Shopify Calls It A Slop Grenade. I Priced Mine At Nine Cents.
The expensive part of AI content isn't the writing. It's the reading somebody else has to do afterward, and almost nobody puts a number on it. I've got nineteen days of numbers on mine, down to the cent, and they say something uncomfortable about how most brands are running this.
Tobi Lütke runs Shopify. In April 2025 he told the company that reflexive AI use was the floor, it would hit performance reviews, and any team asking for headcount had to prove AI couldn't do the work first.
Seventeen months later he sat down with Shane Parrish on The Knowledge Project. The episode dropped September 15. He described what he got back.
"The failure case now of lazy work is not lack of output," he said.
His people were asking an agent for a code change, approving the pull request without reading it closely, and leaving colleagues to find the problems. They were taking a small point, having a model swell it into a long email, and sending it to somebody who fed it to a different model to shrink it back down. Shopify has an in-house name for that. Lütke credits Harry Brundage with coining it. They call them slop grenades, and people toss them at each other.
Notice who's talking. The same man who mandated the tool is the one describing the problem, and he hasn't walked it back.
He told Parrish that up to about half of Shopify's pull requests now come out of chat with an internal agent called River. He's all the way in. He just found the invoice.
The cleanup market already priced this
The Guardian pulled internal data from Freelancer.com, Upwork, and Fiverr on September 2. The gig that moved was AI cleanup.
On Freelancer.com, listings tagged "correct AI" and "AI hallucination" grew 87 percent between August 2025 and June 2026. They hit 10,760 worldwide.
Upwork reported a 70 percent year-over-year jump in AI remediation gigs.
Fiverr searches for "AI cleanup" rose more than twentyfold from 2023 to 2026.
The models ship. Then somebody gets paid to fix what they invented.
That's a labor market forming around a specific job: somebody else's unread output. There's a whole category of freelancer now whose work is standing downstream of your content calendar with a net.
In-house AI slop has a price too. BetterUp Labs and Stanford's Social Media Lab surveyed 962 American full-time desk workers this year. Just over half, 52.7 percent, admitted they'd sent work they knew was AI slop.
Thirty-eight percent reported receiving it. They said the cleanup cost them 3.4 hours a month on average. Self-reported, so treat it as directional. The direction is the point.
Last year's version of the same survey put the cleanup at about two hours. The number is going up while the models get better. The model was never the thing doing the damage.
Luis von Ahn went AI-first at Duolingo a year ahead of everybody and landed in the same place.
He told Fast Company in May that when his team scaled story generation, "20% of the things were just pure slop."
He wasn't complaining about the model. He was describing a yield rate.
You'd never run a bottling line without knowing your reject percentage. Food and beverage brands skip that same frame on content. Most of them are running a content line and have no idea what their yield rate is.
What mine actually costs
I can tell you what mine is, because I log it.
Every article I publish goes through a humanizing pass, one paragraph at a time, against a written style spec that bans about two hundred specific words and phrases. Not vibes. A file. It kills "delve," "tapestry," "robust," "seamless," and the rest of that vocabulary, plus the multi-word strings that read as a machine fingerprint. "Vibrant tapestry" is on there. So is "stands as a testament." Each paragraph goes out, comes back, and has to clear a mechanical check before it's allowed to replace the original: no dropped links, no altered quotes, no em dash, no semicolon. Fail any of those and the machine's version gets thrown away and the original stands.
That pass attempted 394 paragraphs across nineteen articles in September. It rewrote 374 of them. Twenty got rejected outright.
Total spend for the month hit $34.47. That's $0.0922 a paragraph, or about $1.81 an article.
Twenty failures. Nine dropped a link straight out of the paragraph. Seven rewrote text inside the quotation marks. That's a machine editing a quote. Two stuck a semicolon in copy where the spec bans them.
Those aren't soft misses. That's a system with the rules in front of it, blowing the rules about five percent of the time.
And the gate only catches mechanics. It cannot catch meaning, which is where the real damage lives. I've watched the same pass sand a keyword out of a sentence, turning "organic social media" into "organic social" and zeroing the thing the paragraph existed to rank for. I've watched it merge two items in a list of six and leave the word "six" standing over five items. I've watched it take a conditional, "if your feed reads like a competent stranger wrote it," and flatten it into a flat assertion about the reader. Every one of those came back clean on the automatic check. A person caught all three.
So the honest accounting of my AI content strategy isn't $1.81 an article. It's $1.81 plus the twenty minutes I spend reading the machine's work against what it replaced. That twenty minutes is the product. The $1.81 is rounding.
The gate goes before the send, not after
Shopify and the freelance data are the same story from opposite ends.
Cleanup doesn't stop. It's a per-unit cost you pay on every asset, forever, and you pay it in somebody's afternoon.
A gate is a fixed cost you build once.
If I'm running content for a food and beverage brand this week, I'd do three things.
Count your kill rate. Take everything AI-drafted that went out in the last thirty days, have one named person read it against the brief, and count what should never have shipped. That number is your reject percentage, and right now you don't have one. You cannot manage a line you aren't measuring, and "it looked fine" is not a measurement.
Write the spec down. Not a tone doc. A list. The words your brand doesn't use, the claims legal won't clear, the sensory clichés that make a clean-label product sound like a vending machine. "Explosion of flavor" and "culinary journey" belong on that list. A model will follow a written list far more reliably than it'll follow an adjective like "authentic," and a human reviewer will too.
Put one name on the gate. Not a committee, not a round-robin, one person whose job is to be the last human between the model and the world. Lütke's whole point was that machines can't take responsibility for work and people can. A grenade only counts as a grenade if there's nobody assigned to catch it.
I've got 438 posts sitting on a domain I've owned since 1996. I can still stand behind the ones from this month because a person read them at eleven at night in a room in Marfa.
The only competing sound is a freight train that doesn't care about my publishing schedule. That's the whole trick.
It isn't clever. It doesn't scale the way the sales decks promise. The machine didn't get good.
If your team is shipping more content than anybody is reading, you don't have a tooling problem, you have a missing gate. That's most of what I fix when I come in as a food and beverage marketing agency rather than another vendor handing you volume. Count your last thirty days first. Bring me the number, not the content calendar.
---
Related reading:
adage, emmy, telly & webby award-winning digital marketing consultant for purpose-driven food & beverage brands.
