A YouTube video titled "I'm pretty sure this book is AI slop" got 1.2 million views in January 2026. Nobody commissioned a study. Nobody ran a detection scan. Readers just noticed, then said so, then kept saying so until the thread had hundreds of comments. By March, Hachette had pulled Shy Girl from US shelves and discontinued its UK edition. The author denied personally using AI; she said a contractor did. The publisher cited its commitment to "original creative expression." Nobody mentioned a policy.
That is how the new disclosure norm announced itself.
The more formal version exists too, and it tells a different story. In April 2024, the BMJ Group made AI disclosure mandatory for all submissions across 49 of its biomedical journals. Required. In the form. No exceptions. Researchers checked a box. By November 2024, 25,114 submissions had come through. Exactly 5.7% disclosed AI use. Separate surveys in the same period found that somewhere between 50 and 76 percent of researchers use AI in their work. This is not a rounding error. It is a near-complete failure of a mandatory rule, playing out quietly across tens of thousands of submissions, with no enforcement mechanism and no consequences. The arxiv preprint servers hit the same wall from the other direction: banning slop after the fact rather than requiring disclosure up front.
The workplace data is stranger, because there the motive is on the surface. Ivanti surveyed more than 6,000 office workers in 2025 and found that 32% of employees who use generative AI at work keep it from their employer. The largest group, 36%, said they wanted a competitive advantage over colleagues. Twenty-seven percent described what the report called "AI-fueled imposter syndrome": they do not want anyone questioning whether the work is really theirs. This is not a compliance problem. It is a social one. The behavior it most resembles is not fraud; it is the thing people did with spell-check in the nineties, quietly, until nobody cared anymore.
Authors are watching from a different angle. The Authors Guild launched a "Human Authored" certification in January 2025: $10 per title, honor system, no scan required. A seal for your book jacket that says a human wrote the words inside. The UK's Society of Authors followed with its own logo in early 2026. These programs did not come from regulators. Nobody required them. The demand came from writers who wanted a way to draw a line, which is itself a data point: there is now a market for the claim "this one is mine" that did not need to exist before. That the certification runs on the honor system is not a design flaw. It is an accurate description of where the whole conversation is right now. The book that started the loudest recent argument about AI and nonfiction also relied on the honor system.
Journalists are in the same holding pattern. About 70% have used AI to assist with work, according to a 2024 Associated Press survey of newsrooms. Fewer than 20% work at organizations with formal policies. Researchers at Trusting News who study how audiences receive AI disclosures have found a structural problem: when a label is identical every time, it becomes noise. Readers stop processing it. The solution they suggest is to weave the disclosure into the story itself, which means the disclosure needs to actually mean something.
What is forming here is not a norm about disclosure. It is a norm about suspicion. The Hachette case was not resolved by a form or a policy. It was resolved by a crowd that had developed pattern recognition, then acted on it. When people operate without that scrutiny, the data shows most of them stay quiet. The informal etiquette that is emerging is not "say what you used." It is closer to "don't be the one who gets caught not saying it."
Read the rest at Ivanti 2025 Technology at Work Report →
