In May 2025 a summer reading list ran in the Chicago Sun-Times and in at least one edition of The Philadelphia Inquirer, fifteen titles credited to well-known authors, and only five of the books existed. NPR traced it to licensed content supplied through King Features, a unit of Hearst, and to a freelance writer who said afterwards that it was partly generated by AI and that the mistake was his. The Sun-Times said the list had been neither created nor approved by its newsroom. It was wrong once, upstream, and syndication did the rest.
The scaling of errors is what happens when AI executes without a human validation step. A mistake does not stay where it was made: a plausible invented statistic, or a positioning claim that has drifted a few degrees from what the business actually is, passes review once and then gets reproduced across every asset the system generates from it. The error is not just made. It is industrialised.
One post against every asset
A person's wrong fact lives in one blog post. Someone misremembers a figure or puts a quote in the wrong mouth, a reader notices, and one correction fixes it, because the mistake would have had to be carried by hand into every other place it could go, and mostly it never was. A machine's wrong fact lives in every asset built from the same prompt, the article, the email sequence written from the article, the social posts cut from the emails, the sales deck that quotes the posts, and each copy reads as fluently as the true sentences around it.
Automation scaled mistakes long before language models did. At a finance advisory firm, a single ActiveCampaign automation sent 16,000 emails. Speed removes the natural error brake. The slow, human steps that used to catch mistakes, the proof, the second read, the colleague who asks where a number came from, are exactly the steps automation deletes, because from the outside they look like waste.
Two kinds of scaled error
The invented statistic is the visible kind. It fits the argument, it arrives in the same confident prose as the verified facts, and it gets quoted because it is quotable. A hallucination, the machine producing a fact-shaped sentence with nothing underneath it, does not look any different from the truth on the page, and it only shows itself when somebody checks it against a source.
The drifted positioning claim is quieter, because no single piece is wrong enough to stop, but across fifty pieces the business starts promising things it never decided to promise and sounding like its category instead of like itself. Inconsistency has always been what I treat as the tell of amateur work, and drift is inconsistency produced at the speed of the machine. Reading for quality catches neither kind. Checking against the truth catches both.
The cost compounds
Fixing a scaled error means finding every copy, and you usually find them after a customer or a regulator does. In January 2023 CNET issued corrections on 41 of the 77 stories it had published using an AI tool, and the internal review that found them came after outside reporting had already pulled errors out of one of those stories in public.
The customer version is smaller and costs more per word. In 2022 Air Canada's website chatbot told a passenger he could book a full fare to his grandmother's funeral and claim the bereavement discount afterwards, which was not the airline's policy, and when he applied the airline argued the chatbot was responsible for its own actions. The British Columbia Civil Resolution Tribunal disagreed and ordered Air Canada to pay $812.02, on the reasoning that a company is responsible for all the information on its website, whether it comes from a static page or a chatbot. One generated answer became a policy the business never wrote and still owned.
The trust cost runs the same way. My view is that any movement away from the truth eventually forces the customer to explain your words to themselves, and the only explanations left are that you were ignorant, incompetent, inadvertently selfish or malicious. None of those is repaired by publishing faster. On The Trust Algorithm a scaled fabrication is Authority Theatre produced by accident, and it collapses the same way, because the credibility it borrowed was never there.
Q3 is the stage you cannot hand over
Q3 is human validation, the third of the four steps in the loop that runs inside the 4-Quadrant Framework, where a person checks what the machine produced against the truth before anything is deployed. It is the one stage that cannot be delegated. Skip it to save time and you save the time once, then pay for it on every asset the error reaches.
The truth being checked has to live somewhere other than the validator's memory, and that place is the System Seed, the single source of what the business knows, how it sounds and which facts are verified. The seed feeds the machine in Q2 and it is the benchmark in Q3, so the validator is not asking whether a sentence reads well. They are asking whether this number is in the seed, whether this promise is one the business has made, and whether this claim would pass what Traffic + Offer calls the Bullshit Police, applied to machine output: would this statistic survive a fact-check?
When the answer is no, the fix goes to the source rather than the sentence. Correct the seed and the correction travels to every asset built from it, which is the scaling mechanism running in reverse, and it is the job Cascading Updates exists to do.
The seed itself is checked against receipts rather than against memory. A fact that cannot be traced to one does not go into the seed, and a page that needs a fact the seed does not hold gets a question sent back instead of a sentence, which is the validator's rule moved to the one place where a single correction reaches every copy.
The check worth running
Pick one number your operation has published in the last quarter and find every place it now appears, the page, the emails, the posts, the deck, the proposal, then trace it back to a source. If you cannot find every copy, the operation is producing faster than it can correct. If you cannot find the source, none of the copies was ever true. The diagnostic runs the same kind of test against each of the four quadrants.