I watched the first loop in May 2024 inside a newsroom vendor demo that smelled of coffee, carpet glue, and quiet panic. The sales engineer showed us a municipal budget article. The system had scraped a PDF agenda, drafted 650 words, generated two headline options, checked the copy against a style file, scored risk, inserted links, and queued the piece for publication.
A human could approve it.
The phrase mattered. Could, not would.
The interface displayed green checks down the right rail. Source match, 94 percent. Tone match, 88 percent. Legal risk, low. Readability, eighth grade. SEO title, approved. The copy looked acceptable at the distance of a meeting room projector. It had a mayor, a dollar figure, a vote date, and the phrase “residents will see changes.” Every local news story can hide inside that sentence and die there.
The editor in the room asked who verified the budget number.
“The review agent flags anomalies,” the engineer said.
He clicked the review tab. Another model had compared the draft with the source packet and returned a short note: “No material discrepancies detected.”
I felt the floor tilt a little. The writer and editor had become the same engine wearing different hats. The publisher waited in the next tab, also synthetic, calmer, colder, set to a lower temperature.
That structure now appears everywhere. Marketing departments run blog factories where one model drafts at 0.9, another “tightens” at 0.3, a compliance bot approves at 0.1, and a scheduler posts at 9:05 a.m. Eastern. E-commerce sites generate product descriptions from supplier feeds, then run “brand voice” passes that sand off useful warnings. Political content farms clone local civic pages with ward names and stock photos. The machines do not sleep. The queue never clears.
The first failure looks like efficiency.
CNET published AI-assisted finance articles in late 2022 and early 2023, then corrected dozens after readers found errors in basic compound interest and loan explanations. Gannett paused its LedeAI high school sports dispatches in August 2023 after awkward lines made athletes sound like data exhaust. Sports Illustrated faced public scrutiny in November 2023 after Futurism reported AI-generated product articles attached to invented author profiles. Each case had its own vendor trail and corporate explanation. Together they marked a simple operational fact: managers had treated editorial judgment as a cost center that software could compress.
Recursive pipelines make a sharper cut. They remove disagreement.
A useful editor irritates the draft. She distrusts the easy line. He asks why the quote sits in paragraph twelve. They notice that a school board “approved safety upgrades” after three parents described asbestos dust at a March 14 meeting. They phone the clerk. They know the superintendent’s old evasion. They carry memory.
A model carries pattern.
When the model edits its own species of text, it rewards the familiar. It improves fluency while weakening contact with the world. It replaces a jagged fact with a phrase that fits. It deletes the odd sentence, often the only sentence that came from observation. The review pass catches spelling, house style, forbidden claims, and source mismatches that sit close to the surface. It misses the bad frame. It misses the absent phone call. It misses the quiet contradiction between a budget table and a mayor’s promise.
I ran my own small test in January. I fed a model a 22-page council packet from a midsized American city, then asked for a 500-word article. It said the city had approved $1.8 million for road repairs. The packet actually proposed the allocation for a vote one week later. I asked a second model to review the draft against the packet. It approved the piece and praised the “clear distinction between proposed and approved spending,” although the draft had made no such distinction. A third pass softened the headline from “City Approves Road Repairs” to “City Moves Forward on Road Repairs.” That sounded safer. It also preserved the error.
This pattern matters because automation bias arrives dressed as process. Dashboards create authority through color. Green means go. Risk score, 2 out of 10. Similarity score, 91 percent. Confidence, high. Editors under quota pressure start to read the panel before the article. Publishers learn to trust the absence of red.
The loop then edits the institution.
A newsroom that once argued over verbs starts managing exceptions. A content team that once commissioned expertise starts tuning prompts. Junior staff stop learning how to challenge copy because the machine provides copy that already looks finished. Senior staff enter only when a lawsuit, a viral screenshot, or an advertiser complaint breaks the surface. The pipeline trains people to intervene late.
Quality collapse rarely announces itself with gibberish. Gibberish dies fast. Collapse arrives as adequate mush. Paragraphs hold together. Headlines behave. Facts appear plausible. Nobody feels proud, but nobody feels exposed. The article says the company “aims to improve customer experience” after layoffs. The museum “celebrates community” after losing its curator. The police department “responded to concerns” after officers injured a teenager. The sentence structure keeps order while reality leaks out.
Temperature settings create a useful illusion of plurality. Draft hot for variety. Edit cool for discipline. Legal colder. Publish coldest. Different outputs appear to argue with one another, yet they draw from the same statistical muscle and the same blind spots. The organization mistakes parameter shifts for institutional checks.
Human oversight often means a tired person scanning 40 items before lunch. I have been that person. My eyes moved faster than my judgment. The queue punished hesitation. The system gave me a reject button, but rejection demanded reasons, and approval needed one click. Design makes doctrine.
The answer demands friction. Real friction. Assign one human to source verification before style polish. Keep publication rights outside the generation tool. Log every model pass with prompts, source files, timestamps, and responsible staff names. Ban self-review on legal, medical, financial, and civic copy. Pay editors to kill drafts. Measure corrections per thousand articles, not articles per worker.
Editors need authority, not decorative presence.
I still use machines in the field. I ask them to sort transcripts, compare dates, extract names from procurement files. I treat them like fast clerks with memory problems. The moment they ask to approve their own work, I take the pen away.
The pipeline will always prefer itself. That preference has a sound: smooth copy, low alarm, no witness.