Slop was never about the AI

You have met this man. Everyone has met this man. Call him the Deputy Head of Something, and he’s twenty minutes into a lecture nobody asked for. Senior enough that no one can leave, and confident in a way that has never once correlated with being right. He is explaining, to a room of people who know better, how some grand sweep of history proves whatever he already believed when he walked in. The facts are wrong, but it hardly matters, because they aren’t connected to one another anyway. The logic is a rope bridge with no far side. And the room nods along. Someone will later describe him, without irony, as a big-picture thinker.

Nothing he says is his. Every phrase is pre-owned: an extrapolation from the back-cover blurb of a book in a dentist’s waiting room, a keynote he half-watched, a thing a consultant told him in 2019 that impressed him mainly because of what it cost. He is doing next-token prediction over a corpus of prestige noises, and he has been doing it his whole career. If a model produced this monologue, we would screenshot it as proof that the machines are ruining discourse. Because it comes from a man with a pulse and a title, we call it perspective. And nobody in the room says otherwise — because correcting a man like this is a career decision, while sneering at a chatbot is free.

Everyone is worked up about AI slop. My position is that this is a dumb thing to be worked up about. Not because slop is fine; slop is a real problem, and I want less of it in my life as much as anyone does. The dumb part is the adjective. Slop is a human problem wearing a new mask, and the panic over the mask is a clean example of something I have taken to calling human exceptionalism: the mostly unexamined belief that if a person did it, it must be better.

Consider how little of the slop in my life has ever needed a model. A travel insurer I bought one policy from seven years ago has emailed me marketing roughly twice a day ever since. No model wrote those emails, at least not at first, and they have never once been worth reading. Or go bigger. Around 2010, the content farms (Demand Media, eHow, and a long tail of imitators) paid writers about fifteen dollars an article, and often much less, to flood search engines with plausible-looking pages that nobody involved believed in or expected anyone to finish. Demand Media alone was commissioning about four thousand new articles a day. The flood got bad enough that in 2011 Google shipped an update to its ranking algorithm specifically to kill it, one that touched nearly 12 per cent of all search queries. That was “write me a thousand posts and make no mistakes” executed entirely with human fingers, fifteen years ago. Nobody concluded that humans were the problem. We called it spam, and we punished the behaviour.

The exceptionalism runs deeper than content, though. People dismiss these models as “advanced autocomplete”, as though we were something categorically different. We are not. A human being is, among other things, an extremely sophisticated pattern-recognition machine, and the overwhelming majority of what we do is automatic. Your breathing, your heartbeat, and your immune system are running right now without consulting you. So are your habits. Spaghetti goes with meatballs; gin goes with tonic. Try drying yourself after a shower in an order that is not your usual one, and notice how wrong it feels the whole way through. Even the parts of life we experience as active thinking are mostly reproduction — we output previously synthesised material, phrases we have absorbed, positions we have inherited, formats we have seen work before. Unique thought is rare and expensive, and we ration it accordingly. The Deputy Head of Something is not an aberration. He is the default setting, with a salary.

So why did the anger organise itself around the tool? Three reasons, I think.

Firstly, the obvious one. Admitting that a model does much of what we do, faster and for nothing, is uncomfortable. “Advanced autocomplete” is not a description; it is a defence.

Secondly, the models broke a signal we relied on. Substack co-founder Chris Best made this point sharply in a recent conversation with Nate B Jones: a long piece of writing used to be its own proof of work. However bad the thinking, somebody had at least spent the attention to type it out. That signal was never reliable (the Deputy Head puts in the hours too), but it was expensive to fake, and now it is free. People have lost a filter they didn’t know they were using, and they are angry at the thing that broke it. Fair enough, as grief goes. But mourning a signal is not the same as identifying a cause.

Lastly, and I think most tellingly: AI slop is the first slop with no feelings to hurt. Nobody in that room could tell the Deputy Head what his monologue was worth, because hierarchy and politeness protect human slop, and they always have. A chatbot’s output is the only garbage we are socially permitted to name. The culture finally got a licensed target for a frustration it had been swallowing for decades, and then drew the category boundary around the tool, because the tool is the only participant that cannot be offended.

There is one objection that deserves to be met head-on, because it is the strongest one. Best’s version of it is that generated language at sufficient scale becomes a denial-of-service attack on the public square. Even when the output is good, every reader now pays a tax: is this real, is anyone home, will anyone be there if I answer? I don’t wave that away; the economics did change. The Deputy Head could only waste one room at a time, and his drivel at least cost him twenty minutes of his own life. The thousand-post publisher pays nothing, and externalises the whole cost of reading onto everyone else.

But look at what the objection actually condemns. It condemns a behaviour — publishing at volumes you have not read yourself, asking for attention you never paid — and the behaviour does not require a model. The content farms mounted precisely this attack with cheap human labour, and the response that worked did not involve deciding that people were the problem. It identified the behaviour and made it unprofitable. A model is an amplifier; it scales whatever judgement is applied to it, including none. Aim at judgement and you catch the slop no matter who typed it. Aim at provenance and you convict the writer who thought carefully with a model’s help, and you acquit the Deputy Head.

Because the pattern, in my own use, is boring and consistent. When I hand a model poor context and accept its first answer, I get crap, and the crap is mine; I chose not to think. When I give it the full context, argue with it, and iterate until the output says what I actually mean, I get work I am willing to stand behind. The tool has never once been the variable. Jones has a test for slop that I think is the right one: did the author show up? Did they choose the thought, and will they answer for it once it is public? Notice that the test is blind to provenance. It convicts the thousand-post publisher and the Deputy Head in the same breath, and it acquits the writer who used a model at every step and arrived somewhere considered.

The test also has a reader-side half that gets less attention. Slop survives because rooms nod along. Every disengaged audience that mistakes confidence for insight, every inbox that never unsubscribes, is the demand side of the market. Judgement is not only the producer’s obligation.

One last thing worth noticing. Somewhere in the past year you acquired the phrase “AI slop”. You did not derive it. You absorbed it from ambient repetition, the way I did, and the way you absorbed the confidence that human output is the better kind — which is, of course, the exact behaviour the phrase condemns. That does not make the underlying worry wrong. It should make the exceptionalism harder to sustain.

The AI is not the problem. The people are, and the judgement is, on both sides of every page. Spend yours deliberately.