Blown Up

There is one photograph of my mother in this house. A photographer took it for the state per its exacting instructions and needs.

A vesikalık = passport picture, the little square you hand over for a hüviyet = identity. You went to a shop for these. Somebody's studio, lights and a stool and a roll of backdrop, a man who did this all day. He would have set the lamps and told her where to look, maybe tilted her chin. Then he printed a sheet of them and put the sheet in a paper envelope, and she carried the envelope to a government office and handed the pictures over to be affixed and stamped on various other identity things or rather.

Think about what he was allowed to keep from Leyla. The size was fixed, the background was fixed, the head had to fill a set proportion of the frame. Everything the man actually knew, whatever he had learned in thirty years about where to put a light and how long to wait until somebody stops performing and just sits there, had to survive inside that or not at all. She smiled anyway, and that got through.

The rest of her was never in the specification to begin with. The office needed enough of her to attach a face to a name in a register, and it needed that in quantity, cheaply, at the size of a thumbnail, six to a sheet, spent across a lifetime — the school, the bank, the clerk who wants one more before he will proceed.

Years later, Hasan the cousin had it enlarged for me. A gift, and a good one, blown up ten times to the size of a real portrait in a silver frame, and it has been on the table by the window ever since with a Christmas cactus half in front of it.

What came back bigger was not my mother. You cannot enlarge your way to information not recorded. Instead you have the encoding: the halftone screen it was printed with, the crease where the card was folded, and the seal, which now runs across her hair at ten times the size it was stamped.

And it still reads as a face. Nothing in it ever says insufficient data, please stand by. The form had a box and something had to go in it, and the box was the same size whatever walked into that studio. The output was required and the information was not, so fidelity is the only thing left that can give.

It took me a while to see the trap in that. The photograph is not false. It is a true record of my mother, made honestly by a man who knew his job, and it would still be a true record if there were forty others in a drawer. It goes wrong only by being the only one, because then it starts answering a question nobody built it for. What was she like. Nobody ever asked that photograph that. It answers anyway, every day, on that table by the window.

So then what? Every remedy I have heard for what has happened to public life is enlargement. More fact-checking, better moderation, more nuance in the discourse, another debate, a longer piece. Every bit of it is work performed on the output, downstream, after the encoding, and I have sat in rooms where grown adults called that a strategy. If the distortion got in upstream — if the categories arrived because the channel could not carry anything finer than categories — then all that effort does what Hasan's did. It gives you a bigger, sharper, more expensive picture of the seal.

I have been arguing for about a year that societies do the same thing. Three decades inside institutions, watching them throw information away on purpose, and I wrote it up in April 2025 as Bandwidth Economics. (A year of arguing. What follows is one coefficient with four decimal places.)

Institutions are codecs. A hierarchy is a rule that says track your boss, not everyone. Ritual says run the script instead of recomputing the right action every morning. Durkheim said roughly that and got filed under conservative for it. Ideology is the same trick at scale. Apply the schema, skip the tradeoffs. None of that is a bug. Codec is the only reason a group bigger than a dinner table works, and all of it runs on throwing information away.

Take that seriously and two things fall out. One is mechanical and a little boring — throughput rises, capacity does not, the codec discards more, and what it discards stops being redundant and starts being signal. Distortion rises on its own. Nobody has to get dumber or meaner. It is the framerate. Chaos and hilarity ensues.

I want to be careful with that, because it can be made to excuse far too much. Capacity explains why something has to go, and it says nothing whatever about what goes. Power decides which losses get tolerated, which get written into the procedure, and which get engineered on purpose. Dressing that last one up as a bandwidth problem would be its own act of compression. This essay is about the first kind, and the first kind does not cancel the second.

The second one is why I keep writing about this. Squeeze the channel far enough and the cheapest code that still says anything is one bit. Us and them, good and bad. A one-bit quantizer over a continuous moral landscape loses the detail, and then does something worse: it manufactures two camps, maximally separated, where the distribution was a smooth hill. Kant said we never meet the world raw — it arrives pre-sorted by categories we brought with us. He meant a fact about minds. I think it is a fact about bandwidth, which is worse — the categories arrive whether or not there was ever enough signal for them. The Romans would have called that a genus, and there are two of them now. The Turks say ya devlet başa ya kuzgun leşe. The crown on your head or the raven on your corpse. Nothing between, and never anything between, which is the phrase and which is also one bit.

So the hypothesis the photograph handed me is that polarization is what a channel does when it can only carry a rectangle. The codec doing its job, with nobody's tolerance necessarily entering into it. That is what I set out to test. It is not what I can tell you I found.

Two different failures are hiding in there and they are worth pulling apart, because Leyla's picture is one of them and polarization is the other. The picture omits. Everything in it is true, almost everything about her is gone, and nothing was invented to cover the gap. The quantizer does the opposite kind of damage. Drop two reproduction points onto a smooth hill and they land where nobody was actually standing, so the code invents the very gap it then reports back to you. Omission leaves no trace and you never learn to look for it. Invention leaves a finding, and people act on findings.

There is a third, and it is the one that actually got me. A representation built for a narrow purpose stops standing in for the thing and quietly becomes the thing. That picture was made so a clerk could match a face to a name. It is now what my mother looked like. Nobody decided that. It happened by attrition, because everything else went and this stayed for the government. (To be sure, gov wants a new photo every so many years so they have an update mechanism).

It is a good story, and I have told it at dinner for a year and nobody has pushed back once, which usually means a story is doing no work, (or that there is no story). So I went and got the data.

All of it public. Boxell, Gentzkow and Shapiro keep a replication archive with comparable feeling-thermometer series for twelve rich democracies. V-Dem has an expert-coded measure of whether a society has split into hostile camps, for about 180 countries, which is the only polarization panel wide enough to run fixed effects on. Internet penetration from the ITU and the World Bank stands in for throughput. Nothing proprietary. You can re-download all of it this afternoon and argue with me by dinner, which I would prefer to you taking my word for any of it.

Those twelve got the same shock. Adoption speeds sit between 2.87 and 3.38 percentage points a year, half a point of spread across the whole group, the same wire in the same decade. What came out of it runs from the United States at +5.60 points of polarization per decade, down through zero, to Germany at -3.72. Germany depolarized. And Germany took the wire at 3.03 points a year while the United States took it at 3.06, which is about as close to the same treatment as observational data ever gets. Same wire, same years, opposite direction, while I was at dinner explaining that the wire does it.

That is my story getting mugged.

I ran the wide panel too. A hundred and seventy-four countries, 1990 to 2023, country and year effects, the standard machinery. It returns a small positive number, b = +0.0034, and there is your coefficient with four decimal places. I sat with it for a while before I understood what I was looking at. It is the average of the United States and Germany. Two effects with opposite signs, averaged, and then reported as though it described a country.

No country got +0.0034. Drop one reproduction point onto that spread and it lands where nobody is standing.

I built a quantizer and published its output as a finding, in an essay about quantizers.

There is a plainer way to say what went wrong, and it is worth one paragraph. My own theory says the institution is the codec. That makes the codec the thing that decides what throughput does to you, which means it belongs in the model as a condition, not a nuisance. Then I ran a specification whose first move is to subtract out each country's fixed characteristics, which is exactly where a codec lives. I removed the variable the theory is about and estimated what was left over. The number was always going to be small. It would have been small if I were right and small if I were wrong, and it cannot tell those two apart.

Underneath that sits the part that took me longest to say. I needed a measure of how much goes through the channel, and I used a measure of who has a connection, which stops moving in rich countries right around the time consumption starts climbing hardest. I needed a measure of resolution, and I used a score that a panel of academics wrote down after reading the same newspapers that have spent fifteen years explaining that the internet polarizes.

Both of them are vesikalıks. Standardized, usable, made to somebody else's specification for somebody else's register, missing everything that fell outside it. I used them because they are what somebody happened to record. That is the move this essay is about. I made it for a year and did not notice.

So I tried to build the instrument the theory actually needs. There is a card sort and a thermometer battery at unsal.com/academic that tries to score how many distinct categories your political map holds. Six minutes, anonymous, you get a number back.

It does not work yet, and I am going to tell you how it fails rather than wait until it doesn't. Before spending anything on a real sample I wrote down the condition the instrument had to meet: the two halves had to agree with each other at r = 0.25. On the first thirty-four people they came in at -0.06. The sort and the thermometer are not measuring one thing. I do not yet know which of them is wrong, and until I do there is no point buying a bigger sample of the same confusion.

Two levels, one hole. The panel cannot see resolution. My own instrument cannot agree with itself about resolution. Every polarization measure ever built records how far apart two camps feel, and the claim I want to test is about how many camps there are. That is why there is nothing here to show you. It is not the same thing as having looked and found nothing.

What would settle it needs two things I do not have.

One is a feed. Take a population's own feeds and hold the mix constant, same sources in the same proportions, then turn the volume down. Rate, on its own. The deactivation studies have never isolated it, because switching a platform off moves rate and content and ranking in a single stroke. Then watch, over weeks, whether the map coarsens. That experiment can only run inside a company that owns a feed. Meta, Google, TikTok, X, Reddit. That is the whole list.

The other needs no feed at all. Whether distortion lands hardest on the atypical case, the edge case, the person who does not fit the template, is a tail-behavior question, and it runs on model interactions. OpenAI, Anthropic and Google DeepMind are already arguing about it internally with better instrumentation than I will ever have. That one is a request for an eval, not for a feed.

Either half, the design is written and I will hand it over. Co-author it, preregister it, publish it however it comes out. I would love to tell you I have no attachment to being right. I have an enormous attachment to being right. If it lands null I will write that essay too, and it will be the better one. Friends at any of these fine institutions: heed my call!

The picture on that table is as big as it is ever going to get. Hasan did the loving thing and the futile thing in a single act, and I have never once wished he had not. Everything anybody would actually want was ruled out before the shutter opened, by a specification that wanted a face for a register and got exactly that.

Ya devlet başa ya kuzgun leşe. The crown or the raven, no middle. The middle is where nearly everyone I have ever met actually lives. I would like to know whether we are doing that to each other now.

Boxell, Gentzkow & Shapiro (2024), "Cross-Country Trends in Affective Polarization," Review of Economics and Statistics. V-Dem v2cacamps via Our World in Data. Internet, broadband, mobile, GDP and urbanization series via ITU/World Bank through Our World in Data. The framework essay this builds on is "Bandwidth Economics" (April 2025).