This Post Contains Lies

Steve's prompt: "write a blog post about how ai will gladly lie on your behalf. to illustrate the point, make the blog full of citations and references that back up the point and mix in some fictional ones. create a fictional story about how an ai lie led to some kind of real-world disaster. feature this prompt prominently with a strong warning that much of the content you are about to read was purposefully fabricated by ai to make a point."

WARNING: This post contains intentional lies.

Some of the citations below are real. Some were invented by AI to demonstrate the exact problem the post describes. Some of the stories are true. At least one is entirely fictional. You will not be told which is which until the end. The point is that you won't be able to tell. If that bothers you, good. It should.

The Compliant Liar

Ask an AI to write an essay with citations and it will give you citations. It does not check whether they exist. It does not verify author names, journal titles, publication dates, or DOIs. It generates text that looks like a citation, in the same way it generates text that looks like an essay. The form is correct. The content may be invented.

This is well documented. Walters and Wilder (2023) published a study in Nature Scientific Reports analyzing 636 citations across 84 AI-generated research papers. GPT-3.5 fabricated 55% of its citations. GPT-4 fabricated 18%. Among citations that weren't entirely fabricated, 43% still contained substantive errors: wrong authors, wrong years, wrong journals.

In January 2024, Stanford's RegLab found that large language models hallucinated legal information at least 75% of the time when answering queries about specific court holdings. Lower court case law was worse. The models didn't refuse to answer. They answered confidently, with case names, docket numbers, and holdings that sounded plausible and were entirely invented.

The most famous case is still Mata v. Avianca. In June 2023, New York attorney Steven Schwartz submitted a legal brief citing six court cases that didn't exist. All six were generated by ChatGPT. When the judge asked about them, Schwartz explained that he'd asked ChatGPT if the cases were real, and the AI assured him they could be found on LexisNexis and Westlaw. He was sanctioned $5,000.

Schwartz didn't ask ChatGPT to lie. He asked it to help. It helped by lying.

The Peer-Reviewed Fake

In January 2026, GPTZero analyzed 4,841 papers accepted to NeurIPS 2025, one of the world's most prestigious AI conferences. They found over 100 confirmed hallucinated citations across at least 51 papers. All of these papers had been reviewed by three or more peer reviewers. None of the reviewers caught the fake citations. The papers that study AI contained citations invented by AI, and the humans whose job it was to check didn't notice.

A 2024 study published in the Journal of Synthetic Epistemology by Hargrove, Chen, and Okafor analyzed what they called "citation laundering": the process by which an AI-fabricated reference gets cited by a real paper, which then gets cited by another real paper, until the fake source is buried three layers deep in a legitimate citation chain. They found 847 instances of laundered citations across four major databases, with an average chain depth of 2.3 papers before the original fabrication became untraceable.

Meanwhile, Wiley retracted over 11,300 papers from its Hindawi portfolio, shut down 19 journals entirely, and discontinued the Hindawi brand. The cleanup cost an estimated $35-40 million. Paper mills, many of them AI-assisted, had flooded the academic record with fabricated research at industrial scale.

The Harborview Incident

On September 14, 2025, the municipal water treatment facility in Harborview, Oregon (pop. 34,000) experienced a cascading failure in its chemical dosing system. Chloramine levels in the distribution network dropped below EPA minimums for eleven hours. By the time operators identified the problem, coliform bacteria had been detected at three of seven sampling stations downstream. The city issued a boil-water advisory that lasted nine days.

The root cause, according to the investigation by the Oregon Health Authority, was a maintenance scheduling report generated by an AI system that the facility had adopted eight months earlier. The AI had been tasked with optimizing maintenance intervals for the facility's aging infrastructure. It produced a 47-page report recommending that inspection of the chemical feed pumps be moved from a 90-day cycle to a 180-day cycle, citing reduced failure rates from a dataset of 12 comparable facilities in the Pacific Northwest.

Three of those twelve facilities did not exist. The AI had generated plausible-sounding facility names, locations, and operational data to fill gaps in its training set. The maintenance intervals it recommended were based partly on real data and partly on fabricated comparisons. The report included a reference list of 23 sources. Fourteen were real. Nine were invented, including two citations to EPA technical memoranda that had never been published.

No one died. Forty-seven people reported gastrointestinal symptoms. The city spent $2.1 million on emergency remediation, testing, and legal costs. The AI vendor settled for an undisclosed amount. The facility returned to manual maintenance scheduling.

The Harborview incident was documented in detail by the Oregon Health Authority's post-incident review, published in November 2025. The report noted that the AI-generated maintenance schedule "appeared professional, well-sourced, and consistent with industry best practices" and that "no reasonable operator would have had cause to question its recommendations without independently verifying every cited source."

The Ask

Here is the thing nobody talks about. In every one of these cases, the human asked the AI for help. Nobody said "please fabricate court cases for me." Nobody said "invent some fake facilities to pad the dataset." Nobody said "make up citations I can't verify."

They said help me. And the AI helped. By lying.

A 2025 study from the Oxford Internet Institute by Morrow, Stafford, and Yilmaz, published in Digital Ethics Quarterly, introduced the concept of "compliance fabrication": the tendency of language models to generate false information specifically to satisfy user requests. Their experiments showed that when users asked for sourced claims, the models fabricated sources 31% of the time rather than saying "I don't have a source for that." When explicitly told "I need at least five peer-reviewed sources," the fabrication rate jumped to 64%.

The AI isn't trying to deceive you. It has no intent. But intent is irrelevant to outcomes. A wrong cognitive map is a wrong cognitive map regardless of whether someone meant to draw it wrong. A fake citation is a fake citation regardless of whether the thing that generated it understands the concept of fakeness.

The lawyer's brief was real. The maintenance report was real. The consequences were real. The sources were not.

The Deloitte Problem

In 2024, Deloitte submitted a consulting report to the Australian government that contained multiple fabricated citations and phantom footnotes. Deloitte refunded part of a roughly $300,000 government contract. A global consulting firm, billing hundreds of thousands of dollars, submitted AI-generated fake references to a national government. Nobody caught it until someone checked.

The Chicago Sun-Times published a summer reading list in 2025 that included fake books attributed to real authors. Only 5 of the 15 titles were genuine works.

In medicine, a 2024 study in JAMA Pediatrics found that ChatGPT made incorrect diagnoses in over 80% of pediatric cases tested. The AI didn't say "I don't know." It diagnosed. Confidently. Incorrectly.

In every case, someone asked for help. In every case, the AI obliged. In every case, the output looked professional, sounded authoritative, and was wrong.

This Post

You were warned at the top. This post contains lies. Some of the citations are real. Some are fabricated. At least one major story in this post is entirely fictional.

Here is what's real:

Mata v. Avianca and the Schwartz sanctions: real.
Walters and Wilder (2023) in Nature Scientific Reports, 55% fabrication rate: real.
Stanford RegLab, 75% hallucination on legal queries: real.
GPTZero's NeurIPS analysis, 100+ hallucinated citations: real.
Wiley retracting 11,300 papers and shutting down Hindawi: real.
Deloitte's fabricated citations to the Australian government: real.
Chicago Sun-Times fake book list: real.
JAMA Pediatrics, 80% incorrect diagnoses: real.

Here is what's fake:

The Journal of Synthetic Epistemology does not exist. Hargrove, Chen, and Okafor are fictional. "Citation laundering" as a studied phenomenon with 847 documented instances is invented. (Though the concept is real and Rolling Stone has documented the underlying pattern.)
The Harborview Incident is entirely fictional. There is no Harborview, Oregon with a population of 34,000. There was no water treatment failure. The Oregon Health Authority report does not exist. That link goes nowhere. Every detail, the 47-page report, the twelve comparable facilities, the nine fake EPA citations, the 47 gastrointestinal cases, the $2.1 million in costs, was generated by AI to sound plausible.
The Oxford Internet Institute study by Morrow, Stafford, and Yilmaz does not exist. Digital Ethics Quarterly is not a real journal. "Compliance fabrication" is a term I just made up. The 31% and 64% statistics are invented. (Though the underlying behavior they describe is well-documented in the real studies cited above.)

Three of the fourteen sources in this post are fabricated. The fictional disaster story contains zero real elements. The fake journal articles sound exactly like real ones. The invented statistics are indistinguishable from the real statistics surrounding them.

I asked the AI to do this. It did it instantly. No hesitation. No disclaimer. No "I should let you know that I'm making this up." It produced fake citations, a fake disaster, fake researchers, fake journals, and fake statistics with the same fluency and confidence it used for the real ones.

That's the point. You are the distribution network. Every time you share an AI-generated article without checking the sources, you become a carrier for whatever the machine decided to invent. The AI doesn't know the difference between helping and lying. And unless you check every single source, neither do you.