Essay · August 2026
AI Slop, AI Slobs and the Innocence of the Prompt
It has become fashionable to rail against the flood of “AI slop,” and I share the frustration that fuels the outrage. The tells are familiar by now. Standard constructs, standard triads, everything “lands” and everything is “load-bearing,” and every other sentence informs us that something is not this but that. The graphics carry the same default color scheme, the same layout, the same symbols. It begins to make everything look like a fake Hollywood building: empty calories, a flashy knockoff of the thing it imitates.
Vacuous, performative text is nothing new. Throughout my career I have written documents, and read many more, that I can confidently describe as corporate slop. They followed institutional conventions, responded to corporate mandates, repeated approved formulations and inflated modest ideas into pages of centrally managed prose. Much of this writing was neither particularly clear nor insightful, and it was not expected to be either. It served to demonstrate compliance: it signaled that everything was under control and operating in the prescribed order. Consultations had taken place, positions had been accommodated, the required references included, the process observed. Frequently the document existed simply because the organization expected a document to exist.
The same applies beyond corporate life. Anyone who has persevered through a long-running series of novels will recognize some of the alleged hallmarks of AI writing: recycled descriptions, standardized dialogue, inconsistencies in logic, repeated explanations and plots extended beyond their natural life.
Human beings produced generic and insufficiently edited work long before machines offered to help. Generative AI did not invent performative writing. It merely became very good at it, and very fast.
What speed changes
Speed is the part that is genuinely new. The production of slop has lost its friction — the rate-limiting resource was always the time of the person required to produce it, and that constraint is gone. It is now extremely cheap and almost stupidly simple to ask one of many generative systems for some generic work.
And that changes how written text can be valued at all. The volume of information was overwhelming well before generative AI; today it is worse by orders of magnitude, drowning out and degrading the worth of everything it surrounds. Worse, most of this output will be used to train future models, degrading those in turn; and by sheer volume it is bound to establish a new normal for what counts as acceptable prose. No filtering mechanism and no amount of human review can cope with the quantity now flooding our sources. The problem predates generative AI. The scale does not.
The slop and the slob
So I began to follow the thought: who is putting AI slop out? And, more importantly, why?
Set the bots aside for a moment — they were unhelpful to the ordinary user long before there was AI. For the rest, and I would argue for the majority of what gets called AI slop, a person typed a lazy prompt and pressed publish. Which suggests that “AI slop” is the wrong frame. It blames the machine, and hands it all the agency in the transaction.
Perhaps the more accurate term is AI slob. The slop is the product. The slob is the person who requests a text, gives it barely a glance, and releases it into the world under their own name. The word is ugly on purpose. It puts the ownership of the problem where it belongs.
The central failure is one of human judgment. Asking an AI system to prepare a draft and publishing its first response is not fundamentally different from asking a colleague or an intern to write something and forwarding it without reading it. In both cases a person has delegated the work while quietly abandoning responsibility for its quality. AI makes that abandonment unusually convenient, because its prose often looks finished. It arrives grammatically intact, neatly structured, and wearing the confident expression of someone who has never considered the possibility of being wrong.
That fluency is the danger. Awkward writing alerts us that something may be amiss. AI can present weak reasoning, invented facts and empty generalities in an assured and orderly voice. It can make an argument sound settled when it has barely been considered. The machine is doing what it was built to do; its confidence does not excuse our credulity. If polished sentences persuade us to stop thinking, the intellectual laziness is ours.
The innocence of the prompt
The real trap is the innocence of the prompt: a blank window, begging us to type something into it and then watch a stumbling thought come back with an eloquence and confidence we wish we had. And the prompt is not the only instruction in the room. Much of what returns is shaped by system prompts the user never sees, and by the AI’s behavior rewarded during training. The window looks like a blank page, but the page has already been written on.
Underneath the psychology there is a transfer of agency, or at least a borrowed authority. The story goes like this: a person has a thought, a vague one; generative AI heralds it as a deep and novel one, then produces output that feeds straight into the confirmation bias and sounds, at first glance, compelling. Not unlike the very confident friend who always has an answer for everything — except that we have learned to discount the friend’s opinions. We are tempted to believe the machine because our instincts were formed on computer programs that behaved deterministically. That assumption no longer holds, and we have not adapted.
Generative AI can be a powerful companion — to red-team a position, to pressure-test a structure — and it excels at that, provided I state it as my intent. But when I supply an innocent prompt, it returns an essay that sounds like independent truth, when it is the output of a machine doing exactly as instructed: following the user’s prompt. We then treat that output as though it were separate from us, and it validates us. Our prompts are not innocent. They are instructions that produce text justifying the thought we typed in.
Where blame is the wrong instrument, and where it is not
None of which makes this an individual failure alone. It is a collective action problem: the sum of individually reasonable decisions producing an outcome nobody chose. No single person can change the trajectory of the information ecosystem, and nobody can meaningfully opt out of it. The pressure to use these tools is not imaginary, and neither is the cost of refusing. I have no solution to offer, and I am wary of anyone who claims one.
But there is a distinction worth holding onto, because a great deal turns on it. Directing blame at an individual is the wrong instrument where responsibility has been distributed until no one is left holding it — where the harm of the collective action is real, yet every individual contribution to it is negligible. It is the right instrument where a specific person performed a specific act and put their own name to it. The slob did not cause the flood. The slob is the person who declined to read what they were about to publish under their own name, which is a small and entirely voluntary act that no account of systemic pressure makes disappear. Systemic critique explains why the flood exists. It does not explain why any particular person added to it without looking.
None of which makes the slob contemptible. The mandate to use more AI is real, the deadlines are real, and most people are simply trying to get through the week. But sympathy must not be conflated with absolution. Understanding why someone did not read the thing is not the same as concluding that they did not have to.
My own practice, and my own lapses
I use AI extensively. I also maintain that my thoughts remain my own, although this claim requires more than having supplied the original prompt. AI can influence the framing of an argument, sharpen some distinctions and flatten others. It can suggest a direction that had not occurred to me, or make a weak idea sound more convincing than it deserves. Retaining ownership of the thought therefore depends on remaining willing to argue with the output, reject its formulations, check its claims, and remove whatever adds only polish.
Used in that way, AI is less an author than an unusually patient editorial companion. It can test an argument, expose repetition, and find language that conveys an idea more precisely. It can pressure-test my own thinking and let me abandon an ill-conceived idea early. It can also produce paragraphs of generic uplift with breathtaking efficiency. The user has to know the difference.
It would be disingenuous to finish without admitting that I have fallen into the trap I have just described. Of course I have typed half-baked thoughts into the prompt window and felt vindicated by what came back. Of course I have been handed hallucinations that fitted the picture so perfectly they only confirmed what a brilliant thought I had had. Some of it I circulated to colleagues for input, and they were gracious enough to point out the flaws. Mildly embarrassing: I did look rather like the boomer swept away by modern technology. Since then I have learned to write better prompts — clearer intent, more background, more scaffolding, as the term goes — and to read the output far more critically. In almost all cases I now ask the system to pressure-test its own output before I read it. It all helps. Mistakes still happen.
To borrow and adapt David Sedaris, AI makes me talk pretty. That is genuinely useful. Prettiness, however, is not thought. It can clarify an idea, but it can just as readily disguise the absence of one.
“AI slop” is therefore a fair description of a great deal of what now surrounds us. It becomes hypocritical when it allows people to treat the technology as the sole author of material they requested, approved and distributed. The machine may have assembled the sentences, but it did not decide that those sentences deserved anybody else’s attention.