Call it what it is: the internet is drowning in machine-made filler, and most of it isn't even trying to fool you anymore. "AI slop" only entered common usage in 2024, but by 2026 it has a body count of data behind it — hundreds of millions of interactions on garbage Facebook pages, tens of billions of views on YouTube channels that never touch a human hand, and Amazon guidebooks that could get a mushroom forager killed. Spotting it is no longer a parlor trick. It's basic media literacy.
Where the flood is actually coming from
Start with Facebook. Stanford Internet Observatory researchers tracked more than 120 Pages posting surreal AI images — the "Shrimp Jesus" genre, Christ fused with crustaceans and sea life, captioned with bait like "Say Amen for 7 years of luck" — collectively pulling in hundreds of millions of interactions. Nobody is claiming those images are real. The business model is the algorithm's inability to tell the difference, and it works.
YouTube's version is quieter and worse for who it targets. A New York Times investigation in March 2026 found that roughly 40 percent of videos recommended to children, on both the main platform and YouTube Kids, look like AI slop — Cocomelon-style visuals with nobody behind the camera. Separately, Kapwing analyzed the 15,000 most popular YouTube channels and found 278 that are pure AI slop, with a combined 63 billion views and 221 million subscribers. That's not a niche corner of the platform — it's a content category with its own audience.
Amazon has its own strain, and it's the one with real physical risk: AI-generated mushroom foraging guides sold on the platform contain misidentifications serious enough to get a forager killed. On reviews, an analysis of roughly 26,000 Amazon listings found AI-generated reviews up 400 percent since ChatGPT's launch, with extreme one-star and five-star reviews 1.3 times more likely to be AI-written than moderate ones — outrage and rapture both sell, and both are cheap to fake.
Then there's the industrial version: content farms. Cybersecurity firm DoubleVerify identified a network called "AutoBait" in March 2026 — over 200 websites running templated prompts through a language model to churn out articles and images purely to harvest ad revenue. NewsGuard and Pangram Labs put the broader count at 3,006 AI content farm sites as of that same month, growing by 300 to 500 new sites monthly. This isn't a scattered problem. It's a manufacturing pipeline.
What still gives it away — and what doesn't anymore
Forget the mangled-hands trick. Image models closed most of that gap in 2026: a 2025 study in Science found human accuracy at telling real photos from AI ones had fallen to 38 percent — below the 50 you'd get guessing at random. A Microsoft study spanning 600,000-plus images put human judges at 62 percent, and a consumer study found people correctly pick the AI image 71.63 percent of the time, varying by image type. Automated tools do better — 89 to 94 percent accuracy on photorealistic images — but carry an 8 to 12 percent false-positive rate, meaning roughly one in ten real photos gets wrongly flagged. There is no clean signal, only probabilities.
Text is a different fight, and people lose it differently. After a five-minute conversation, participants misidentified GPT-4o's writing as human 77 percent of the time. Academics separating real research abstracts from AI-written ones scored 44 to 76 percent depending on the field. What holds up as a tell isn't any single word — it's density and rhythm. "Delve," "tapestry," "meticulous," and "multifaceted" show up constantly in model output and almost never in human writing. The construction "it's not just X, it's Y" appears in LLM text at rates researchers put over 1,000 times more common than in human prose. One em dash proves nothing, but fifteen in a 600-word piece, paired with uniform sentence length and a transition word opening every paragraph, is a pattern worth trusting.
The tell was never the flaw. It's the flatness — writing with no friction, images with no accident, content built to clear a threshold rather than say something.
The labels are arriving — actually use them
The infrastructure to settle this argument is finally shipping. Google's SynthID is embedded by default in every Imagen image and every Veo video, and it survives cropping, resizing, and compression. C2PA Content Credentials — backed by Adobe, Microsoft, OpenAI, Meta, and camera makers including Leica, Sony, Nikon, and Canon — attach a cryptographically signed provenance record to a file's entire edit history. Meta has attached "AI Info" and "Made with AI" labels across Instagram and Facebook off that metadata since early 2024. YouTube went further this May: a video carrying C2PA metadata marking it fully generative gets a permanent AI disclosure label in YouTube Studio — the creator can't toggle it off, and the usual appeals path doesn't apply. CEO Neal Mohan named reducing slop and catching deepfakes platform priorities back in January.
None of this makes detection effortless. It makes it checkable. A reader who wants to separate signal from slop has three real moves:
- Check for a Content Credentials or "AI Info" badge before trusting an image's origin story — don't rely on your eyes alone.
- Read for rhythm, not vocabulary: uniform sentence length, a transition word starting every paragraph, and zero specific detail are worse tells than any banned word list.
- Follow the incentive. A "channel" with no face, no name, and a caption begging for engagement is a business model, not a creator.
The slop isn't going away — the pipeline generating it is getting cheaper by the month. But the provenance tools built to counter it are real, they're shipping now, and they work better than squinting at a photo ever will. Use them.
