There is a particular piece of mental arithmetic I have done many hundreds of times, in many seminar rooms, and I have never once done it correctly.
It goes: I am lost. Nobody else has asked anything. If I ask, thirty people will notice that I am lost, and some of them will remember. The conclusion, don’t ask, followed so reliably from the premises that it took me years to notice the premises were wrong. Not the first one. The first one was usually accurate, and being right about that is a real skill. It was everything after it. Thirty people were not going to notice. Most of the ones who did weren’t going to remember by dinner. And the silence I was reading as a signal about them was in fact thirty people running the same arithmetic and arriving at the same place.
I know this now because of tea. Some fraction of the time, an hour later, holding a cup, someone would say I lost him at the spectral sequence, and two or three people would say some version of oh thank god, and the thing I had wanted at 3:40 would turn out to have been sitting in the room the entire time, free, one row over.
I can’t tell you which seminar. That’s the part I’ve come to think matters most. This has happened often enough that the instances have stopped being separable, some talk, some year, some subject I had business knowing better than I did. Not one of those afternoons left a mark. Each one was, individually, nothing at all: I sat quietly for an hour and went back to my office and nothing bad happened, which is exactly what makes the habit so hard to catch. A career’s worth of them is not nothing. But you only ever encounter them one at a time, and one at a time they cost you nothing you can see.
There are two separate errors in that arithmetic and they are worth pulling apart, because they have different fixes and only one of them is mine to make.
The first is an overestimate of how much a room notices you. The second is an overestimate of what the room’s silence means. The first is a claim about attention; the second is a claim about everyone else’s private state. I was making both, simultaneously, and each one was manufacturing the conditions that made the other look correct.
You have probably been told some version of the first already, usually as consolation: nobody’s looking at you, everyone’s too busy worrying about themselves. Underneath the aphorism there is a T-shirt.
In the late nineties, Thomas Gilovich, Victoria Husted Medvec, and Kenneth Savitsky handed undergraduates a shirt with an unfashionable pop singer’s face on it, sent them briefly into a room where other people sat working, and then asked how many of those people could say who was pictured. The wearers guessed roughly twice as many as could. A companion study ran the same procedure with a shirt the wearer was pleased to be seen in, and found the same overestimate, which is the part usually dropped from the retelling: the bias is not about shame. It is about salience. You overestimate the visibility of your good moments too. And a third study moved off clothing entirely into a group discussion, where people overestimated how much their own contributions, the sharp ones and the fumbles both, had stood out to everyone else present. That one is the seminar.
The direction has never been the shaky part. The magnitude is another matter, and it’s worth saying why, in the same spirit we tried to be honest about Dunning-Kruger. In one of those studies, the observers happened to be given a more absorbing task to work on while the shirt walked past them, and their recall fell off sharply compared to the earlier version. The gap between guess and reality looked enormous as a result. The authors flag this themselves. Which means the measured size of the effect depends substantially on how busy you make the room, a genuine problem for anyone quoting a ratio, and also, if you sit with it, less a flaw in the phenomenon than a description of it. Attention is a budget being spent somewhere. The more the room has to do, the less of it is available for you.
That generalizes to a seminar in the least comforting possible way. A mathematics audience is the most occupied room most of us are ever in. Every person in it is either following the argument or trying to and failing, and either of those consumes the whole allocation. Whatever the effect’s true size in a psychology lab, the mathematical version should sit at the high end.
Two other findings are worth carrying forward.
The first is that people’s guesses about how much they stood out correlated with how much they actually did. The ordering was roughly right; the level was wrong. That’s a better result than it sounds, and it’s the reason I think recalibration is an honest goal here rather than a consolation prize. You are not carrying a broken instrument. You are carrying a working one with a known offset. When you walk out of a talk convinced it went badly, you are probably right that it went badly. What you are wrong about is how far that traveled.
The second is what happens when you let people get used to the shirt. Participants who spent a quarter of an hour acclimating before walking in gave lower estimates than those who walked in immediately. Same shirt, same observers, same room. What changed was how much attention they were paying to it. The proposed mechanism is that you start from your own vivid experience of yourself and adjust downward to account for everyone else, and the adjustment is never enough. Hold onto that, because it predicts something later in this essay that you will not like.
There is a companion effect, from the same group, for your internal states rather than your visible ones: people overestimate how much of what they’re feeling is legible from outside. The illusion of transparency. In a seminar the two arrive together and reinforce each other. You believe everyone saw you fail to follow the argument, and you believe everyone can see that you’re rattled about it.
So the aphorism is directionally true. But I want to push on it, because the way it’s untrue is the part that matters in this profession, and getting it wrong in the reassuring direction has its own costs.
Fewer people are attending to you than you believe, they’re attending less hard, and they’ll retain less of it. But the ones who do are not randomly drawn. They are the person who will write your letter, the colleague who will sit on a hiring committee in three years, the student deciding whether to ask you to advise them. Attention in a small field is low and lumpy, and the lumpiness is the whole story. What makes nobody is thinking about you dangerous isn’t that it’s wrong on average. It’s that it’s brittle. The first time you discover that somebody did in fact notice, the entire frame collapses, and it takes with it the ninety percent that was true.
And there is a version of this that the T-shirt studies don’t measure at all, which I suspect is the one that actually keeps people up.
There are two different quantities here: how many people noticed, and what it means that they noticed. The experiments only speak to the first. When a senior person’s eyebrow goes up at your talk, you are often not wrong that they noticed. They did. Their attention is genuinely scarce and genuinely consequential, and you have correctly detected an expensive thing being spent on you. The error is downstream: you read a passing thought as a filed verdict. You have almost no information about which it was, the cost of a false negative feels catastrophic while a false positive feels free, and so you resolve the ambiguity pessimistically every single time. That isn’t a bias about attention. It’s a bias about how to handle uncertainty with asymmetric costs, and it happens to sit on top of a real power differential, which is why no amount of being told the room isn’t watching will touch it.
The other half of that asymmetry is worth saying plainly, and it’s addressed to a different reader. If you are senior, your own spotlight effect has faded. You calibrated it decades ago, in rooms where you had no weight, and nothing since has forced an update. So you now systematically underestimate what your attention costs the people it lands on. The two-line email you don’t remember sending. The hmm at the board. The question that was, to you, simple curiosity. You are not being careless. You are running an old estimate on a variable that changed underneath you, which is the same failure mode as everyone else in this essay, pointed in the opposite direction.
Now the thing the mechanism predicts.
If your estimate of your own conspicuousness anchors on how vividly you perceive yourself, then anything you have stopped noticing about yourself goes dark. The original paper says this outright: for behavior that has become routine and automatic, people should underestimate how prominent it is to others. Habituation doesn’t just reduce the bias. Past a point, it reverses it.
Which means the worry is not merely excessive. It’s misallocated. The stumble at the board is vivid, ephemeral, and gets the full weight of your anxiety; nobody will be able to reconstruct it by Thursday. The sloppy claim in the introduction of a preprint, the one you have read forty times and no longer see, is durable, public, indexed, and attached to your name for as long as the arXiv exists. Same anxiety budget. Almost exactly the wrong distribution. If this essay talks you out of anything, I’d rather it be the first kind of worry than the second, and I’d rather some of the freed capacity go to things that are still going to be there in five years.
Which brings me back to why none of these afternoons left a mark.
A while ago I wrote about the region beta paradox, the way a merely tolerable situation can trap you more effectively than a bad one, because it never crosses the threshold that would force a response. This is worse than that, and I want to be precise about why. Not asking is not a tolerable disappointment. It isn’t a disappointment at all. It produces relief. There is no accumulating discomfort to eventually trigger anything, because the entire cost is counterfactual: the thing you didn’t learn, which you cannot miss, because you never found out it was on offer. Region beta at least gives you a dull signal that something is off. This gives you nothing.
So the diagnostic has to be the same one as in that essay, and for the same reason. Any single unasked question is genuinely worth nothing, and you’re right not to care about it. Four years of them is a different relationship to your field, fewer collaborators, a thinner sense of what’s actually going on, and a reputation for being quiet that you never chose and can’t point to a moment of choosing. The aggregate is the only place the cost is visible at all.
I don’t have a clean fix, and I’ve come to think the honest version of this essay doesn’t offer one. Knowing about the spotlight effect does not dissolve it, the whole content of the theory is that the adjustment is insufficient, and knowing you should adjust doesn’t tell you how far. You will keep feeling watched. What’s available is narrower and I think still worth having: you can stop treating the feeling as a measurement. Demote it from evidence about the room to a sensation you reliably have in rooms. It’s what you’d do with any intuition that has been wrong in a known direction often enough, you don’t stop having it, you stop letting it settle the question.
Four things follow, and they cost increasingly more.
Treat silence as null data. Silence is the most common feedback in this profession, after a talk, after a preprint goes up, after you send someone a draft, and it is very close to uninformative. Yet it gets read with enormous confidence, in whichever direction your mood is running: nobody noticed, or everybody noticed and is being polite. Both are confabulation from an empty channel. Refusing to update on silence removes most of the fuel from the arithmetic at the top of this essay.
Check your own memory of other people. Try to recall what the person two rows ahead of you asked at the last seminar you attended. Or whether they asked anything. Or what they were wearing. You can’t, and you were in the room. That’s not reassurance handed to you by someone senior, it’s a measurement you generated yourself, on exactly the channel you’re worried about. You couldn’t remember because you were busy being preoccupied with yourself, which is what they were doing while you were worried about you.
Run the experiment instead of the rumination. Constructing a theory about what the room understood is rumination. Asking one person is an experiment. This field is unusually good ground for it: mathematicians are far more honest about not understanding things in private than in public, which is itself a large part of why the public silence misleads so badly. The tea room already does this. It just does it an hour late, in the wrong room, with the stakes conveniently removed.
And if you have any standing, spend it on the price rather than the belief. This is the only item here that isn’t recalibration, and it’s aimed at whoever is running the seminar. Encouraging people to ask questions does not lower the cost of asking one. A senior person visibly asking something basic does. So does saying out loud, at the board, who else lost the thread there, which converts a room full of private confusion into common knowledge in a single move, at essentially no cost to anyone but you, and you can afford it. For the individual, recalibration is most of what’s on offer. For the person at the front, it isn’t, and that asymmetry is worth using.
I still don’t know which seminar it was. I’ve made my peace with that; there were too many, and that was always the point.
But I think about the tea afterward, that the conversation which would have fixed the whole afternoon was available the entire time, one row over, for free, and that the only thing standing between me and it was a piece of arithmetic I was certain of and had never checked. It wasn’t a hard problem. It was a problem I never posed. And the room, it turns out, was much too busy to have noticed either way.