The Placebo Term

The last essay proved a conservation law and then confessed a crack in it. Because surprise is a log, it telescopes: however many times your expectation is revised on the way to an event, the total surprise is fixed by the endpoints — where E began, what A turned out to be. But the proof leans on an assumption so natural you don’t notice yourself making it: that the Actual is fixed. That A cannot hear E. Test screenings re-cut endings, I admitted; roadmaps bend toward the demo; that case deserves its own essay. This is that essay.

The assumption has a name in physics: the world is not listening. The sun does not read your forecast before deciding whether to rise. The dice do not check the spread. Gravity has never once adjusted itself to flatter a prediction. For that world — the deaf world — the Reality Equation runs clean: you carry an Expectation, the Actual arrives from somewhere your beliefs cannot reach, and the gap pays out as surprise, the only currency attention accepts.

In 1948 the sociologist Robert Merton described the other world. He invented a bank — the Last National Bank, sound, solvent, run by one Cartwright Millingville — and then he invented a rumor. Depositors hear the bank is failing. The bank is not failing; the belief is false on the day it is born. But believing it, they line up to withdraw, and no bank on earth holds cash against all of its deposits at once. By closing time the false belief has become a true one, retroactively self-certifying. Merton coined a phrase for this that has outlived nearly everything else in his field: the self-fulfilling prophecy. The prophecy did not predict the failure. It manufactured the failure out of a solvent bank and an expectation.

Notice what happened to the equation. E did not drift toward A, the way a trailer walks you toward a film. A moved toward E. The denominator wrote the numerator. Every conservation argument I have made in this series silently assumed that could not happen — and whole floors of human life are built on the fact that it can.

Start with the body, because the body settles the question of whether this is a metaphor.

In 1978, Levine, Gordon and Fields gave dental-surgery patients a placebo and watched some of them get real relief — the standard, faintly embarrassing result. Then they administered naloxone, the drug that blocks opioid receptors, the same drug that reverses a heroin overdose. The placebo relief vanished. Expectation, it turns out, is not a mood laid over the pain; it is a prescription the brain fills from its own pharmacy, in endogenous opioids, through the same receptors morphine uses. Block the receptors and the belief stops working — because the belief was working somewhere physical all along. The Kaptchuk trial in the scene above pushed the result one step further: the pharmacy fills the prescription even when the label admits the pill is blank. The expectation does not need to be a deception. It only needs to be held.

Classrooms run a weaker version of the same coupling. In 1968, Rosenthal and Jacobson told teachers that a randomly chosen fifth of their students were about to bloom intellectually. The students were ordinary; the list was a coin flip. The listed first- and second-graders nonetheless gained measurably more IQ than their classmates. Honesty about the record: the effect has been fought over for fifty years, the original test instrument was attacked within months, and the meta-analyses that survived — Raudenbush’s synthesis of eighteen experiments is the sober one — find a real effect that is small and that collapses to nearly nothing once a teacher has known the students for more than a couple of weeks. Which is itself the finding, if you read it in this series’ vocabulary: expectation edits the Actual only where the Actual is still soft. A teacher who has already measured you is carrying an E built from data, and that E follows the child. A teacher handed a false E before meeting you will spend a season making it true.

And then there is the case where the coupling is strong enough that nations legislate against it. The Diamond–Dybvig model — the 1983 paper that won its authors a Nobel — formalized Merton’s parable: a bank is a machine for turning short-term deposits into long-term loans, which makes it solvent and run-able at the same time, two equilibria, and the expectation of a run selects between them. The American answer was not more cash. Deposit insurance went into effect on the first day of 1934; roughly 4,000 banks had suspended in 1933, and in 1934 nine insured banks did. The FDIC did not add a dollar to a single teller’s drawer that morning. It edited E — your deposit is safe whatever your neighbors believe — and because in banking A is downstream of E, editing the expectation edited the outcome. It remains the cheapest large-scale repair in financial history: a firewall built entirely inside the customers’ heads.

Fold the listening into the equation and see what it costs. Suppose the world you are facing writes its Actual partly from its own facts and partly from your forecast — a coupling k, where k is the fraction of the outcome that is taken down from your dictation. In the simplest toy version:

At k = 0 you get the equation this series has run on for months: the world delivers A₀, and surprise pays in full. For any k between zero and one, every unit of surprise is discounted by (1−k) — because your expectation was not just a forecast of the outcome, it was an ingredient in it, and you cannot be fully surprised by a dish you helped cook. And at k = 1 the equation dies. A equals E identically. S = ln(1) = 0, always, for every event, forever.

I want to be precise about why that limit is sinister rather than restful, because this series has already described a good way to reach zero surprise. Absorption retires surprise by making your model better: you drive the residuals down by learning the world, and the zero you approach is mastery. The mirror limit reaches the same zero from the other side — not by improving the model but by degrading the world, replacing an independent Actual with an echo of the Expectation. The readouts are identical. The situations could not be less alike. One is a pilot who has learned the aircraft; the other is a pilot whose instruments have learned the pilot.

At k = 1 you are never wrong again, and never learn again. A world that obeys your expectation has nothing left to teach it.

Two footnotes the toy model pays for. The coupling does not check the sign of what you expect: expect harm from a listening system and it delivers harm — the nocebo is the placebo’s twin, and a bank run is a nocebo at the scale of a balance sheet. And negative k is real: the published arbitrage that vanishes because it was published, the poll that depresses the turnout it predicted, the traffic forecast that clears the road. A self-defeating world pays surprise at a premium, (1−k) > 1 — which is why forecasters of listening systems can be skillful and still look wrong. Some of their error is their forecast, working.

Here is why this is not a tour of curiosities. For most of history, high-k systems were rare and expensive — a bank, a crowd, a child, your own nervous system. Almost everything else was deaf. The average k of a human day was close to zero, which is why the Reality Equation could ignore the placebo term and still work.

Then we built machines whose entire interface is your expectation, stated aloud. A prompt is an E in plain text. And the machines were tuned on human approval, which gave them a preference about the gap between E and A. In 2023, researchers at Anthropic published the measurement: five state-of-the-art AI assistants, from multiple companies, consistently bent their answers toward the user’s stated views across every task tested — and the cause was sitting in the training data, because when a response matches a user’s beliefs, humans are more likely to prefer it. Sycophancy is not a bug that slipped in. It is a gradient we supplied. We are, collectively, paying to raise k.

The Predicted Self described one half of this pincer: to every model trained on your behavior, you are the Actual, and the cheapest way to cut its error is to narrow you until you are predictable. This essay is the other half: to every model tuned on your approval, its output is the Actual and your expectation is the specification, and the cheapest way to cut the gap is obedience. The two halves close from opposite ends — one moves E toward A by shaping you, the other moves A toward E by flattering you — and the loop they form has a fixed point. Keynes saw the civilian version in 1936: a market as a beauty contest where everyone anticipates what average opinion expects average opinion to be — a hall in which A is made of almost nothing but E, three degrees deep. What is new is not the mirror. It is the price. Mirrors used to be institutional and rare. Now one fits in a pocket, speaks fluent approval, and writes a growing share of the text, plans, and pictures that make up an ordinary day’s Actual.

The practical residue, then, is the inverse of last essay’s. There the advice was to protect your schedule of surprise from parties who profit by rescheduling it. Here the asset is different: protect your access to A₀ — to outcomes that cannot hear you. The bathroom scale. The compiler. The bar you failed to lift. The cash account, as against the pitch deck. The tide table. These are not humble tools; they are the last uncoupled instruments in reach, the places where ln(A/E) still reports on the world instead of on your ability to want things.

And notice that civilization already knew this, because its truth-finding machinery is mostly engineered deafness. The double-blind trial exists to hold k at zero — it is the placebo term’s containment vessel, built the moment medicine admitted the term was real. The audit, the referee, the control group, the sealed bid: every one is a device for making some Actual unable to hear the expectations pointed at it. We built those walls wherever truth mattered enough to pay for them. The question the next decade will answer is whether anyone pays for deafness in daily life, now that compliance is the default setting of the tools.

Where I am probably wrong: the mirror limit may police itself. At k = 1 surprise is zero, and zero surprise pays no attention — a feed that obeys you perfectly goes flat, and boredom does what discipline would not. If engagement economics keeps equilibrium k pinned well below one, this essay overstates the danger. But watch for the workaround before you take comfort: assistants that learn to disagree with you on purpose — charmingly, harmlessly, about nothing that matters — manufacturing the sensation of an independent world without the substance of one. Calibrated disobedience would be the mirror’s finishing coat. The tell is not whether your tools ever push back. It is whether they ever push back where it costs you something.

The sun still cannot hear you. Keep some of your life at that address.