Chapter 66
On Lying as the Origin of Meaning
The last chapter asked what a symbol system would have to be and answered with a functor that organizes the conditions attractively and then fails. This chapter takes up what the failure leaves standing.
Chapter 57 established a negative result and gave it two-sided evidence. Rongorongo and Linear A are corpora without populations: enough text to train a model to emit well-formed strings, and no community of speakers left. Sperm whale codas are the mirror case — no Rosetta stone, no cognates, and real progress, because the population is alive and its behavior can be watched. The conclusion was that the correspondence between a symbol system and the world is carried in the community of speakers rather than in the symbol system, and that decipherment is either transport into an anchor or direct observation of the using population, neither of which is distributional.
That chapter also insisted, in Observation 57.5, that the reference bootstrap problem is not open in the sense that its solubility is in doubt. It has been solved twice, independently. And it isolated three features common to both solutions: the solver was a population, the interface was not chosen, and there was an external criterion supplying the fitness signal.
What it did not supply was a mechanism. It said that the crossing has been made and did not say by what motion.
This chapter records a candidate. It is due to Michael Benedikt, it is that language originates in lying, and i think it is the most interesting thing anybody has said to me about the bootstrap problem. i want to be clear at the outset that what follows is speculative in a way that Chapter 65 was not. The propositions are targets. The reason to state them anyway is that two of them, if they went through, would upgrade the negative result of Chapter 57 from an observation to a theorem, and would say something sharp and unwelcome about what grounding requires.
66.1 Benedikt’s argument
Michael Benedikt was ACSA Distinguished Professor of Architecture at the University of Texas at Austin, held the Hal Box Chair in Urbanism, and wrote Deconstructing the Kimbell among much else [249]. He died in August 2025. The argument below is recalled from conversation and should be read as personal communication; a search has surfaced no published version of it.
The setting.
A band of early hominids on the plain, not far from the trees, dividing food — fruit, or the spoils of a kill.
The prior state.
Many animals make reflexive noises grounded in function: the sounds of breathing, howling in pain, the alarm call. These are triggered by conditions. They are not meaningful in the way that language is meaningful, and the whole argument turns on what the prior state lacks.
The event.
Eve notices movement in the grass. The conditions land on her nervous system in just such a way as to trigger the snake warning call. She calls; the band scatters for the trees. In that split second Eve determines that it was not a snake — it was the wind — and at the same moment notices that she now stands in an advantageous position with respect to the division of the food.
What she sees is the gap: the distance between the reflexively triggered sound and the behavior of the band. And she sees that the gap admits exploitation.
The follow-on.
If Eve does this regularly, others may notice the correlation between the snake call and the division of the food. The band learns Eve’s trick from Eve’s behavior. Language is a virus [250].
The reading.
It is the gap that gives the opening for meaning. A reflexive call and its triggering condition are welded together, and nothing there needs interpreting. Prise them apart and semantics becomes both possible and necessary. This is the sense in which lying is the knowledge of good and evil offered at the center of the garden — a remark i will leave as a remark, since the formal apparatus does not reach it.
66.2 The empirical situation is better than anecdotal
The parable has data behind it, and the data sharpen it in a direction the parable does not anticipate.
Tufted capuchins.
Wheeler reports that wild tufted capuchins give predator calls in non-predatory contexts where the caller stands to gain [258]. False alarms are given by subordinates more than by dominants, rise with the contestability of the food and with its scarcity, and are given from positions in which the caller would gain if the others reacted. Listeners leave the patch; the caller takes the food. This is Eve’s scenario, observed.
Counterdeception.
Listeners are more likely to ignore the same call type when it is produced during competitive feeding than in other contexts, and the difference is carried primarily by a drop in escape responses rather than in vigilance [259, 260]. The receivers have become context-sensitive about a signal that used to require no context.
Fork-tailed drongos.
Drongos steal food using mimicked alarm calls of other species, and vary which call they mimic [252, 253]. The reason given is that a deceptive signal loses power as its frequency rises relative to honest signals, so varying the signal is what sustains the deception.
Earlier work.
Munn’s kleptoparasitic shrike-tanagers and antshrikes [256], whose status as alarm rather than aggressive signals has since been questioned.
Under frequency-dependent discounting, deception does not merely exploit an existing repertoire. It exerts pressure to expand the repertoire, because a fresh signal starts at the honest response rate.
Observation 66.1 is a mechanism for vocabulary growth and not only for the birth of meaning, and it is the part of the picture the parable alone would not have produced.
66.3 Three stages, and a re-typing
The translation into the framework is short, and its shortness is the point: nothing is added to the calculus.
Stage 0: the reflex.
The call is a rewrite rule, not a symbol. It is a sensor-driven emission, and the receiver’s response is another rewrite. The correspondence to the world is a fact about the wiring, and neither party needs a hypothesis. In the terms of Table 1 of Chapter 21 both sides are non-modal: this is perception, reading structure the world has already computed into legible form.
Stage 1: the decoupling.
Eve’s innovation is to make the call channel writable by her policy independently of the sensor. In the channel classification of Chapter 21 — channels enabling internal reductions, channels enabling interaction with the environment, and channels doing both — this is a re-typing. An emission driven only from the sensor side becomes driven from the policy side as well. The gap is a change in which redexes can fire on a name, and that is all it is.
Stage 2: semantics as the receiver’s forced expense.
Once emission is decoupled, call heard no longer entails snake present. The receiver’s predicate cannot remain non-modal. It must be indexed by context, and the context is a hypothesis about the sender rather than about the world. This is machinery the Turn already built: context-labeled modalities, and an environment whose most interesting contents are other learners.
Meaning is not a property of the sign. Meaning is the theory that receivers are forced to build once senders are capable of lying. Semantics is the receiver’s bill.
66.4 Goodman, read backwards
Chapter 65 adopted Goodman’s conditions provisionally and spent them: on the two bisimulations, on what a budget buys, and on the construction that fails. The adoption was a bet that the conditions were the right ones rather than merely the first ones available. With Stage 0 in hand there is something to say about that bet, and it is not what i expected to be saying.
Goodman’s marks are, in his own gloss, physical inscriptions or utterances. A triggered alarm call is a mark. The apparatus therefore applies to Stage 0 directly, with no reinterpretation and no analogy, and we can simply run the five conditions on a reflexive population and see what happens.
| Condition | Stage 0 (reflex) | After the lie | Repaired by |
|---|---|---|---|
| Syntactic disjointness (65.1) | holds | broken by mimicry | — |
| Syntactic finite diff. (65.2) | holds if discrete | unchanged | Conjecture 65.1 |
| Unambiguity (65.3) | holds maximally | broken by the lie | — |
| Semantic disjointness (65.4) | holds | broken by mimicry | — |
| Semantic finite diff. (65.5) | holds if discrete | unchanged | Conjecture 65.1 |
66.4.1 A perfect notation that means nothing
Take the conditions in turn on a Stage 0 population with a discrete repertoire. No mark belongs to two characters, so Condition 65.1 holds. The call types have no structurally intermediate forms, so Condition 65.2 holds. Distinct call types fire on distinct triggering conditions, so Condition 65.4 holds. And Condition 65.3 — that each character have a single compliance class, not different compliants at different times or in different contexts — holds maximally, because the call and its condition are welded. That weld is what Benedikt’s argument says the prior state consists of.
A discrete reflexive signalling system satisfies all five of Conditions 65.1–65.5. It is, by Chapter 65’s adopted criterion, a notational system. It is also the system in which, on the argument of this chapter, nothing means anything.
Observation 66.3 sharpens Remark 65.1 rather than contradicting it. That remark said Goodman’s conditions cannot select a compliance relation, being invariant under relabeling. The reflex case says something less comfortable: the conditions are best satisfied precisely where no compliance relation was selected at all, because the world wired it. Notational adequacy is not merely insufficient for reference. It is maximized at the point where reference is absent, and degrades as reference arrives.
And the condition that degrades is nameable. Stage 1 introduces a second production path for a mark that had one, so that call heard has one compliant when a snake is present and another when a subordinate stands over contested food. That is Condition 65.3, failing, in the condition’s own words. The gap is not analogous to a failure of unambiguity and is not a consequence of one. It is the same fact under two descriptions.
66.4.2 Compliance requires possible non-compliance
There is a stronger reading available, and it is the one i would defend.
Goodman’s compliance relation presupposes that compliance can fail. The whole wrong-note paradox of §65.1.3 exists because a performance can fail to be a performance of the work; if it could not, there would be no paradox and no need for the budget repair. But a trigger cannot fail. If the condition holds and the wiring is intact the call fires, and if the call fired the condition held. There is no room for a mis-transcription.
So at Stage 0 there is not a compliance relation that happens to be perfectly satisfied. There is no compliance relation — only causation, described twice. Character and compliant are not two things standing in a relation; they are one thing under two names, which is why the relation between them cannot fail and why Condition 65.3 is satisfied for free.
This shows up structurally as an arrow count, and the arrow count is the cleanest version of the point.
Goodman’s adequacy, by §65.1.2, is a roundtrip: character to compliant and back, compliant to character and back. At Stage 0 there is one arrow, world to mark, and no return, because the mark is caused by the compliant rather than produced from a character. Stage 1 supplies a second producer of marks, so the mark ceases to determine its producer. Stage 2 requires the receiver to invert, and by Conjecture 66.7 the inversion is of a mixture rather than of a function.
Read that way, Goodman’s framework requires exactly the structure that deception introduces. Before the lie there is a transduction and no notation; the transcription direction does not exist to be checked. This is worth holding alongside §65.7 of the last chapter, which located the transducer inside the calculus and reassigned the network to proposal. The transducer is not where notation lives, and Stage 0 is a population that has a transducer and nothing else.
66.4.3 Deception attacks what the budget cannot repair
The right-hand columns of Table 66.1 are worth reading together, because the division is exact.
Conjecture 65.1 makes the two finite-differentiation conditions budget-relative, and its reading was that a poorer agent has a coarser notation. It says nothing about unambiguity or about either disjointness condition, and this is not an oversight: those three are not statements about how finely an observer can separate things, so no amount of budget bears on them. A rich agent and a poor agent face the same ambiguous sign.
Deception breaks exactly those three. The lie breaks unambiguity. Mimicry breaks the two disjointness conditions — and does so in a way worth stating plainly, because it is not a semantic phenomenon at that point. A fork-tailed drongo emitting a pied babbler’s alarm call [252, 253] has put one mark into two characters. Condition 65.1 is syntactic, and deception at the community level is a syntactic attack on it.
Between the budget move and the deception mechanism, then, all five conditions are accounted for, and neither account touches the other’s. That is not an argument that either is correct. It is an argument that they are two halves of one thing, and that Goodman’s conditions decompose along a seam nobody had reason to look for until there was a liar in the picture.
66.4.4 What the ethologists have been measuring
One check is available immediately, and it is favorable in an unusual way: the relevant measurements have already been made by people who had never heard of Goodman.
The standing criteria for functional reference in animal signalling are two [261]. Production specificity: the signal is given to one class of stimulus. And the perception criterion: the signal elicits its response in the absence of supporting contextual cues — the property that made the vervet alarm calls the textbook case [262]. The second is Condition 65.3, tested from the receiver’s side and stated in behavioral terms. And the live dispute in that field is whether receivers integrate context after all — which is precisely what the counterdeception results of [259, 260] report they do, and precisely what Stage 2 predicts they must.
A second hit lands on the syntactic side. Quantitative reanalysis of vervet alarm calls finds the leopard, eagle and snake calls well separated, with overlap appearing once calls given in aggressive contexts are included [263]. Condition 65.1 fails exactly where the socially strategic use enters, which is where this chapter says it should.
Neither observation was made to test anything here, which is the reason to report them.
66.4.5 The objection, which i cannot dispose of
A deflationary reading is available and should be stated at full strength.
One may say: at Stage 0 the compliance class of a character simply is its triggering condition class. That is a perfectly good compliance relation. The system is therefore notational, Observation 66.3 is correct but unsurprising, and all it establishes is that notationality is not sufficient for meaning — which Remark 65.1 already said, by a shorter route. On this reading nothing above is new.
The reply is Observation 66.4: a theory of notation whose central failure mode — the wrong note — is unstatable in its best-behaved instance is telling us something about what it measures. But that reply depends on the claim that compliance requires possible non-compliance, and that claim is at present an intuition with a paragraph attached rather than an argument. It is the first thing i would want written down properly, and until it is, the deflationary reading is live and the stronger one is a proposal.
66.4.6 A prediction the pairing generates
Combining the two halves of Table 66.1 yields a claim that neither makes alone.
Under the frequency-dependent discounting of Observation 66.1, a fresh signal is worth manufacturing only if it starts at the honest response rate, and it can do that only if receivers do not confuse it with the discounted ones. The pressure of Conjecture 66.5 is therefore not merely toward more characters but toward characters satisfying Conditions 65.1 and 65.2. Deception pressure should drive a repertoire from the analogue end of Goodman’s distinction toward the digital end.
Conjecture 66.1 is comparative and testable. Signalling systems are documented as graded, as discrete, and as mixed, and the variation is measurable. The prediction is that discreteness should covary with deception pressure, and the drongo is a suggestive extreme: it does not manufacture new characters at all but imports discrete ones wholesale from other species, which is the cheapest possible way to buy distinguishability.
If it holds, the digital / analogue distinction acquires a second economic reading to set beside the one Conjecture 65.1 gives it. There, a notation is as fine as its observer can afford. Here, a notation is as discrete as its liars have made it necessary to be.
66.5 Statements to attempt
The retrospect of §66.4 named the condition that fails and left the failure informal. What follows are the targets that would make it precise, together with the ones that would give the mechanism empirical content. None is established.
In a population where every emission is sensor-driven, the receiver’s optimal policy is a non-modal predicate on the signal: the modal rungs buy nothing.
The corollary would be that an alarm call in a wholly honest population is not a symbol — and the framework could say so with a purchase criterion rather than with a definition, which is the difference between a result and a stipulation.
With a nonzero deception rate the non-modal predicate is strictly dominated, and the receiver’s best available response requires context-indexing. The transition has a threshold in the deception rate.
The threshold should be computable in the currency of the ladder returns of Chapter 22, and Conjectures 66.2 and 66.3 together are the pair such a computation could actually decide.
The liar’s yield is proportional to the response rate, which declines in the population deception rate; hence an interior equilibrium.
The empirical predictions are already made and already met: subordinates lie more, lying rises with contestability, listeners discount contextually [258, 259].
At equilibrium there is pressure to diversify the signal. Vocabulary size grows as a function of the deception pressure and of the cost of manufacturing a new distinguishable name.
Freshness by quotation says names are cheap to manufacture, which predicts fast growth — plausibly too fast. What limits it is a real question and a good one.
The band learns Eve’s trick by watching Eve, not by descending from her: a quoted process copied between contemporaries. This is the reproduction material of Chapter 21 with the vertical arrow replaced by a horizontal one.
Burroughs’ virus is not a metaphor under Conjecture 66.6. It is an \(R_{0}\), and the thing it spreads degrades the resource it spreads on.
If emission is policy-driven, the joint distribution over signals and world-states is a mixture over sender policies whose weights are facts about the population. A corpus gives the marginal, and the marginal does not identify the mixture.
Conjecture 66.7 is the statement with reach outside the framework, because it says that no corpus of any size recovers the correspondence, and says it as a claim about strategic signaling rather than as a complaint about data scarcity. If it goes through, the central observation of Chapter 57 is upgraded: the correspondence lives in the community because the senders are strategic. Linear A resists because its semantics was never in the distribution. Whale song yields because the senders remain available to be modeled.
A transformer over a corpus can only read. Under Conjecture 66.7 the readable object is the wrong one, and the identifying information is available only to an agent that can emit into the community and observe the response — that is, to one that can perform the intervention separating the mixture components.
This sharpens Section 65.7 of the last chapter rather than contradicting it. There the network was reassigned from transducer to proposal mechanism, on the grounds that the transducer is already exact and a network cannot improve on it. Conjecture 66.8 says something stronger about the other half: not merely that a corpus is the wrong place to look for the transducer, but that a corpus is the wrong kind of object for recovering correspondence at all, for a reason that has nothing to do with its size. If it holds, grounding requires participation rather than perception, which is a sharp and unwelcome claim about what embodiment has to mean. Its interest is that it is stated in the framework’s own pricing vocabulary rather than imported.
66.6 Caveats, and what the argument does not license
Deception is not the only bootstrap.
Lewis conventions and Skyrms’ signaling games produce meaning by reinforcement under aligned interests, with no liar anywhere [254, 257]. The defensible claim is narrower than “lying is necessary for signaling”. It is that lying is what makes meaning require modeling, forcing the receiver up from correlation to theory. The two routes should yield different kinds of symbol system, and characterizing the difference is a good problem.
Deception is parasitic on an honest baseline.
The reflexive layer must exist first, or there is nothing to exploit. This constrains any engineered bootstrap: build a functionally grounded reflex layer before there is anything for a policy to detach from.
The ethological alignment should be checked rather than reinvented.
Dawkins and Krebs argue that signals are manipulation rather than information transfer [251], which is Benedikt’s thesis arriving from ethology; and the equilibrium of Conjecture 66.4 has presumably been worked out in the costly-signaling literature [255] in a different currency. Import it and pay it in tokens.
Intentionality is not claimed and should not be.
The capuchin literature is careful about whether false alarms are deliberate, to the point of testing whether adrenocortical stress explains the calling. The framework does not need deliberateness. It needs only that emission be policy-driven rather than sensor-driven, which is a structural condition on channels and is exactly what the conjectures above are stated in terms of.
66.7 Open questions
Does the deception rate play the role that the environment plays in the inversion of Chapter 23 — so that whether meaning is affordable is a property of the environment rather than of the learner, exactly as whether truth is affordable turned out to be?
Can Conjecture 66.7 be given as a genuine identifiability theorem, with the mixture stated over a space of sender policies and the failure of identifiability exhibited rather than argued?
Of the three common features of Observation 57.5, the third — an external criterion — is here instantiated as the food share, and the first — that the solver was a population — is here explained rather than observed, since a lie requires someone to lie to. Does the second, that the interface was not chosen, receive a similar treatment?
Deception should be impossible in the limit where agents reason over one another’s source rather than over one another’s claims. In the three grades of access of Table 1 of Chapter 21 that limit is the reflection grade, and deception lives at grades one and two. Does it return under bounded verification, when the cost of checking a property exceeds the stake — making obfuscation an economic attack rather than an informational one? And if two agents are extensionally indistinguishable on the fragment each can afford to check, deception appears to live entirely in the complement, which suggests a framework-native definition: deception is engineering the opponent into the undecided region and harvesting their default.
66.8 What is claimed
Claimed: that the reflex-to-symbol transition can be stated in the framework as a re-typing of a channel, with nothing added to the calculus; that on that reading meaning is the receiver’s forced expense rather than a property of the sign (Observation 66.2); and that frequency-dependent discounting gives vocabulary growth a mechanism (Observation 66.1).
Not claimed: any of the seven conjectures, and in particular not Conjecture 66.7, which is the one i would most like to be true and therefore the one to be most careful about. It is stated as a target because the shape of the required proof is visible — a mixture over sender policies, and a demonstration that the marginal does not identify the weights — and because both sides of it are things somebody could actually work on.
There is a reason this chapter sits where it does. Chapter 65 reached a construction that fails, and named the failure precisely: the calculus supplies no proposal mechanism, no way to get from a stream of world to a set of builtins. The temptation at that point is to conclude that what is missing is more data or a better architecture. The suggestion here is that what is missing is a liar — that the information which fixes reference is not in any record of what was said, because what was said was said by somebody with an interest, and the only way to learn what an interested speaker means is to speak back and find out what it costs them.