language model uses the word "water," what does it refer to? To H₂O, or to statistical neighborhoods in a text corpus? If two systems have identical weights but are deployed in different worlds, do they mean the same thing? If a contract speaks of "personal data" and two lawyers have the same thing in their heads but work in different legal jurisdictions – are they talking about the same thing? In 1973 Putnam built a thought experiment that sharpens exactly these questions. It is called Twin Earth.

This article takes you the whole distance: from the two silent assumptions of the classical theory of meaning through Twin Earth and its variants, via the division of linguistic labor and Kripke's rigid designators, on to Burge's social sharpening of the argument, to the internalist counterattack – and finally to what all of this means for models that see only form and no world.

Part 1: The Two Silent Assumptions

Putnam does not begin with a thought experiment but with a diagnosis. The theory of meaning, he writes, came to rest on two "unchallenged assumptions." He formulates them on page 700 as (1) and (2):

"(1) That knowing the meaning of a term is just a matter of being in a certain psychological state (in the sense of 'psychological state,' in which states of memory and belief are 'psychological states'; no one thought that knowing the meaning of a word was a continuous state of consciousness, of course). (2) That the meaning of a term determines its extension (in the sense that sameness of intension entails sameness of extension)."

Translated and stripped down:

Assumption 1 (internalism): To know the meaning of a word is to be in a particular psychological state. Not a state of consciousness – Putnam expressly includes states of memory and belief – but in any case a state that belongs entirely to me.

Assumption 2 (meaning determines extension): The meaning fixes what the word applies to. Same meaning, same set of things you are talking about. In the terminology: sameness of intension entails sameness of extension.

Both assumptions look harmless. Together they are not. Because together they claim: what goes on in my head fixes what I am talking about. Putnam's sentence immediately after the two assumptions is correspondingly terse: "I shall argue that these two assumptions are not jointly satisfied by any notion, let alone any notion of meaning." Not that one of the two is false, but that there simply is no notion that delivers both at once.

It is worth appreciating how radical this diagnosis is. Putnam does not say, "meaning is more complicated than we thought." He says: no notion satisfies both assumptions together. I am of the opinion that this is best read as the diagnosis of a conflation: what the tradition called "meaning" carries two roles that need not coincide – the role of inner understanding and the role of reference to the world. Putnam's program is to break that conflation apart.

In the longer follow-up paper The Meaning of 'Meaning' (1975), Putnam exposes a further, still deeper assumption – not a second name for assumption (1), but an assumption about what counts as a psychological state in the first place. He calls it the assumption of methodological solipsism: "When traditional philosophers talked about psychological states (or 'mental' states), they made an assumption which we may call the assumption of methodological solipsism." And he defines it this way:

"This assumption is the assumption that no psychological state, properly so called, presupposes the existence of any individual other than the subject to whom that state is ascribed."

A psychological state in the strict sense, on this assumption, presupposes the existence of no one and nothing besides the subject. Putnam even sharpens the point: traditionally the assumption was so strong that not even one's own body could be presupposed – a "disembodied mind" would have to be able to be in the same state. The assumption, he writes, is "pretty explicit in Descartes, but it is implicit in just about the whole of traditional philosophical psychology." So it is not a straw man of a single school.

Part 2: Twin Earth

Now the thought experiment. It has an almost childlike simplicity, and that is precisely where its force lies.

Imagine a planet identical to Earth except for one small thing. Putnam describes that small thing as follows:

"One of the peculiarities of Twin Earth is that the liquid called 'water' is not H₂O but a different liquid whose chemical formula is very long and complicated. I shall abbreviate this chemical formula simply as XYZ. I shall suppose that XYZ is indistinguishable from water at normal temperatures and pressures."

So on Twin Earth it is not H₂O that fills the lakes, rivers and glasses, but another liquid with a long and complicated formula that Putnam abbreviates as XYZ. At normal temperatures and pressures XYZ is indistinguishable from water. It quenches thirst, it boils, it freezes, it rains. The inhabitants of Twin Earth call it "water."

Now the decisive move. Putnam turns back the clock:

"Now let us roll the time back to about 1750."

Before 1750, no one on Earth knew the formula H₂O. The chemical composition of water was unknown – on both planets. Putnam picks two speakers:

"Let Oscar1 be such a typical Earthian English speaker, and let Oscar2 be his counterpart on Twin Earth."

Oscar1 lives on Earth in 1750, Oscar2 on Twin Earth in 1750. Both are typical English speakers. When they say "water," both have exactly the same ideas: a clear, thirst-quenching, tasteless liquid that flows in rivers and boils in pots. Every description Oscar1 can give of water, Oscar2 can repeat word for word – and vice versa. Their inner states are, insofar as "inner" means anything, identical. Putnam puts it by saying they "understood the term 'water' differently in 1750 although they were in the same psychological state."

And now the punch line. In 1750 Oscar1 calls water "water," and the word applies to H₂O, because on Earth the stuff is H₂O – even though nobody knows it. In 1750 Oscar2 calls XYZ "water," and the word applies to XYZ. The extension of the two words differs. From this Putnam draws the conclusion that is the core of the whole argument:

"Thus the extension of the term 'water' (and, in fact, its 'meaning' in the intuitive preanalytical usage of that term) is not a function of the psychological state of the speaker by itself."

Note the precision of the formulation: not a function of the psychological state of the speaker by itself. That is not rhetorical surplus, that is the exact strength of the claim. The psychological state does not suffice; something else is needed. And that something else is the world.

One important detail often gets lost: Putnam argues explicitly that the extension was fixed already in 1750. "Yet the extension of the term 'water' was just as much H₂O on Earth in 1750 as in 1950; and the extension of the term 'water' was just as much XYZ on Twin Earth in 1750 as in 1950." The chemical discoveries of the late eighteenth century did not change the meaning, they uncovered it. That is a substantive thesis, not merely an observation – and it is the point at which critics take aim (see Part 6).

The Second Variant: Aluminum and Molybdenum

Putnam builds the same machine again with two metals – he introduces it simply as a modification of the science-fiction story. This time it is not unknown chemistry but switched vocabulary that carries the difference:

"In particular, we shall assume that 'aluminum' pots and pans are made of molybdenum on Twin Earth. Finally, we shall assume that the words 'aluminum' and 'molybdenum' are switched on Twin Earth."

On Twin Earth the words are swapped: what is called "aluminum" there is molybdenum, and what is called "molybdenum" there is aluminum. Putnam stipulates that molybdenum pots and aluminum pots can be told apart only by an expert. Oscar1 and Oscar2 are again indistinguishable – neither is a metallurgist, neither has ever noticed a difference. And yet:

"Again we see that the psychological state of the speaker does not determine the extension (or the 'meaning,' speaking preanalytically) of the word."

This variant shifts the burden: it does not depend on whether water has a clean chemical essence (which Part 6 disputes), but on there being somebody in the linguistic community who applies the words correctly. So it buys that independence by depending all the more heavily on expert deference – a mechanism that is itself contested. Which brings us to the next building block.

The Third Variant: Elm and Beech

The finest version is the one in which Putnam does not have to build a spaceship, because he uses himself as the example:

"Suppose you are like me and cannot tell an elm from a beech tree."

Putnam admits – with a wink he already recorded in 1973, on page 704, as "(I blush to confess)" – that his concept of an elm is exactly the same as his concept of a beech: a large deciduous tree, elms are not beeches, beeches are not elms. That is all there is in him. And yet "elm" in his mouth refers to elms and "beech" to beeches. Then the doppelgänger:

"Moreover, suppose I have a Doppelganger on Twin Earth who is molecule for molecule 'identical' with me. [...] It is absurd to think his psychological state is one bit different from mine: yet he 'means' beech when he says 'elm,' and I 'mean' elm when I say 'elm.' Cut the pie any way you like, 'meanings' just ain't in the head!"

Here is the sentence, and here is its context. It falls not with the water but with the trees – in a case where the work is done not by chemistry but by the social web of word use.

Part 3: The Division of Linguistic Labor

If Putnam's concept of elm and of beech is the same, why does his "elm" nevertheless refer to elms? Because he does not speak alone. Putnam develops the thought with gold:

"Let us shift the example; consider gold. Gold is important for many reasons: it is a precious metal; it is a monetary metal; it has symbolic value (it is important to most people that the 'gold' wedding ring they wear really consist of gold and not just look gold); etc."

Gold is important, so many people need the word. But not everyone who needs it must be able to recognize gold. Putnam describes the linguistic community here with a workplace analogy: some people have the "job" of wearing gold wedding rings, others of selling them, and still others the job of telling whether something really is gold. From which it follows:

"But they engender a division of linguistic labor: everyone to whom gold is important for any reason has to acquire the word 'gold'; but he does not have to acquire the method of recognizing whether something is or is not gold."

That is the division of linguistic labor: acquiring the word and acquiring the method of recognition are two different things. I can speak about gold with full standing without being able to test gold, as long as somebody in my linguistic community can – and as long as, in case of doubt, I would fall back on their judgment.

Putnam turns this into a general hypothesis, which he sets in capitals:

"HYPOTHESIS OF THE UNIVERSALITY OF THE DIVISION OF LINGUISTIC LABOR: Every linguistic community exemplifies the sort of division of linguistic labor just described; that is, it possesses at least some terms whose associated 'criteria' are known only to a subset of the speakers who acquire the terms, and whose use by the other speakers depends upon a structured cooperation between them and the speakers in the relevant subsets."

Every linguistic community, on this hypothesis, possesses at least some terms whose criteria are known only to a subset of speakers, and whose use by everyone else rests on a structured cooperation with that subset. Note how modestly the hypothesis is phrased – "at least some terms," not "all concepts." It does not claim that meaning is always social, only that every language has places where it is.

To anyone working in IT this should sound familiar. The meaning of "personal data," "critical infrastructure," "zero-day" or "production sign-off" is, in no organization, entirely inside any single head. It is distributed across lawyers, auditors, operators and documents, and ordinary use rests on a willingness to defer when in doubt. Putnam's hypothesis is, note, formulated descriptively – every linguistic community exhibits such terms – and not as a condition an organization would have to satisfy. I am of the opinion that it can nevertheless be read as a diagnostic tool: on this view deference is not a lack of competence but its normal case.

Part 4: Indexicality and Kripke's Rigid Designators

That leaves the question: how exactly does the world get into the meaning? Putnam's answer is surprising. He treats natural-kind terms as hidden pointing words.

"Our theory can be summarized as saying that words like 'water' have an unnoticed indexical component: 'water' is stuff that bears a certain similarity relation to the water around here."

"Water," on Putnam's account, does not mean "clear thirst-quenching liquid" but: the stuff that stands in the right similarity relation to the water here. The "here" is the indexical core. This quotation is on page 709. The relation itself Putnam had already introduced seven pages earlier, on page 702, immediately after the Twin Earth case: he calls it sameL ("same liquid as") and stresses its theoretical character:

"The key point is that the relation sameL is a theoretical relation: whether something is or is not the same liquid as this may take an indeterminate amount of scientific investigation to determine."

Whether two portions are "the same liquid" is not a question of appearance but an open-ended scientific question. The meaning of an everyday word is thereby tied to a research process whose outcome speakers need not know – and in 1750 did not know.

At this point Putnam's analysis meets Saul Kripke's. In 1970 Kripke delivered three lectures at Princeton University whose transcript appeared in 1972 in Semantics of Natural Language, edited by Donald Davidson and Gilbert Harman (Dordrecht: Reidel, pp. 253–355 and 763–769), and in 1980 as the standalone book Naming and Necessity from Harvard University Press. Kripke's central tool there is the rigid designator. The Stanford Encyclopedia of Philosophy ("Rigid Designators") characterizes it as follows:

"A rigid designator designates the same object in all possible worlds in which that object exists and never designates anything else."

Kripke himself makes clear that no necessary existence is being claimed: "All I mean is that in any possible world where the object in question does exist, in any situation where the object would exist, we use the designator in question to designate that object." (This sentence is not from Naming and Necessity; the SEP quotes it from Kripke 1971, p. 146 – the paper Identity and Necessity.)

Putnam sees the connection and names it himself:

"Kripke calls a designator 'rigid' (in a given sentence) if (in that sentence) it refers to the same individual in every possible world in which the designator designates. If we extend this notion of rigidity to substance names, then we may express Kripke's theory and mine by saying that the term 'water' is rigid."

And on page 710:

"It should be clear, however, that Kripke's doctrine that natural-kind words are rigid designators and our doctrine that they are indexical are but two ways of making the same point."

Two ways of making the same point: Putnam via indexicality ("stuff like the stuff here"), Kripke via rigidity ("the same in all worlds"). Putnam ascribes the surprising effect to this shared theory – not to an inference of Kripke's – and names Kripke as its first observer: "What Kripke was the first to observe is that this theory of the meaning (or 'use,' or whatever) of the word 'water' (and other natural-kind terms as well) has startling consequences for the theory of necessary truth." The consequence: there are necessary truths that can only be discovered empirically. "Hesperus is Phosphorus" is necessarily true because both names point rigidly at the same planet – but that they do so is something astronomy had to find out. The category of the necessary a posteriori, the SEP writes, was before the debate about rigidity "generally thought to be an empty category."

Caution is called for here, and I think it is important to say so. The SEP expressly marks this step as contested: "So it is controversial, though common, to say that rigidity assures that the proposition Hesperus = Phosphorus is necessarily true and a posteriori." The version about statements counts as less problematic than the one about propositions. So anyone presenting "water is H₂O is necessary a posteriori" as a settled result is overreaching. It is a widespread and influential position, not a theorem.

Part 5: Burge's Sharpening – When the Community Fixes the Content

Putnam's argument concerns words and their extension. Tyler Burge turns the screw further: it concerns thoughts too. His paper Individualism and the Mental appeared in 1979 in Midwest Studies in Philosophy, volume 4 (the SEP gives the page range as 73–121, PhilPapers as 73–122).

Burge's case needs no second planet, only a different dictionary. It runs as follows (I follow the SEP's presentation):

The actual case. Larry believes he has arthritis in his thigh. This is false, since arthritis affects only joints. His belief is nevertheless an arthritis belief – precisely because he lives in a linguistic community whose members use "arthritis" that way and would defer to experts if in doubt. So Larry has a false belief about arthritis.

The counterfactual case. Larry is, individualistically considered, entirely identical – molecule for molecule, memory for memory, behavioral disposition for behavioral disposition. Only his community differs: there, physicians and lexicographers use "arthritis" for inflammations of the joints and various other rheumatoid ailments.

The result. In this second case Larry's belief about his thigh is true – but on the SEP's presentation it is "not an arthritis-belief"; it is a belief about something else, which the literature calls "tharthritis." Since Larry is internally identical in both cases, the content of his beliefs varies with his community.

The difference from Putnam is crucial and often blurred: in Putnam's case nature makes the difference (H₂O vs. XYZ), in Burge's the linguistic community – and the concept is not a natural-kind term but one about which Larry is simply incompletely informed. The SEP in fact traces the social version back to Putnam himself (to the elm/beech case) and writes that "this social version of content externalism was later developed, powerfully, by Burge (1979, 1986)" – so Burge built it out rather than invented it. And it notes a further difference: Burge "explicitly endorses" extending the argument to propositional attitudes – Putnam does not.

Part 6: The Counterattack – What Speaks Against the Argument

An article that merely presents a philosophical thought experiment has done half the work. The more interesting half is where it breaks.

Objection 1: The Point Is Weaker Than the Slogan Sounds

The mildest objection is also the most consequential: the slogan overextends the conclusion. The SEP ("Externalism About the Mind") speaks of "some initially over-exuberant formulations of Putnam (1975) of the 'Cut the pie any way you like, meaning just ain't in the head', variety" – so of some exuberant formulations, not of Putnam's position as a whole – and makes clear what is really at stake: the conclusion "does not concern the location of (some) mental states but, rather, a claim about their individuation. External individuation does not entail external location."

That is a fine but consequential difference. The argument does not show that meanings are lying around somewhere outside the head. It shows that you cannot individuate mental states – tell them apart and type them – without referring to the environment. My thought can take place entirely in my head and still be identifiable only by reference to my world. Anyone who reads "meanings are not in the head" as a thesis about location reads it wrongly.

The SEP goes further and makes the dependence on a prior assumption explicit: Putnam's answer "will be found inviting by anyone who thinks of meaning as a matter of reference, or truth-makers, or truth-conditions." Which means: those who understand meaning primarily as reference to the world find the argument convincing. Those who understand meaning primarily as what a speaker masters can dispute the intuition. The thought experiment does not settle the question independently of that prior decision.

Objection 2: Narrow Content

The systematically developed counterposition is called narrow content. Its strategy, according to the SEP, consists in "accepting the force of those arguments but limiting their significance": yes, there is a wide, world-involving content – but alongside it there is a narrow one that depends only on what happens inside the individual, and this narrow content is the psychologically interesting one. In the years right after 1975 people mostly thought of that as phenomenal content (the SEP names Fodor 1981, 1982 and McGinn 1982). Jerry Fodor pursued a different strategy in Psychosemantics (1987): physically identical individuals have different wide contents but identical narrow contents – where narrow contents are "functions from contexts to wide contents." (Important: the SEP presents these positions as different strategies, not as a change of mind on Fodor's part; anyone speaking of a "rethinking" is inventing a narrative.)

The same basic idea carries two-dimensional semantics, which the SEP links to Lewis 1981, Jackson 2007 and Chalmers 2012: one separates a dimension that depends on the environment from one that does not. Not everyone thinks this is workable – Yli-Vakkuri and Hawthorne (2018) argue, according to the SEP, that "there is no form of content that can be easily cordoned off from the world."

Objection 3: Self-Knowledge

An objection from an entirely different direction concerns self-knowledge. We normally assume that I have privileged access to my own thoughts – that I know what I think without investigating the world. The entry "Externalism About the Mind" states the problem tersely: "On the face of it, externalism threatens the existence of such privileged knowledge."

The sharpest form comes out of the debate around McKinsey. The separate SEP entry "Externalism and Self-Knowledge" presents the argument this way: externalism and privileged self-knowledge "jointly entail an obvious falsehood, namely, that Oscar can know from the armchair contingent facts about the external world." If my thought "water is wet" is this thought only if there is H₂O in my environment, and if I know a priori that I am thinking this thought, then I would know a priori that there is H₂O in my environment. That cannot be. The same entry: "EXT and SK apparently lead to an absurd sort of apriori knowledge, and are thus incompatible." The incompatibility argument is attributed there to McKinsey (1991, 1994, 2002, 2007) as well as Brown (1995) and Boghossian (1997, 1998). The entry does not give McKinsey's first name.

Objection 4: Water Is Not Simply H₂O

A further objection is chemical. The thesis "water = H₂O" presupposes that water has a clean essence. The SEP on the natural-kinds debate reports several dissenting voices: Needham and van Brakel argue that "we cannot consider water to be just a compound composed of a collection of H₂O molecules"; Hendry points out that every liquid sample of water, "however pure," contains not only H₂O molecules but also H₃O⁺ and OH⁻ ions. Salmon argues that only a "trivial" essentialism can be inferred from the semantic arguments; Hugh Mellor regards the assumption of essences as "a gratuitous assumption" in the first place.

If that is right, it hits the water variant of Twin Earth directly: what exactly would be the fixed extension that "water" had on Earth in 1750? I am of the opinion that Putnam's aluminum/molybdenum and elm/beech variants are less exposed to this objection, because their difference does not hang on a question of essence but on the community's vocabulary and on expert deference. But that is no free pass: those variants depend all the more on deference – and that is precisely what, according to the SEP, Crane (1991) disputes, seeing no change of concept in Burge's case at all. Shifted, not solved.

Objection 5: The Intuitions Themselves

Finally, it is disputed that the intuitions are as clear as presented. The SEP names Boghossian's "Dry Earth" case (1997), Crane's objection (1991) that no concept changes at all in Burge's case, methodological criticism from Cummins (1991) and Chomsky (1995), and work in experimental philosophy (Mallon et al. 2009) testing whether the claimed intuitions are in fact shared.

Objection 6: The Question Itself

Brie Gertler (2012) goes one step further. Her objection concerns not the intuitions but the distinction on which the whole debate rests: according to the SEP there is "no understanding of the distinction between internal and external properties," and she therefore recommends "that philosophers drop the issue in favor of more well-defined questions" – replacing the question with better-defined ones.

The upshot of this part is not that Putnam has been refuted. It is that externalism is a strong, influential and contested position. On its significance the SEP says soberly that it is "potentially far-reaching though often unclear and disputed."

Part 7: What This Means for Machines That See Only Form

Here it gets interesting for anyone working with language models. Because Twin Earth, read correctly, is a machine for separating two things that are constantly confused in practice: mastery of a word and reference of a word.

Oscar1 and Oscar2 master "water" equally well. They refer to different things. What makes the difference is nothing that occurs inside them – in the water case it is the composition of the substance itself, in the elm/beech and gold cases it is cooperation with the speakers who know the criteria. Two different mechanisms, the same point: mastery does not fix reference by itself.

A related distinction – that between form and meaning – stands at the center of a much-discussed critical paper in recent NLP literature. Emily M. Bender and Alexander Koller presented the position paper Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data at ACL 2020 (Proceedings of the 58th Annual Meeting of the ACL, pp. 5185–5198; awarded Best Theme Paper). They formulate their core thesis in the abstract as follows:

"In this position paper, we argue that a system trained only on form has a priori no way to learn meaning."

And they explain why they consider this urgent – with their own epistemic hedge in the sentence: "However, we find that these successes sometimes lead to hype in which these models are being described as 'understanding' language or capturing 'meaning'."

Three things need to be kept clean at this point. First: Bender and Koller themselves call their work a position paper; they argue for a thesis, they do not prove a theorem. Second: they speak of systems trained only on form – a restriction that does not simply carry over to multimodal systems or systems embedded in tools and an environment. Third: they do not invoke Putnam. Anyone quoting the sentence as "AI cannot mean anything" is quoting it wrongly.

The connection between Putnam and this debate is not drawn by Putnam, and I should mark it as my own reading. I am of the opinion that Twin Earth is the most precise conceptual tool for taking apart the question "does the model understand something?": it shows that even with perfect inner sameness the question of reference is still open. From that follows a practically useful reformulation. The question is not whether a model "knows what water is." The question is whether there is, between its token water and the stuff in the world, a relation of the kind Putnam calls sameL – a connection established through causality and expert deference. I am of the opinion that in a model whose entire contact with the world consists of text, this relation is mediated by the people who wrote the text. Whether a chain mediated in that way counts as reference is itself contested – in the externalist literature inherited, deferential reference is precisely the normal case, not a failure. "Mediated" and "absent" are not the same thing.

Two concrete consequences, which I state as assessments and not as findings:

I am of the opinion that Putnam's division-of-labor hypothesis is a better metaphor for retrieval, tool and agent architectures than "knowledge in the model." A model that consults a source when uncertain is not an incomplete model – it is a speaker within a division of labor, exactly like Putnam with gold. Deference is not a workaround, it is the normal case of linguistic competence.

I am of the opinion that externalism offers an explanation of why pure form benchmarks can overestimate. Two systems can score identically on every textual test and still differ in what they refer to – that is exactly the Oscar1/Oscar2 constellation. What makes the difference measurable is not more text but contact: tools, sensors, checkable actions, feedback from the world.

And a warning in the same direction: Burge's case shows that the community's dictionary alone shifts the content. If a model is trained on corpora in which "arthritis," "critical infrastructure" or "consent" are used differently than in the community where it is deployed, its outputs are not merely stylistically off. I am of the opinion that Burge's case supplies the fitting analogy here – with one caveat I do not want to skip: Burge argues about the contents of the propositional attitudes of a subject who is prepared to defer. Whether a model has such attitudes is a question this article leaves open. So the analogy describes the phenomenon that practitioners call domain drift; it proves nothing about it.

The Central Takeaway

The central, practically usable insight of Twin Earth is not the slogan. It is this sentence, and it is deliberately phrased more weakly than the quotation on page 704:

Mastery of a concept and reference of a concept are two different things, and the first does not fix the second by itself.

Putnam's own conclusion says exactly that and no more: the extension is "not a function of the psychological state of the speaker by itself." That "by itself" is the whole content of the argument – and also its limit.

What follows for practice? Three test questions. They are not an inference from Putnam's text but my attempt to apply his tool; I am of the opinion that they bite on concepts that are repeatedly argued about inside organizations.

First: whose criterion decides? Putnam's hypothesis is descriptive – every linguistic community exhibits terms whose criteria are known only to a subset of speakers, and whose use by everyone else rests on a structured cooperation with that subset. Note: "at least some terms," not all. The practical value lies not in Putnam demanding an authority, but in his making visible what the division of labor looks like for concepts that work: if nobody on your team can say who decides, in case of doubt, whether something is "critical," "done," "personal data" or an "incident," then precisely the subset that Putnam's speakers defer to is missing. On this reading the remedy is a named owner and an escalation path, not a longer glossary.

Second: do we share the word or the world? Two teams can master the same word and talk about different things. What matters is which of the two mechanisms is at work: if both point at genuinely different objects – different systems, data fields, process steps – it is the Oscar constellation. If they use the same word with different community criteria, it is the Burge case. The remedy differs in each: in the first, point together at the same object; in the second, reconcile the lexicon.

Third: what does reference hang on? Before you ascribe "understanding" to a system – or to a document, or to yourself – ask through what chain the word is connected to the thing. Bender and Koller argue that a system trained only on form has "a priori no way to learn meaning"; that is a reasoned position, not a theorem, and it concerns only systems that see nothing but form. The test question remains useful nonetheless: is the chain to the thing your own or inherited – and if inherited, from whom?

A Question to Reflect On

Take a concept that comes up constantly in your work and that you assume everyone involved uses the same way – "done," "secure," "critical," "in scope," "in production." Then ask yourself the Twin Earth question: Is there a person or an authority in your organization whose judgment settles, in case of doubt, what falls under this concept – and would everyone who uses the word daily know whom to ask? If not: is it still the same concept when two teams use it, or are they merely using the same word? And if the latter – what would it cost to make the division of labor visible, instead of re-fighting the argument every time?

Cross-References in the Vault

Fact Check

After drafting, this article was put through a structured fact check. 130 checkable claims were extracted and held against the primary sources – 28 hard facts, 47 quotations, 25 source attributions and 30 synthesis claims made by the article itself. Result: 86 confirmed, 39 corrected or weakened (7 of them factually wrong), 6 not checkable. Of the 47 English quotations, 46 were verbatim correct; the weaknesses lay almost entirely in the draft's own generalizations.

The three most important corrections:

  1. "Methodological solipsism" was misassigned. The draft treated the term as Putnam's name for assumption (1). In fact Putnam introduces it in 1975 as a separate, deeper assumption about what counts as a psychological state in the first place.
  2. The hook asserted a location thesis that the article itself refutes. The SEP makes it clear: "External individuation does not entail external location." The hook was accordingly pulled back from "meanings are partly outside" to an individuation thesis.
  3. "The Central Takeaway" was stronger than Putnam's hypothesis. Its wording is descriptive ("Every linguistic community exemplifies … possesses at least some terms"); the draft had turned that into a "needs" and a "for every definition." Both have been brought back to what the sources support and marked as the author's own application.

Further corrections concerned: the first name Brie (not Barbara) Gertler and her substantive placement; the authorship of the SEP entry Externalism About the Mind (Rowlands, Lau, Deutsch – not Deutsch alone); the date of the chemical discovery of water's composition (late eighteenth century instead of "1800"); the provenance of the Kripke quotation (Kripke 1971, Identity and Necessity, not Naming and Necessity); the attribution of the "startling consequences" (according to Putnam to the theory, with Kripke as its first observer); restoring the epistemic "we find that" in the Bender/Koller quotation; and withdrawing the claim that the aluminum/molybdenum variant is "argumentatively more robust" – it merely shifts the burden from the question of essence to expert deference.

Limitations that belong to transparency: in the available scan of the 1973 paper, only the page numbers 700, 701, 702 and 711 are visibly printed; the references for pages 703 to 710 are inferred from running heads and the flow of the text (the central reference, p. 704, was verified twice). The first names Nathan Salmon and Michael McKinsey do not appear in the cited SEP entries and were cross-checked externally; the article therefore no longer gives them. The bibliography sections of the entries Natural Kinds and Externalism and Self-Knowledge were not fully retrievable.

Automated AI check against primary sources; no human expert review.

Sources and Further Reading

← All articles