A senior colleague told me, more than once and always kindly, that if I wanted a raise I should go get an outside offer. He was not being cynical. He was doing me a favor; passing along, unprompted, the kind of thing nobody writes down and everybody is supposed to somehow already know. It was the method that had worked for him, and he thought I should have it too.
Here is what took me an embarrassingly long time to notice: the advice was almost certainly correct. Outside offers really do move salaries. That isn’t folklore, and I’m not going to pretend I discovered he was wrong. What I eventually noticed was something else, and it has bothered me ever since. I had received a fact about how institutions work. What I had actually been handed was one person’s career, described accurately, by someone with no way to know how representative it was, and no particular reason to wonder.
Think about who I could have heard that from. The person who got an offer and got matched is still down the hall, and will tell you the story, because it’s a good story with a clean ending. The person who got an offer, wasn’t matched, and left is at another university now and is not in my sample at all. The person who got an offer, wasn’t matched, stayed anyway, and now has a chair who knows they were shopping, that person is extremely present in the hallway and completely absent on this topic, because nobody tells that one. And the person who never tried is invisible to everyone, including themselves.
So the advice is an estimate produced from a sample where the negative cases have been removed, by a mechanism that guarantees you will never meet them. Not noise. Structural, one-directional truncation, selected on the outcome. Every mathematician reading this knows what that does to an estimator. Almost none of us apply it to the things we believe about our own workplaces, because those beliefs don’t arrive labeled as estimates. They arrive as how it works here.
Two errors, pointed in opposite directions
There are two well-studied mistakes underneath this, usually taught separately even though they’re the same problem seen from two sides.
The first is pluralistic ignorance: most people in a group privately reject some norm, but each believes the others accept it, so everyone publicly goes along, and their compliance becomes the evidence everyone else uses to infer that the norm is widely held. The classic demonstration is a campus drinking study from the early nineties, in which students were privately more uneasy about the drinking culture than they believed their peers were. Nearly everyone was more moderate than the average person they imagined.
The second is the false consensus effect: overestimating how many people share your view. It looks like the opposite, and in a sense it is. But they are the same estimation problem with different inputs. You are trying to learn a population parameter, what do people around here actually think. When there’s no public signal, you fall back on the one observation you have complete access to, which is yourself, and you over-weight it. That’s false consensus. When there is a public signal, you trust it, and you shouldn’t, because public behavior is not an independent draw from the private distribution. It’s a function of everyone’s estimate of that distribution. The thing you’re measuring is contaminated by the fact that it’s being measured.
My senior colleague was running the second error. I was running the first. Both of us in good faith, at the same table, producing between us a durable belief about the institution that neither of us had checked.
The mechanism connecting them was worked out by Dale Miller and Cathy McFarland in the eighties, and it’s precise: identical actions by yourself and others get read as reflecting different internal states. Their test case was a classroom. Students given incomprehensible material and told they could ask for clarification didn’t ask, and then explained their own silence by embarrassment and everyone else’s by comprehension. Situational for me, dispositional for you. Their summary of the resulting state is the best one-line description of a seminar room I know: no one is certain, but everyone is certain that everyone else is certain.
Two things are worth knowing about the state of this research, and then I’ll leave it alone. The measurements are mostly self-report, which is a weak instrument, but there is at least one good field experiment, correcting men’s beliefs about what other men thought, in a setting where the belief governed a real decision, changed what they actually did months later. And on the other side: the interventions built on this idea, the most students drink less than you think poster campaigns, have a poor record. A large review of the alcohol literature found no benefits worth the name. Broad impersonal correction does close to nothing; correction aimed at a specific person facing a specific decision does better.
That combination, real mechanism, failed fix, is the most useful thing here, and there’s a structural reason for it. Pluralistic ignorance is not a shortage of information. It’s a shortage of common knowledge. Everyone knows the emperor is naked; not everyone knows that everyone knows. Reassuring five people privately creates mutual knowledge among five people and moves the equilibrium not at all, because each still doesn’t know the others were reassured. What’s missing isn’t a fact. It’s a fixed point of the “everyone knows that” operator, and only a public act produces one.
The number is visible; the map is not
Salary looks like the exception that should break all this, because at a great many institutions it is simply public. If you’re at a state university, your colleague’s compensation is likely sitting in a searchable database and you can have it in thirty seconds. Private institutions disclose far less, the federal filing covers officers and a handful of the highest-paid employees, which in practice means administrators, not the analyst down the hall.
But transparency here is narrower than it looks, and in a specific way. You can see that a colleague makes eleven thousand dollars more than you do. You cannot see why. Outside offer? Retention package? A better negotiation in a hotter hiring year? An administrative stipend that was never removed? Compression, or the absence of a fix for it? The level is public. The mechanism that produced the level is entirely private.
That’s worse than opacity in one particular respect. You now hold a precise, credible, vivid observation that demands an explanation, and the only explanations available to you are exactly the folklore this essay is about. Transparency without mechanism is a rumor generator.
One experiment on this, run across the University of California, found the effect of learning peer salaries to be one-sided: people who discovered they were below the median for their unit reported lower satisfaction and more interest in leaving, while people above it reported no improvement at all. Publishing the number doesn’t correct the norm, because the norm was never a claim about numbers. It was a claim about how the place works.
The invisible cases everywhere else
I’ve spent this long on money because it’s the one place where the truncated sample is legible, the counterexamples physically left the building, and their absence is easy to point at.
Everywhere else the missing cases are merely quiet, which is much harder to notice. The person who published less than the folklore says you must and was fine. The person who declined the service request and found that nothing happened. The person who kept ordinary hours through the whole clock and got tenure anyway. They are not gone. They are down the hall, and you will never hear a word about it, because nobody tells a story about a thing that didn’t go wrong. There is no genre for it. So the stories in circulation are drawn, systematically and without anyone intending it, from the tail where something dramatic happened, and everybody’s model of the standard is fit to that tail.
Tenure is the sharpest instance, and not because the process is unfair. It’s because the process is designed, for defensible reasons, since candor requires confidentiality, to emit exactly one bit and no reasons. The field is trying to learn a decision boundary from a small number of labeled examples, with the features unobserved and the labels noisy. When the loss is that asymmetric, the safe estimate is maximum effort, and everyone converges there independently, without ever comparing notes. That isn’t irrationality. It’s the correct response to a channel that carries almost nothing.
And the committee is running the mirror image. Each member privately thinks the case is fine but assumes the others hold a firmer line, so everybody hedges slightly in their written language, because nobody wants to be the one who appeared to have low standards. The document that comes out is more demanding than the room that produced it. Then that document becomes evidence, for the next cohort, about what the field expects.
What the transmission actually looks like
Here is the part I find genuinely unsettling.
I learned how this place works the same way my colleague did: someone senior told me, in good faith, and they were projecting too. And whoever told them was projecting. You can follow that chain back as far as you have patience for and you will never arrive at a measurement. At each hop, a personal experience is reported as an institutional fact by someone whose standing makes the report sound like data, and the confidence increases rather than decays, because the alternative, saying I only know this happened to me once, is a visible reduction in your own authority, and almost nobody volunteers that.
So a norm governing how thousands of people spend their weekends propagates entirely by inference and is verified nowhere.
It also explains the durability, which otherwise ought to be puzzling. Pluralistic ignorance among peers is fragile, one tea room conversation can crack it, which is roughly what the last essay was about. This isn’t peer to peer. It runs across a power gradient, where the junior person has every reason to accept the estimate and no standing to test it. That’s a much more stable arrangement, and it survives generations of people who each privately thought it was too much.
I should say the obvious thing, which is that I am now on the other end of this. People ask me how things work here and I answer, and I mean well, exactly as he did. I would like to offer you the anecdote in which I caught myself doing it. I don’t have one, and I’ve come to think that’s the honest report rather than a gap or more generous telling; the whole claim here is that you cannot detect your own projection from the inside. An author who confidently produces the instance where he did it has quietly refuted his own argument. What I have instead is the structural knowledge that I have no access to my own sample size, and that some fraction of what I say with confidence is one career, reported as the world. I don’t know which fraction. Neither did he.
The second-order problem
The Princeton drinking study followed its students across a semester and found that private attitudes moved: some resolved the tension by internalizing the norm they believed everyone else held, others kept their private views intact and went on conforming anyway. Either way the norm survived, and in one of those routes it stopped being a misperception at all.
Run that forward in a department. The population that remains is a filtered sample, selected on willingness to meet a standard that nobody chose. Ask them what the expectation is and they’ll tell you, accurately, that it’s very high. And they’ll be right, because the belief manufactured the consensus it was originally wrong about.
There’s a distributional problem sitting on top of it. Go get an outside offer is not equally executable. It requires mobility, and mobility is distributed by a partner’s job, a child’s schooling, a visa, an elderly parent, and whether anyone happens to be hiring in your subfield this year. It looks like universal advice. It is a lever a subset of people can reach. And the people who can’t reach it do not generally conclude that the norm is partial, they conclude that they are paid less because they are worth less, which is precisely the inference the missing half of the sample is built to produce.
What to actually do
The difficulty is that this essay has spent most of its length arguing that the usual fixes fail; private reassurance, costless statements, published averages, individual scrupulousness. So anything practical has to survive all of that, and what does survive splits by who can do it, which is part of the point.
If you are the one receiving the advice:
- Ask about the sample, not the claim. The useful question isn’t whether the advice is true. It’s how did that go for other people you’ve watched try it? This is askable, it sounds like interest, not doubt, and it does the entire job, because it forces the sample size into the open. Often the honest answer turns out to be: once, on me.
- Get multiple estimates and read the spread. This is the aggregation argument, executable by someone with no power at all. Ask several people independently, and treat the disagreement as the finding. Different accounts of the standard means there is no standard, and the confidence in any single telling was unearned. Convergence is informative too, but you can’t tell which case you’re in from one data point, and one data point is what most people are working from.
- Ask about the mechanism, not the level. Especially where numbers are already public. How do raises actually get decided here? is procedural, answerable, and not awkward. It’s also the question that matters when you find yourself below someone: not am I valued less but was that produced by a process that could also apply to me? Nearly all the damage lives in attaching the first interpretation to an observation that only supports the second.
If you are the one giving it:
- Report the mechanism, not the encouragement. Telling a junior colleague they’re doing fine is cheap, and gets discounted accordingly. Telling them how a committee actually reads a file is expensive and doesn’t.
- Say what you don’t know. This is what happened to me, once, and I don’t know how representative it is costs you something real, a visible reduction in your own authority, which is exactly why it registers. Cheap talk fails here. This isn’t cheap talk.
- Aggregate, and publish the spread. If you have standing: poll the department anonymously about what people believe the expectation is, then show everyone the distribution at the same time. Not the mean, a mean is a target, and people move toward targets. The spread is the finding, and it has to come with a plain statement that the low end is fine, or you’ve just handed everyone a number to drift toward. Revealing it to everyone simultaneously is the only move here that produces common knowledge instead of mutual knowledge.
And then the limit, which I don’t want to leave out, because leaving it out would do its own damage.
Some of this is not a misperception. The outside-offer route really is unavailable to some people, for reasons that have nothing to do with belief. The service load really is distributed unevenly. Some departments really do hold the standard the folklore says they hold, and if you run the anonymous poll you may find out that they do, and then you’ll know, and you won’t be able to un-know it. None of the tools above touch any of that. Telling someone that their accurate observation of a structural asymmetry is a distortion is its own particular harm, and it’s a harm that essays like this one are unusually well positioned to cause. Distinguishing the two cases is most of the work, and I don’t have a clean method for it either.
I don’t think my colleague did anything wrong. He gave me the best information he had, generously, at no benefit to himself, and it was true, and there is no version of that conversation where he does better by being more careful, because he would have had to know something he had no way of knowing. What stays with me is that I did the same arithmetic he did, got the same answer, and passed it on with the same confidence, and never once in that entire chain did anyone check. It was a fact about the institution assembled entirely out of things people had said to each other. And it governed, for years, how I thought about a question I could have asked much more directly, of many more people, at a cost of nothing.