Artificial intelligence was sold to every institution it touched as a way to do more with fewer people. The same week the Vatican's first AI encyclical urged governments to "slow down," three unrelated stories showed where the technology's danger actually lives — and none of them involved a machine that was too smart. They involved machines that were confidently wrong, at a volume no human institution was built to absorb.
A study found the rate of fabricated references in biomedical papers has grown more than twelvefold since 2023 — by early 2026, one in 277 papers cited at least one source that does not exist. A surge of AI-drafted lawsuits from self-represented litigants is swamping court dockets, consuming resources faster than clerks can process them. And as AI coding tools like Mythos produce what one report called a "bug-pocalypse," cybersecurity job postings rose 11% and executive search firms began turning away clients.
Three institutions — science, the courts, software — facing the same problem from three directions. AI was sold to each of them as a way to do more with fewer people. Each is now spending more people to clean up after it.
What the Tool Was For
The promise was always subtraction. AI would draft the brief, so you needed fewer paralegals. AI would summarize the literature, so the researcher could skip the library. AI would write the boilerplate, so the engineer could ship faster. Every pitch was a labor pitch: the machine takes the tedious work, the human is freed for judgment.
That promise assumed one thing — that the machine's output was trustworthy enough to use without checking. The entire labor savings depended on the verification step being cheap, or better, unnecessary. You don't save a paralegal's hours if a lawyer has to read every AI citation to confirm the case is real.
This is the load-bearing assumption, and it is the one that failed. Large language models do not retrieve facts; they generate plausible text. A fabricated citation looks exactly like a real one. A hallucinated case name has the right format, the right reporter abbreviation, the right year. The output is indistinguishable from correct output until someone checks — and checking is precisely the labor the tool was supposed to eliminate.
The Web Was the Warning
This was visible three years ago, in a lower-stakes system. In May 2023, the Washington Post reported that AI text generators were quietly authoring more of the internet. A month later, The Verge documented how AI was "killing the web" — spam sites, mass-produced articles, AI-generated junk on Etsy, Reddit, and Wikipedia that "exhausts moderators." The word was already there: exhaust. The volume of plausible-looking machine output had begun to outrun the humans whose job was to vet it.
That same month, a New York lawyer faced sanctions for filing a brief full of "bogus judicial decisions, with bogus quotes and bogus internal citations" generated by ChatGPT. It read as a curiosity in 2023 — one careless attorney, one viral cautionary tale. It was actually the first court appearance of a structural problem. The lawyer's mistake was not using AI. It was trusting the output without checking it — doing exactly what the tool's value proposition told him he could do.
The web absorbed the first wave because the web has no quorum, no docket, no standard of proof. A polluted comment section degrades gracefully. The institutions that came next do not.
The Institutions Don't Degrade Gracefully
By 2025, the same dynamic had moved into systems with consequences. In May, Anthropic — a company whose entire brand is AI safety — apologized to a court for one of its own expert witnesses citing a fabricated source generated by Claude. In June, the UK High Court warned lawyers they could face criminal charges for submitting AI-hallucinated case law. By October, judges across Latin America were struggling to rule on cases built on AI-generated fabrications. By November, a group of lawyers had documented 533 separate cases of AI misuse in legal filings.
The courts could not respond by degrading gracefully, because a court's function is to be a quorum — a place where claims are tested against a standard of proof. You cannot have a court that processes unverified claims; that is just a complaint box. So the verification burden did not disappear. It landed on judges and clerks, one fabricated citation at a time.
Science had the same structure and the same outcome. The peer-review system was designed for a world where producing a plausible-looking paper with a full bibliography took months of human effort — the cost of fabrication was itself a filter. Lower that cost to near zero and the filter stops filtering. A 12x increase in fabricated references since 2023 is not a story about dishonest scientists. It is a story about a verification system calibrated for a fabrication cost that no longer exists.
The Reversal, in a Job Posting
Which brings us to the bug-pocalypse. AI coding assistants were the clearest labor pitch of all: write more software with fewer engineers. And they did write more software — including, it turns out, more bugs and more vulnerabilities than the humans who shipped them could catch. The result is not fewer jobs. It is a documented 11% rise in cybersecurity job postings, recruiters turning away Fortune 100 clients, a hiring market for the specific human skill of finding what the machine got wrong.
The technology sold as labor subtraction created a labor category. The job is checking the machine.
This is the reversal stated cleanly. A tool adopted to reduce the number of people needed to do a task has increased the number of people needed to verify the task was done correctly. The savings on the generation side are real; the new cost on the verification side is also real, and for high-stakes institutions it is larger. The net is not subtraction. It is a transfer — from the work of doing to the work of checking — plus a tax for the volume in between.
You can see the same transfer in the week's stranger stories. Executives are building AI "digital twins" of themselves to handle correspondence and presentations — which means someone now has to verify the twin did not say something the executive never would. Every act of automated generation creates a downstream act of human verification. The only question is who pays for it, and whether they can.
What the Pope Saw
That the encyclical's "unpredictability" passage reportedly bears the influence of Anthropic — whose own model once forced the company to apologize to a court for a fabricated citation — has implications of its own. The point here is narrower. The people closest to these systems are not warning about a distant superintelligence. They are describing a property the technology has right now: it produces output that is confident, fluent, and unverified, and it does so faster than any human process can keep up.
"Slow down," the Pope wrote. Whatever one makes of a papal encyclical as technology policy, the phrase identifies the actual mechanism. The danger in the day's data is not that AI is too capable. It is that AI's generation speed has outrun society's verification speed, and the gap between them is filling with fabricated citations, swamped dockets, and unfound bugs. Slowing generation is one way to close the gap. The other — building verification capacity that scales with the output — is the one no institution has yet figured out how to afford.
The Tax No One Priced
For three years, the cost of AI has been measured on the generation side — tokens, GPUs, capex, the price per query. Those numbers are real and enormous. But the day's stories point to a cost that has never appeared on any balance sheet: the verification tax. Every plausible-looking output that an institution cannot trust without checking imposes a cost on whoever has to check it — the clerk, the reviewer, the security engineer, the judge.
The web paid this tax quietly, in moderator burnout, because the web could afford to be wrong. The courts, the journals, and the codebases cannot. They were built on the assumption that producing a credible-looking claim was expensive enough to be self-limiting. AI removed that limit, and the institutions are now discovering what they cost to run without it.
In 2023, an AI-written legal brief with fake citations was a single lawyer's embarrassment. In 2026, it is one in 277 medical papers, 533 documented legal filings, an 11% hiring surge, and a Pope. The tool was supposed to do the work. It turns out the work was never the generation. It was the checking — and the machine made more of that, not less.