On the Rise of Work-Slop and the Decline of Critical Thinking

There is nothing so useless as doing efficiently that which should not be done at all.

By now it is hard to get through a working day without touching AI, and mostly I am glad of it. I use it constantly. I am faster with it, I make fewer of the small mechanical mistakes I used to make, and I have become better at stating what I actually want (writing for a machine turns out to be decent training for writing for people). I would be dishonest if I claimed the gains were not real. They are real for me, and I assume they are real for most people who use these tools with any seriousness.

But everyone I know has had the other experience too. The document from a co-worker that is polished, complete, and somehow empty. The reply that answers at length without answering. The task turned around so fast that you do not trust it, and you catch yourself checking it more carefully than if the person had just said, give me until Friday. Something arrives that has the shape of work. Whether any work happened inside it is a separate question, and lately it is a question you have to ask. There is a name for this now. Researchers at BetterUp and Stanford started calling it workslop (in their survey of 1,150 American workers, 41 per cent had received some in the previous month), and the name stuck because everyone recognised the thing immediately. Naming and measuring are useful. But the numbers describe the cost of the phenomenon, and I think the mechanism of it is both more interesting and worse. Start with what work slop actually is, because "low quality AI output" misses it. The defining property of slop is that it is rarely bad in any visible way. Bad work announces itself; you read two sentences, you see the problem, you send it back, and little harm is done beyond the wasted minutes. Slop reads well. The formatting is right, the sections are all present, the tone is confident and professional. What is missing is hard to point at, and for a while the best I could do was to say the document had no spark.

I think I can do better than spark now. What is missing is safety.

A project requirements document was never valuable as a stack of pages. Nobody has ever read a PRD for pleasure. Its value was in what it certified: that a person had walked through the project in their head before anyone spent money on it, had fallen into the pits, and had marked them. When you build on such a document you are borrowing someone's traversal of the territory. You get to move fast precisely because someone else moved slowly. That was always the economic function of these documents, and the function is invisible, which is the problem, because the visible part (the pages, the headings, the confident tone) is the part AI reproduces perfectly.

A slop document is a map of terrain nobody walked. Which makes it worse than no document at all. A missing map tells you to go carefully. A false one invites you to build at full speed on ground no one has checked.

People sometimes tell me that handing work to AI is no different from delegating it to a junior colleague, and that managers have delegated since management existed. I like this objection because it is almost right, and the way it fails is the whole point. Watch what happens in ordinary human delegation. A manager hands off a task, usually with the words this should be simple. It is never simple. The subordinate discovers, one by one, the hundred things the manager never thought about, and the discovering is loud: questions come back, objections come back, somebody appears at a desk saying this cannot work the way you described it. The manager's thinking was incomplete, but the gap gets filled by another mind, and the company as a whole ends up holding the completed thought.

Hand the same task to AI and nothing comes back. The model does not show up at your desk to say, you have not considered the legacy billing integration. It produces. It fills every gap with something plausible, because producing plausible surface is the thing it most reliably does. And the hundred unconsidered things (of which, in my experience, ninety are so specific to your company that no model could currently know them, though you would have, had you done the walking) go unconsidered. Not by you, not by the machine, not by anyone. The junior colleague redistributes the thinking. The model deletes it. Some of this is a capability problem, and I want to be fair about that. Today's models really do lack the context; they were not in the meetings, they were never burned by the vendor, they do not know that the CEO has a thing about dashboards. Presumably this improves. Every company is presently racing to pipe its meetings and messages and tickets into these systems, and some of the hundred unthought things will eventually get thought by machines. If that were the whole story, work slop would be a passing annoyance of a transitional technology, and not worth an essay.

It is not the whole story, because the durable part is human. At some point in the last two years a great many people quietly moved from using AI as an instrument to using it as a ghost-worker. The instrument user remains the author of the work: the machine drafts, the human decides, and every claim in the final document has passed through a mind that could defend it. The ghost-worker user is closer to a courier. A request goes in, an artefact comes out, the artefact gets forwarded. The courier may skim it. The courier cannot defend it, because the thinking it gestures at never happened anywhere. To understand why couriering spreads, you have to feel what it is like from the inside, because from the inside it feels wonderful. You ask for the document and twenty minutes later it exists. It has heft. Scrolling through it feels like surveying a day of your own effort, and sending it off feels like completing a day of work, and by every measure visible to you, you have. For our entire working lives, polish was a reliable signal of effort (a clean thirty-page document could only exist because someone had spent days on it; there was no other way for such an object to come into being), and our sense of our own productivity was calibrated against that signal. The signal is now broken. Polish is free. The calibration has not caught up, and I am not sure it ever will, because the feeling of having done a full day's work can now be manufactured in under an hour, and it is a good feeling, and nothing about it announces its own falseness.

Corporate life produced hollow documents long before AI (anyone who has watched a consulting deck get assembled knows this). What changed is the price. Hollowness used to cost nearly as much effort as substance, which kept its volume down. Now it costs nothing, and we are drowning in the volume.

And nothing arrives to correct the feeling, which is the detail I keep returning to. The producer of slop may be the only person in the chain who never experiences it. The whole point of sending work away is that its consequences land elsewhere. The receiver wrestles with the document, works around it, quietly redoes it, curses you privately at their desk, and you learn nothing. About half the people who received slop in that survey said they afterwards saw the sender as less capable and less trustworthy than before. Almost none of them said so to the sender. The feedback exists. It accrues in a ledger you never see, until enough of it accumulates to become a reputation, and reputations do not announce their arrival either. I should concede what the researchers themselves stress. In their follow-up work, more than half of employees admitted to sending slop at least occasionally, and the authors trace this to a squeeze: leadership mandates AI use without saying what good use looks like (41 per cent of employees reported receiving exactly this kind of instructionless mandate) while workloads keep growing, and they call blaming individuals a fundamental attribution error. They are partly right, and it costs me nothing to grant it. The pressure explains why a tired person reaches for full delegation. It does not explain why full delegation feels costless. That part is explained by the broken feedback, and no reorganisation I can picture will make a receiver walk over to a colleague and say, your work has become an imposition on mine. Because that is what slop is at the receiving end, an imposition. The afternoon the producer saved did not vanish; it was transferred, with interest. The same survey put the cost of each instance at just under two hours of the receiver's time, and I would guess that is an undercount, because verifying an alien document is slower than writing your own (you are reverse-engineering intentions that were never formed). And there is a subtler charge on the same invoice. After you have been slopped a few times, speed itself turns suspicious. A fast turnaround used to impress; now it triggers inspection. I catch myself doing it, reading the quick ones twice as hard, which means the sender's saved time has become my spent time, twice over. Trust between colleagues was a technology for not having to check each other's work. It may have been the single most productive technology a company could own, and it never appeared on any balance sheet, and slop is spending it down.

Now widen the frame to the whole company and watch the ledgers. Every producer's felt productivity is up; the documents flow as never before. Every receiver's real productivity is down. Since nearly everyone is both, most people are feeling faster while being slowed, and feeling wins, because the speedup is experienced directly and the slowdown arrives ambiently, as a fog with no obvious source. If this sounds too neat, there is at least one clean measurement of the gap. METR ran a controlled experiment on experienced open-source developers and found that with AI tools they completed their tasks 19 per cent slower, while estimating afterwards that the tools had made them about 20 per cent faster. Call it a forty-point error in self-perception, in a field where reality pushes back within days. I do not hold out much hope for the self-assessments in the parts of a company where reality pushes back within quarters, if it pushes back at all.

Up to here this is an essay about waste, and waste is survivable; companies have always run at some ratio of theatre to substance. What actually alarms me is what the practice does to the people inside it.

I have written elsewhere, about machines and thinking in general, that the artefact was never the point of the work, and the argument lands with particular force on office documents. Writing the PRD was the mechanism that forced you to think the project through. The hour spent staring at the empty section titled Risks was the work. The document you handed over afterwards was closer to a receipt. Skip the writing and you skip the thinking, and this might be tolerable if some replacement thinking appeared, but mostly none does. Prompting is not that replacement (describing a task at a high level is precisely the manager's move of calling it simple, performed before the hundred things have surfaced to be described). Reviewing is not it either, or not enough of it. Reading a generated document and finding it reasonable is recognition, and recognition is a far thinner activity than construction. You can nod along to a document whose gaps you would only have discovered by writing it yourself.

The early research points the same way. When Microsoft surveyed knowledge workers about how they think with AI, the pattern was that the more a person trusted the machine, the less critical thinking they reported doing, while people confident in their own skills thought more, not less. Trust the machine, retire the mind. And the retirement compounds quietly, because the capacities that lapse first are exactly the effortful ones (holding a complicated structure in your head across days, noticing the assumption everyone in the room shares, staying with a problem past the point of comfort), these being the capacities that delegation most gratefully relieves you of.

The cruellest part is what happens when the slop finally fails. The failures come late and expensive, surfacing months downstream where nobody can trace them back to the document that seeded them. And even when the producer is standing in the wreckage, they cannot learn much from it, because learning requires a model of what went wrong, and the model is the thing they never built. The blow-up teaches them, at most, to hope for better luck. I suspect the receiver's thinking decays too, on a delay, though here I am past what anyone has measured and you should take it as suspicion. A person who spends their days sorting substance from plausible non-substance is doing triage, and triage, whatever its virtues, is not the sustained construction that keeps a mind in condition. There is also what happens to your will to work carefully once you realise your careful work is being read by people who did not do theirs. Why walk the terrain for a receiver who will skim your map, or feed it to a model for a summary? The rational response to a slop economy is to stop paying full price into it. Slop begets slop, not through malice but through the perfectly reasonable refusal to keep subsidising other people's shortcuts, and the equilibrium of that game is a company in which documents circulate freely and nobody is thinking. A Gresham's law of cognition, bad thought driving out good because it is cheaper to mint.

I argued once that delegating thought to machines is a private harm, a cost your present self quietly bills to your future one. Work slop is the same act viewed from the outside: the atrophy stays yours, and the burden goes to your colleagues. One act, two invoices, neither itemised. Your mind does not notify you when it dulls, and your co-workers do not notify you when they stop trusting what you send. Everything visible (the output, the speed, the feeling of a full day) reports that the arrangement is working. Everything that reports otherwise is delayed, or diffuse, or lands on someone else's desk.

I doubt exhortation fixes any of this; nobody audits their soul before hitting forward. But I have started doing one small thing when a suspicious document reaches me, which is to ask the sender a question about it. Any question, so long as it is specific and comes from somewhere in the middle. The people who did the thinking answer at once, usually at too much length, and the conversation that follows is often the best part of my day. The people who did not go quiet, or say they will get back to me, and more than once now, what they have got back to me with is a second document. I mind the silence less than I mind the second document. The silence, at least, knows something is missing.