Someone asks a question in a channel. Forty seconds later a colleague drops in nine hundred words under a heading that says Key Considerations, with three bolded subheads, 75 em-dashes (ps; how the fuck do you even type out an em dash) and a closing paragraph explaining that the right approach depends on your specific requirements.
Nobody replies. What would you even say? The message has no position in it. Somebody has to read the whole thing to discover that, and now that somebody is everyone in the channel.
Your colleague thinks they helped. They spent forty seconds, and they mailed a small invoice to every person in the room.
The bill is measured
BetterUp Labs and the Stanford Social Media Lab named this in HBR last September: workslop. AI output that looks finished but carries too little substance to move the task, so the receiver does the thinking the sender skipped. Their survey of ~1,000 US desk workers:
- 40% received workslop in the past month.
- Each instance costs 1 hour 51 minutes to deal with.
- That’s about 20 minutes longer than if the sender had done the work themselves.
Sit with the third one. This isn’t a trade where you save time at my expense. The total labor goes up. You invented negative productivity and shipped it as a favor. Call it $186/month per employee, roughly $9M a year at 10,000 people, smeared across everyone’s afternoon so it never shows up on a dashboard.
And 53% admit some of what they send is probably slop. So we know. We do it anyway, because your own output never feels like slop to you. You were there for the prompt. To everyone else it’s a stranger’s exhaust.
The prompt is missing
An LLM response is half a conversation. It’s shaped by your question, your context, and the four turns where you corrected it. It’s a reply to you.
Paste it into Jira,Slack,Teams,whatever and you strip all of that. The reader gets an answer with the question amputated. They can’t see what you asked, what you ruled out, or what you’d already discarded before you hit copy. They have to reconstruct your prompt by reading the answer backwards, and they’ll get it wrong, because the information isn’t in the artifact. So they guess, then act on the guess, and now we’re four messages deep negotiating what a machine meant by “consider implementing” to a person who left the conversation.
Distillation was the whole job
Writing something short is expensive because you have to know what you think first. You read the thing, decide what matters, throw away 90%, and put your name on what’s left. That last step is the job. Everything before it is preparation.
When you paste the raw output, that work doesn’t vanish. It fans out. One person declines to spend ten minutes deciding what matters, so eight people each spend six minutes guessing what the first person meant. You didn’t remove the work. You multiplied it by the channel size and made every copy worse, because those eight have less context than you did and can’t ask the model follow-ups.
Three sentences that say I looked at this, here’s what we should do, here’s the one risk beat any volume of correct, comprehensive, beautifully formatted output. Not because brevity is a virtue. Because those three sentences contain a decision, and a decision is the only thing anyone in that channel needed from you.
The tell that you skipped it is “that’s just what the AI said.” It means nobody owns the message. You can’t interrogate it, build on it, or even disagree with it, because disagreeing requires someone holding a position. You’ve made yourself a router, and routers don’t get asked their opinion. Half the recipients in that study rated the sender less capable and less reliable. You didn’t look fast. You looked like someone who doesn’t check their work.
The test is density, not disclosure
Sean Goedecke gets to the rule I couldn’t:
It’s okay to show someone AI-generated output when the content density in that output is the same as what you’d get from a competent human.
His example: “can u add caching” expanded to three paragraphs. The model fills the gap with what it has, which is nothing, so you get caching is important to avoid unnecessary work. What you’d have written by hand: we re-fetch the user’s config every request and it costs ~20ms, this service already has Redis sitting there, and the config service is scaling badly. Same request. One version has three facts, the other has three shapes where facts go.
Density also kills the fix everyone reaches for first. A disclosed three- paragraph answer to “can u add caching” is still garbage. Labelling handles the deception and does nothing about the cost, and the cost was the problem. It cuts the other way too: if a colleague writes in their language and a model puts it into English, density is identical and the thinking is theirs. Any rule that punishes them while waving through your nine hundred words of Key Considerations has it backwards.
The rules
- Run the density check: would a competent human use this many words to say this much? If not, you’re sending packaging.
- Read it before you send it: all of it. If that’s too much reading, notice you were about to make eight people do it instead.
- Put your name on the claim: not “the AI suggests” but “I think.” If you won’t say “I think,” don’t send it. That reluctance is information.
Use the model. This argument is downstream of finding it useful. But it’s a tool for your thinking, and thinking isn’t a transferable artifact. The output is scratch paper. Nobody wants your scratch paper.
That part is still your job. It was always the only part that was.
pps; I used AI to help proofread and format/cataolog my thoughts on this, and yes, i’ve read it all…
Sean Goedecke’s “Don’t feed me AI slop” is the better half of this argument. Workslop data: BetterUp Labs and the Stanford Social Media Lab, HBR, September 2025.