The most productive people in your office right now are the ones sending documents they haven’t read. They get the credit. Somebody else gets the bill.
A founder of a small tech agency wrote on LinkedIn the other week about the requirement specs he gets from prospective clients. Twenty pages, thorough, professional. Then he asks about the content, and the client can’t explain why half of it is there.
Nobody wrote it, exactly. Nobody read it either. Now he has to.
That’s verification debt, and it’s piling up like CDO’s did in 2008.
A term from the techies
Werner Vogels, Amazon’s CTO, put the phrase on the big stage in December 2025: “Code arrives instantly, but comprehension does not.” “That’s what’s called verification debt.”
He didn’t invent it. A Canadian lecturer, Kevin Browne, wrote it up five months earlier, for all AI output, not just code. Since then it has mostly stayed in software, the one industry already built to deal with it.
I circulated it internally. We all discussed it vigorously thinking we had time to brace ourselves. instead we find that our personal context is increasingly being billed by it.
Anyhow, technical debt means you know you cut a corner and you will pay for it later. Verification debt means nobody knows whether a corner was cut. It builds up quietly, because AI output looks finished, and someone else usually pays it.
In Swedish I’ve started calling it granskningsskuld. As far as I can tell, nobody else has yet (first!!1).
It’s already everywhere
Courts keep the best numbers. US filings with AI-invented material went from 37 in 2024 to 524 in 2025 and 668 in the first seven months of 2026. Lawyers filed more than a third of this year’s. Four involved judges.
Outside the courtroom:
- At Eskilstuna tingsrätt, counsel cited four Supreme Court rulings that don’t exist, and had the fee cut.
- Deloitte refunded more than A$97,000 of a A$440,000 Australian government report with a fabricated court quote.
- Tromsø withdrew a school plan built on research that doesn’t exist. The head of education resigned.
Then there’s the enormous middle: email, memos, decks. Stanford and BetterUp call it workslop, and four in ten US desk workers got some last month. In a follow-up, more than half admitted sending it. Senders know.
The receiver pays
Generating scales with compute. Checking scales with human hours. That’s the whole story.
Anthropic measured it on itself in October. By its own count, non-technical staff would need half their working week just to read what their agents produce.
flowchart TD
accTitle: The workslop Ponzi
accDescr: Each reader is paid with the next reader's time. The founding investor generates twenty pages in thirty seconds and looks productive. Reader two has no time to read it, so they generate a reply and also look productive. Reader three summarises the summary. The last reader, a judge, an auditor or a customer, has to actually read it, and the scheme collapses.
A["<b>Founding investor</b><br/>generates 20 pages in 30 seconds"] --> R1(["Return: looks productive"])
A --> B["<b>Reader 2</b><br/>no time to read, generates a reply"]
B --> R2(["Return: also looks productive"])
B --> C["<b>Reader 3</b><br/>summarises the summary"]
C --> D["<b>Last reader</b><br/>judge, auditor, customer"]
D --> X{{"Collapse: somebody has to read it"}}
Every reader gets paid in looking productive, funded by the next reader’s time. Economists call it a “Trojan Horse externality”.
It compounds
Unchecked output gets copied, so one error ends up in ten places. Another AI summarises it and turns a doubtful claim into a confident one. Eventually it’s fed back into the models.
Errors are the easy part, because someone can at least catch them. The harder part is choices. Wording decides things: a “must” instead of a “should”, a requirement that rules out half the bidders, a definition that moves a boundary without saying so. Nobody made those choices. They came with the phrasing, and whoever passed it on unread signed them anyway. That’s the client with the twenty-page spec. Nothing in it had to be false. He just couldn’t say why it was there.
“Just review more carefully” doesn’t fix it. Experienced radiologists dropped from 82 to 45.5 per cent accuracy when an AI suggested the wrong answer. Fluent prose makes it worse: it removes the rough edges reviewers use to know where to look.
The post-tragic bit
There are two comfortable responses. The next model will fix it, or AI is ruining everything. Pick either and you get to stop thinking.
Cheap generation isn’t going away, and it shouldn’t. But when people stop writing things themselves, they stop building the understanding that came free with the writing. That was Vogels’ actual point. You don’t get it back by banning tools. You decide who’s responsible.
This has happened before. After Enron, the answer was a named person who signs off. For AI, the flood and the scandals were 2023 to 2025. Insurers and buyers are moving now. Lawmakers, as usual, are last. That makes 2027 and 2028 the years when someone has to put their name on it.
Semi-Wild-Ass Guess: 2027
Some of 2027 is already scheduled, just waiting for distribution.
The EU AI Act’s transparency rules have applied since August 2026. They care about labels, not truth, with one telling exception: AI text on matters of public interest needs no label if a human has reviewed it and takes responsibility. A named reviewer just became a compliance asset.
Sweden’s new product liability law, which brings software and AI under strict liability, is scheduled to apply from 1 January 2027. Consultants will feel it through contracts and insurance, and insurers are already writing AI out of standard liability cover.
| Forecast for 2027 | Confidence |
|---|---|
| The volume of AI-generated documents, email and code keeps rising. Nobody pulls back because of verification debt. | High |
| Swedish agencies and consultancies get questions about AI review in their insurance renewals. | Medium |
| At least one Swedish public body or supplier has to retract or correct a report built on AI-invented sources. | Medium |
| Some Swedish public buyers start asking suppliers how they use and check AI. Almost none score the answers. | Medium |
SWAG: 2028
By 2028 nobody asks whether something was made with AI. Everything is, and the labelling rules will say so. The question is who checked it, and whether they answer for it.
| Forecast for 2028 | Confidence |
|---|---|
| A “reviewer of record”, a named person who signs off, becomes a contract norm for at least one kind of non-legal work: public reports, tender answers, audit-adjacent analysis. | Low to medium |
| Knowledge services start pricing generation and verified sign-off separately. The draft is cheap. The signature isn’t. | Low |
| Firms outside tech notice they’ve stopped training juniors who can actually verify anything. | Low, but it’s the one that worries me |
That last one is the trap inside the fix. A named reviewer needs to be able to review, and reviewers used to learn by doing the work that’s now generated. It has already started where AI is used most: at Anthropic, juniors’ share of technical hires fell from 27 to 9 per cent since early 2025. A reviewer of record who can’t verify is just the old debt with a signature on it.
What to do on Monday
- Don’t forward what you haven’t read.
- When you buy reports or analysis, ask who checked them, not whether AI was used.
- Put a name on every document that leaves the building.
- Keep paying people to learn the work, not just to approve it.
The fix doesn’t have to wait for a new law or a better model. It needs someone to admit that reading is work, and to budget for it.
Somebody always has to read it. The only question is whether it’s the person who made it, or the person who got it.
I’ve written separately about what this looked like when I wrote a novel with AI as editor. Short version: half a word of rules per word of prose.