Where the Model Lies

Five lessons in, you have a beautiful machine. Costs that only the honest can pay; peacock tails that are honest because they’re absurd; separating equilibria that snap shut on a single inequality. It is elegant, it is powerful, and — like every powerful model — it is a liar. Not a total liar. A partial one, which is worse, because partial liars are the ones you trust too far.

This is the lesson where we turn the theory on itself. Where does signalling fail as an explanation? When does invoking it make you dumber, not sharper? A model you can’t criticize isn’t a model, it’s a religion — and you paid for the expert tier, not the sermon. So let’s be even-handed, rigorous, and occasionally unkind to the very framework we just spent a course admiring.

Dishonest mimicry when the cost gap narrows

Intuitive analogy. A castle’s moat keeps out attackers only while it’s wide and deep. Fill it in a little — or hand out cheap boats — and the same moat that once meant “impregnable” now means nothing, even though the castle looks identical. The defense was never the shape of the moat. It was the cost of crossing it, and costs erode.

Precise definition. A separating equilibrium is honest only while faking stays expensive for the bad type. The instant the low type’s cost of producing the signal falls — call it c_L dropping toward c_H — the Spence window B/c_L < s ≤ B/c_H narrows, then closes. Liars flood in, both types pool on the same message, and the signal’s information content decays toward zero. Formally this is an eroding (unstable) separating equilibrium: it was never a permanent property of the signal, only a temporary property of the cost gap sustaining it. This failure mode is dishonest mimicry — the bad type learning to counterfeit the honest type’s display.

Fully worked example. Recall our numbers: reward of being believed B = \$60, and per-unit costs c_H = 12, c_L = 30, receiver demanding s = 3. The low type nets 60 − 30×3 = −30, so it refuses — clean separation. Now a “signal factory” opens and drops the low type’s per-unit cost to c_L = 15. Re-run the low type: 60 − 15×3 = +15. Positive. The liar now happily fakes it. Both types send s = 3, the receiver’s posterior after seeing the signal collapses back to the prior, and the diploma / logo / warranty conveys nothing. Nobody changed the signal. Somebody changed the price of forging it, which is the only thing that ever mattered.

The Red Queen line. This never settles, because signallers and fakers co-evolve: honest senders escalate the display, fakers catch up, honest senders escalate again — an evolutionary arms race where, like the Red Queen, both run flat out just to stay in the same relative place. Batesian mimics (the harmless hoverfly wearing the wasp’s warning stripes), counterfeit luxury goods, fake blue-check accounts, and gamed KPIs are all the same story: the signal was honest until forgery got cheap, and forgery always eventually gets cheaper.

Common pitfall. Treating a signal’s honesty as a fixed attribute — “a degree is a costly signal, therefore a degree is honest, full stop.” Honesty is not a property of the signal; it’s a property of the current cost gap, and cost gaps decay. Any signal worth faking will eventually be faked, and the more valuable the separation, the harder someone works to counterfeit it. Yesterday’s costly signal is tomorrow’s cheap talk on a long enough timeline.

When to use it. Before you trust a signal, ask not “is this costly?” but “is this still differentially costly, and for how much longer?” When you see a metric being gamed, a credential inflating, or a status symbol going mass-market, you’re watching the cost gap close in real time. Reach for this lens whenever a once-reliable signal seems to be losing its punch — you’re not imagining it; the moat is filling in.

The deadweight waste of pure signalling arms races

Intuitive analogy. Everyone at the concert stands up to see better. For one person, standing works — they see over the crowd. When everyone stands, nobody sees any better than when everyone sat, and now everyone’s legs hurt. The standing was individually rational and collectively pointless: pure burned effort that reshuffled nothing.

Precise definition. Even when a signal does separate honestly, the system can still be collectively wasteful. In a positional competition, what a signal buys is relative rank, not absolute value — and rank is zero-sum. If everyone escalates their signal to hold the same place in the queue, the extra cost is deadweight: a prisoner’s-dilemma-like burn where each player’s best response is to spend more, but the equilibrium leaves everyone poorer at the same relative standing. Crucially, distinguish two cases:

• Productive signalling — the burned cost also improves matching: the right workers reach the right jobs, good borrowers get loans, healthy mates pair off. The separation itself has social value even if the cost is high.
• Pure zero-sum handicap race — the signal only sorts rank and builds nothing; every dollar of escalation is matched by rivals, so the sorting stays identical and the collective spend is pure friction.

Most real systems are a blend, and the honest analytic question is: what fraction of the burn is buying better matching versus just funding an arms race?

Fully worked example — the education arms race. Suppose a diploma perfectly reveals ability but teaches zero usable skill (Spence’s own unsettling case). Ten workers, ranked 1–10 in ability, compete for jobs ranked by pay.

• Round 1: everyone gets a bachelor’s. It separates ability (able find it less painful), so matching is correct. Cost: 4 years each. Society learns who’s who.
• Round 2: to stand out, the top few add a master’s. Now employers expect a master’s for the good jobs, so everyone chasing those jobs adds one too. The ranking is unchanged — the same person is still #1 — but now everyone burned 6 years instead of 4.
• Round 3: repeat with PhDs, certifications, unpaid internships.

The added years past round 1 revealed nothing new (the ranking never moved) and built nothing (the schooling taught no skill, by assumption). It’s the concert crowd standing up: 2 extra years × 10 workers = 20 person-years incinerated to reproduce the exact sorting round 1 already achieved. That’s the deadweight signature — real cost, zero incremental information, zero incremental skill.

Common pitfall. Confusing “this signal is expensive” with “this signal is wasteful,” or its opposite. Expense is necessary for separation (drop it and the signal dies). Waste is about whether the expense buys anything beyond the separation. A costly signal can be expensive-and-useful (first degree sorts workers) or expensive-and-wasteful (the tenth credential in an arms race). Same word, opposite verdict.

When to use it. Reach for the arms-race lens whenever you see escalation without improvement: credential inflation, ad-spend wars between near-identical brands, luxury-goods one-upmanship, résumé-padding. Ask “if everyone dialed this back by 20%, would the sorting actually change?” If no, you’re funding a positional race, and the socially optimal move is a coordinated cap that no individual can afford to make alone. That’s why these systems are so sticky.

Revelation vs causation — the deep confound

Intuitive analogy. A thermometer reveals the fever; it doesn’t cause it. But now imagine a magic thermometer that also, a little, warms you up while it measures. Is the reading telling you about a fever that was already there, or about heat the thermometer itself added? If you can’t take the temperature without the thermometer nudging it, separating “what was there” from “what the measuring did” becomes genuinely, maddeningly hard.

Precise definition. This is the theory’s single hardest, most honest limitation. When a costly signal correlates with quality, there are two rival explanations, and they are brutal to disentangle:

• Signalling / revelation — the signal reveals a pre-existing quality the good type already had. The degree sorts; it doesn’t build. (Spence.)
• Human capital / causation — the signal causes the quality by building it. The degree teaches real skills, so graduates are genuinely more productive because of the schooling. (Becker.)

Both predict the same surface fact — “degree-holders earn more” — so the correlation alone cannot tell them apart. Reality is almost always a mixture: some revelation, some causation, in proportions that vary by field, and economists have argued about the split for fifty years without settling it. That’s not a failure of effort; it’s a genuinely hard identification problem.

Fully worked example. Graduates earn a \$20,000/year premium over non-graduates. How much is signalling vs human capital?

• Pure-signalling reading: the degree taught nothing usable; it just certified people who were already more able. Estimated causal effect of the education itself on skill: near $0. The premium is the market paying for the revelation.
• Pure-human-capital reading: the degree built $20,000/year of genuine skill; a randomly chosen person made to complete it would gain the full premium. Causal effect: the whole $20,000.
• How you’d actually tell them apart: you need a shock that changes the credential without changing the underlying ability — natural experiments like the “sheepskin effect” (a big pay jump for the final year that grants the degree versus an equal year that doesn’t, holding schooling roughly constant, hints at signalling) or randomized/quasi-random access to schooling (which tends to show real skill gains, hinting at human capital). The honest empirical verdict across studies: both are real, the mix depends heavily on the field (a welding certificate builds more than it signals; some elite generalist degrees signal more than they build), and confident all-or-nothing claims are almost always overreach.

Common pitfall. Using a signalling explanation to make a causal claim — e.g., “since a degree is just a signal, closing the university would lose nothing.” That only follows in the pure-revelation extreme, which the evidence rarely supports. The model tells you a signal can be pure revelation; it does not license you to assume it is in any particular case.

When to use it. Whenever someone attributes a person’s competence to a credential — or dismisses the credential as “just a signal” — pump the brakes and ask: is this revealing quality that was already there, or building it? Then ask what evidence could distinguish the two. If you can’t name that evidence, hold your conclusion loosely. This is the frame that keeps you from being either a credentialist or a contrarian bore.

Over-reading noisy, one-shot signals

Intuitive analogy. A single coin coming up heads doesn’t prove the coin is biased. A costly signal fired once, in a noisy world, is a single coin flip — informative in expectation, but wildly over-read when you treat one observation as proof.

Precise definition. A signal is probabilistic evidence, not proof. It shifts the receiver’s belief toward the good type; it does not certify it. In noisy or one-shot settings, a costly display can be luck (the good outcome happened for reasons unrelated to type), a fluke (rare variance), or a rare successful fake (the mimic that slipped through this once). The likelihood ratio can be strong and still leave a real chance you’re wrong — especially when the good type is rare to begin with.

Fully worked example — base rates bite. A costly signal is fired by 90% of true experts and only 5% of frauds — a strong, honest signal. But suppose only 2% of the applicant pool are true experts (the base rate is low). You see one applicant fire the signal. What’s the chance they’re a real expert?

• Experts firing it: 0.02 × 0.90 = 0.018
• Frauds firing it: 0.98 × 0.05 = 0.049
• P(expert | signal) = 0.018 / (0.018 + 0.049) ≈ 0.27

Just 27%. A strong, genuinely costly signal, and one observation still leaves you more likely wrong than right — purely because experts were rare to begin with. Treat that single display as proof and you’re confidently mistaken almost three times in four.

Common pitfall. Reading a costly signal as a guarantee and forgetting the base rate — the classic base-rate neglect / miscalibration error (the exact failure the calibration mental model warns about). One tightrope walk, one killer quarter, one viral post: strong evidence, not certainty. The rarer the good type, the more a single impressive signal should update you less than it feels like it should.

When to use it. In any one-shot, high-noise, high-stakes read — a single interview, one investment track record, a lone heroic anecdote — deliberately weaken your conclusion. Ask for a second independent signal, or downweight for the base rate. Repeated, independent costly signals compound; a lone one is a coin flip that landed your way once.

When cheap talk actually works

Intuitive analogy. “The fire exit is that way.” Free to say, instantly believed, and correct — because the person shouting it has zero incentive to send you the wrong way. Not every message needs a moat. Some messages are trustworthy precisely because nobody benefits from lying.

Precise definition. Signalling theory is a theory of adversarial, conflicting-interest communication. Its “words are worthless” conclusion holds only when the sender’s interests clash with the receiver’s. When incentives are sufficiently aligned — coordination games, giving directions, a teammate calling the play, a doctor telling you which pharmacy line is shorter — cheap talk transmits information perfectly well, because no one gains by lying, so the honest message is also the self-interested one. This is the Crawford–Sobel result in one breath: cheap talk is informative when preferences are close and degrades as interests diverge.

Fully worked example. Two people trying to meet up both want to end up in the same place; there’s no conflict, only a coordination gap. One texts “let’s meet at the north entrance.” It’s costless. It’s also instantly credible and acted upon, because lying serves neither party — they both just want to converge. Compare a used-car seller saying “it’s a gem”: same zero cost, but now interests clash (they profit from your mistake), so the identical form of message (cheap talk) flips from reliable to worthless. The message didn’t change; the alignment of interests did.

Common pitfall — over-applying the course. Walking out of a signalling course convinced that “all talk is worthless, demand a costly signal for everything.” That’s a category error: you’d waste enormous effort demanding proof in cooperative settings where a free word was already trustworthy. The lesson isn’t “distrust words.” It’s “distrust words when interests conflict.” In aligned settings, cheap talk is a feature, not a bug.

When to use it. Before demanding a costly signal, check the incentive alignment. Do our interests point the same way (coordination, teammates, “the exit is there”)? Then take the free word. Do they conflict (a sale, a plea, a defense)? Then, and only then, look for the cost a faker couldn’t pay. Matching the tool to the alignment is the whole skill.

Reflexive, self-defeating signals

Intuitive analogy. The moment everyone learns that the “quiet luxury” crowd signals wealth by not showing logos, someone starts selling logo-less-but-obvious status goods, and the signal that meant “I’m so rich I don’t need to prove it” starts meaning “I read the same trend piece you did.” A signal everyone can read is a signal everyone can game.

Precise definition (light touch). A reflexive or self-defeating signal is one that loses or inverts its meaning once it becomes widely understood. Because signals live in a strategic system, the act of decoding a signal changes how people send it: the readable signal gets exploited, imitated, or deliberately subverted, so its information content decays through the very act of being understood. It’s the arms race from the first section, turned inward — the signal defeats itself as knowledge of it spreads. (We met the elegant cousin of this in the previous lesson: countersignalling, where the very top skip the signal because they can. Here it’s the darker twin — the signal eaten alive by its own legibility.)

Worked micro-example. A startup signals confidence by leasing a lavish office (“we’re doing so well we can afford this”). Investors learn to read it. Now failing startups lease lavish offices to look confident and lure funding — so a lavish office starts signalling desperation as often as success. Once the code is public, the signal flips. The smart move becomes reading the second-order game: who’s signalling, who knows you know, and who’s exploiting your knowledge that they know.

When to use it. Whenever a signal has become common knowledge — everyone knows what it “means” — assume it’s being gamed and read one level up. Ask not “what does this signal say?” but “what does someone who knows I can read this signal gain by sending it?” The most legible signals are the most manipulable.

Sorting fair use from over-reach

You now have a full failure catalog. Time to use it as a filter. For each statement, ask: is this a fair application of signalling theory, or does it over-reach — treating a probabilistic sort as proof, mistaking revelation for causation, ignoring an eroding cost gap, or over-applying “talk is worthless”?

Revelation, causation, and the arms race — one more pass

How to use the model well

The failure catalog isn’t a reason to abandon signalling theory — it’s the theory finished. A model without its failure modes is a half-model. Here’s the disciplined three-question protocol for using it without getting used by it.

Question 1 — the classic cost test. Would this signal cost the liar as much as the honest party? If a bad type could send it just as cheaply, it’s cheap talk wearing a costume; trust it only if interests are aligned. This is the engine from lesson 1, and it’s still the first thing you check.

Question 2 — the value test. Is the cost buying real information, or just burning value in a race? A signal can separate perfectly and still be a positional arms race that reproduces the same ranking at enormous collective cost. Run two ledgers: private value (does it separate?) and social value (does the burn buy better matching, or just relative rank?).

Question 3 — the confound test. Am I mistaking revelation for causation? Does the signal reveal a quality that was already there, or build it? Usually it’s a mix, the data rarely settles it cleanly, and confident all-or-nothing claims in either direction are the mark of someone who skipped this lesson.

And two guardrails riding along: signals are probabilistic, not proof — down-weight one-shot noisy displays for the base rate — and the “words are worthless” verdict applies to adversarial settings only; in aligned ones, take the free word and save your effort.

Putting it together

You came in trusting the machine; you leave able to audit it. Costly signals are real, and the cost gap is genuine, and the peacock’s tail is honest — and every one of those truths has a boundary where invoking signalling makes you dumber than staying quiet. The expert isn’t the person who sees signals everywhere. It’s the person who knows exactly where the model lies, and reaches for it only inside its own borders.

Next up: the Final Exam — one irreversible run across everything, from cheap talk and the Spence window through handicaps, markets, social signals, equilibria, and now the failure modes. No Back button, no retry: submitting locks each answer for good, exactly like a signal you can’t take back once you’ve sent it. Read every question the way you now read the world’s costs — carefully, and one level up.