Unchecked AI output doesn’t remove work, it moves the work onto whoever reads it. Call them AI slop grenades. The fix is one sentence in your working agreement. Whoever hits send owns every word.
Forty hours a month. That’s what a 50-person firm could lose to AI slop grenades if BetterUp’s survey numbers hold at your place, and no line in the budget shows it. The term comes from Shopify CEO Tobias Lütke, who used it in a September interview to describe AI output tossed at a colleague unchecked. The forty hours is our arithmetic, below. The habit is real, and we think it’s the most expensive one a small team can pick up this year.
Key points
- What are AI slop grenades? AI-generated work passed to a colleague without the sender checking it.
- What does it cost a 50-person firm? Roughly 40 hours a month, on one set of assumptions.
- Why won’t a better model or a ban fix it? Because neither touches the cause. A better model still produces output nobody has read, and a ban gets routed around, so you lose the real gains with the junk.
- What rule stops it? One sentence in the working agreement: if you send it, you’ve read it, and you answer for every claim in it.
What are AI slop grenades?
AI-generated work passed to a colleague without the sender checking it. Lütke, speaking on The Knowledge Project, said, “We call those ‘slop grenades’ that people toss at each other.” The sender saves minutes. The receiver has to read it, fix it, or feed it to another model to shrink it.
He asked why we invented “decompression and recompression,” and called the result terrible.
Checking someone else’s work is slower than checking your own, because the receiver can’t tell which parts are wrong without reading all of it. That’s the transfer. Payroll never sees it.
What does it cost a 50-person firm?
Roughly 40 hours a month, on one set of assumptions. BetterUp’s survey found 40% of US desk workers got workslop in a month, and each case took about two hours to fix. Apply that to 50 people and 20 of them lose a working week of one person’s time between them.
BetterUp Labs and the Stanford Social Media Lab surveyed 1,150 full-time US desk workers in August and September 2025 and named the pattern workslop. 40% said they’d received some in the past month, and each case took about 1 hour 56 minutes to resolve. 53% admitted sending it at least some of the time. BetterUp puts the cost at $186 per affected person per month, based on average salary.
Twenty pieces at about two hours each is 40 hours spent undoing other people’s shortcuts, and about $3,720 using BetterUp’s per-person figure (in US dollars). That’s our arithmetic on a year-old US survey. Every input is self-reported, so your number may be a third of that or double. Axios covered the original study when it came out, if you want the source chain.
Why won’t a better model or a ban fix it?
Because neither touches the cause. A better model still produces output nobody has read, and a ban gets routed around, so you lose the real gains with the junk. Accountability is the missing piece. BetterUp’s own page says 41% of employees report leadership encouraged AI use without explaining how or why. Nobody was told who answers for the result.
Lütke’s line points the right way. “Humans take responsibility,” he said, and added that machines can help us take more responsibility because they inform us better. Shopify isn’t an AI sceptic. He says its internal agent, River, handles a large share of production code pull requests, possibly as much as half. The rule is about who answers for output, not whether to use it.
What rule stops it?
One sentence in the working agreement: if you send it, you’ve read it, and you answer for every claim in it. The sender owns the result whether a person or a model wrote the first draft. Then back the sentence with habits that make it cheap to follow.
Anything over a screen long carries a two-line summary written by the sender, not the model. Numbers, names and dates get a source or a “checked by” mark before they leave your desk. And receivers may bounce unreviewed work back with no hard feelings, which only works if a team lead says so out loud first. Without that permission the rule dies in the first week, because nobody wants to be the one who returns the boss’s memo.
Our piece on the hidden time tax of AI covers the wider version of this problem. If your team hasn’t settled which assistant it leans on, the big-three comparison is the place to start. The rule also fits next to a written rule on who wrote what, and it works best where leaders already use the tools themselves, a gap we looked at in our column on leader adoption. Cleanup hours belong in the hidden time tax on AI productivity, so count them with the rest.
Isn’t this just bad email with a new name?
Partly. Bad memos predate chatbots, and BetterUp’s term is a label for an old failure. What changed is polish and volume. An AI draft looks finished, so the reader trusts it longer before spotting the gap. A sloppy human email announces itself, and a slick wrong one doesn’t, which is why old etiquette won’t cover it.
Our evidence is thin in one place. Nobody has measured workslop inside a firm like yours, and Lütke’s remarks come from one very large, very AI-heavy company. If a month of counting rework in your own team shows no rise since AI use took off, the rule is cheap insurance and not a fix. Health data needs its own rule, because ChatGPT Health privacy promises don’t cover your staff policy.
Count it. Ask each team lead to log every item bounced back for a month. If the total is nowhere near 40 hours, you’ve spent one sentence in a working agreement.
Frequently asked questions
What is a slop grenade?
It’s Shopify CEO Tobias Lütke’s phrase for AI-generated material that someone sends to a colleague without checking it, so the colleague has to review, fix or summarize it.
What is workslop?
It’s BetterUp Labs and the Stanford Social Media Lab’s term for AI-generated work that looks finished but lacks the substance to be usable. In their survey, 40% of US desk workers said they got some in a month.
How can a small business stop AI slop at work?
Make the sender responsible for everything they send. Require a human-written summary on long items and a source on every number, and let receivers return unchecked work.
How much time does workslop cost?
In BetterUp’s survey, each case took about 1 hour 56 minutes to resolve, and BetterUp estimates $186 per affected employee per month. Those are self-reported US figures, so treat them as a direction.
The cost example is our own arithmetic on a published survey, not a measurement of any firm. Written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
