The AI Slop Economy Runs on Unpaid Verification

There is a story going around about artificial intelligence and trust. It goes like this: AI floods the world with cheap, plausible content, so the genuine article becomes scarce, and scarcity means value, so credibility is the new moat. Build trust and win. I believed a version of this but after doing some research, my views have changed. The flood is real, but nobody is paying more for the alternative.The flood, with actual numbersRoughly half of new articles published on the web are now primarily AI-generated. That figure comes from Graphite, which sampled 55,400 pages from Common Crawl and averaged three detectors: 49.9 percent in the first quarter of 2026, and roughly flat near half for five straight quarters.Music is further along. Deezer reported that at its peak in June 2026 it was receiving about 90,000 fully AI-generated tracks a day, more than half of everything uploaded.Books, the same shape. A team of researchers tracked 14,419 self-published genre titles against daily Amazon sales through June 2026. Selling titles grew 19.2 times. Revenue grew 8.9 times.Now hold those two Deezer numbers next to each other, because this is where the comfortable story gets its evidence. More than half of daily uploads are fully AI. Those tracks earn between one and three percent of streams.That is a twenty-to-one decoupling between how much gets made and how much gets heard. You see it in text too. AI writes about half of new articles and holds about 14 percent of Google's ranking set, taking only 7 percent of first positions. In books, titles with substantial AI text are 20 percent of the catalog and 11.3 percent of revenue. Among the top five percent of bestsellers, 73 percent contain no AI text at all.Volume and attention have come apart. So far, so reassuring.Where the story falls apartIf scarcity were raising the price of the real thing, you would see it in the price. Go look.In that same book study, revenue per title fell in seven of eight genre clusters for books with no AI text whatsoever. Not for the slop. For the human-written ones. Flooding did not create a premium for the authentic, it compressed the economics for everybody standing in the aisle.Then there is the cleanest test case available. Shutterstock's entire business is verified, rights-cleared, provenance-documented, human-made content. If credibility were becoming scarce and valuable, this is the first place it would show up on an income statement. Its second quarter: revenue $221.8 million, down 17 percent. The data and distribution line, which is where AI licensing lives, down 16 percent for the quarter and 28 percent for the half. Net loss of $155.9 million against $29.4 million of net income a year earlier. The merger that was supposed to consolidate the market died on July 7 and the stock fell more than 30 percent.And the licensing gold rush is thinner than it looks from the headlines. Wiley, sitting on one of the largest scientific corpora on earth, has booked $110 million or more in lifetime AI revenue. Only $8 million of that is recurring. The number that stopped me sits in its fiscal 2026 fourth-quarter investor deck, filed with the SEC in June: Wiley reports four LLM training customers. Four, worldwide, for the entire body of peer-reviewed science.That is not a market. That is a handful of buyers who bought a corpus once.What actually changedHere is the part I think most people have backwards.The cost of producing a plausible artifact went to zero. The cost of checking one did not move. And crucially, nobody is paying a premium for the checking. So verification did not become a product. It became an expense, and it landed on whoever was standing closest.Watch where it landed.In July, security researchers at JFrog found that a single new GitHub account had filed 55 vulnerability reports, and 54 of them were fabrications. Not exaggerations. Inventions. One described a use-after-free in a function that does not exist in the version it named. These advisories reached the National Vulnerability Database carrying critical severity scores, with government enrichment attached, and one was initially scored 10.0.JFrog's explanation of how is the most important sentence in this whole argument: "Because no step in today's system actually requires a proof-of-concept or bug reproduction, a plausible-sounding fake advisory can slide right through the pipeline."Read that again. The global vulnerability pipeline never required reproduction. It did not have to, because writing a convincing fake security advisory used to be expensive enough that almost nobody bothered. The system was not running on verification. It was running on the cost of lying, and that cost just went to zero.Every institution built on that same quiet assumption is now discovering it at once, and paying for it. curl shut down a seven-year-old bug bounty in January after its confirmation rate collapsed from over 15 percent to under 5. arXiv began requiring prior peer review for survey papers. Wikipedia added a speedy-deletion criterion for AI-generated drafts. The National Vulnerability Database moved its entire pre-March backlog to unscheduled.None of those are investments. They are all costs, incurred defensively, by organizations that will not earn an extra dollar for incurring them.The honest complicationI should tell you what cuts against me, because leaving it out would be a small demonstration of exactly the problem.There is no measured penalty for AI content. Ahrefs looked at 600,000 pages across 100,000 keywords and found the correlation between AI-written percentage and ranking position was 0.011, which is to say none at all. Independent trackers show AI's share of top-20 Google results rising steadily, straight through multiple core updates. Google's spam policy is explicitly indifferent to authorship, targeting scaled valueless content "whether automation, humans or a combination are involved." Anyone telling you Google punishes AI writing is selling something.And the best-known slop story has an ending nobody quotes. After curl killed its bounty, Daniel Stenberg reported in April that "the slop situation is not a problem anymore." Report volume doubled again. The confirmed-vulnerability rate returned to 15 to 16 percent, which is above the pre-AI 2024 baseline. Nearly every report now involves AI assistance, and the good ones are better than what people produced unaided.So it was never the technology. It was the money. Remove the bounty and the incentive to mass-produce plausible garbage disappears, and what remains is a genuinely better-equipped researcher. That is the most hopeful fact in this entire piece, and it points at the real lever.What this means if you make thingsThe instruction people take from all this is "add value, be authentic, build trust." That advice predates the flood by fifteen years and it is not what the data supports.What the data supports is narrower and more useful.Stop expecting a premium and start budgeting a cost. Proving your claims is now a line item, not a differentiator. I ran a 21-trial controlled experiment for one article earlier this month, and the writing was the cheap part by a wide margin. Nobody paid me extra for the trials. The alternative was publishing an assertion, which is what the rest of the coverage did, and assertions are now free and worth what they cost. I run my agency's work on that assumption and I would rather state it plainly than pretend rigor sells itself.Optimize for reproducibility, not for authorship. The evidence is indifferent to whether a human typed it and quite sensitive to whether a claim can be checked. Publish the method, the numbers, the seed, the config. That is the property machines and humans both reward, and it is the only one AI cannot manufacture for you.Watch who absorbs the verification cost in your own field, because that is where the next failure is. It will be whoever was quietly relying on production being expensive.One last image, which I cannot stop thinking about. On August 2, the EU began legally requiring AI output to be marked in a machine-readable format and detectable as artificially generated. On August 14, Anthropic shipped text watermarking for Claude and noted that the detection API does not exist yet, and that the mark works worst on short passages and on factual text.Detectability became a legal obligation before it became a shipped capability. That gap is not a scandal. It is just the shape of the whole problem, sitting out in the open: we mandated proof faster than anyone built the means to provide it, and nobody has worked out who pays.