The Role of Human-Curated Directories in LLM Training Data

In April 2025, the SEO platform Ahrefs ran 900,000 freshly published web pages through its in-house AI-content detector. The result: 74.2% of them contained AI-generated text. Only about a quarter were judged purely human-written. That is the raw material now feeding the systems most people use to find anything — and it points to a problem the AI industry has known about for two years but rarely says out loud at the dinner table.

Answer engines have quietly taken over the front door of the internet. Google's AI Overviews now reach around 2.5 billion monthly users, its conversational AI Mode crossed a billion, and ChatGPT sits near 800 million weekly users. These tools don't hand you ten blue links and let you sort it out. They read the web, decide what's true, and tell you the answer. Which means the quality of what they read is no longer an academic concern. It is the whole game.

And here is the uncomfortable part: when these systems learn from a web that is three-quarters synthetic, they start to rot.

When AI Eats Its Own Tail

The technical term is model collapse, and it isn't a blog-post theory. In July 2024, Nature published research by Shumailov and colleagues showing that when a generative model is trained on content produced by earlier generative models, it suffers “irreversible defects.” Each generation of training smooths away the rare, the specific, the unusual — what the researchers call the tails of the distribution — until the model converges on a bland, confident, increasingly wrong average.

Think of it like photocopying a photocopy. The first copy looks fine. By the fortieth, the faces are gone. The danger isn't that AI invents one bad fact; it's that the unusual-but-true details — a niche provider, a regional business, a fact that only a handful of sources ever recorded — disappear first, because there was never much human signal holding them in place.

So you have a feedback loop. AI floods the web with text. The next model trains on that text. Accuracy degrades. The web fills further. The only thing that breaks the loop is a steady supply of data that a machine didn't write — verified records of real things, reviewed by real people.

Do AI Search Engines Actually Prefer Human-Written Content?

Yes — and this is the finding that should reshape how businesses think about visibility. A 2025 study by Graphite, reported by Axios, analyzed which content actually gets surfaced. The numbers are striking. Of articles ranking in Google Search, 86% were human-written. Of articles cited by ChatGPT and Perplexity, 82% were human-written. When AI-generated articles do appear, they tend to rank lower.

Read that again. The engines built on AI are disproportionately reaching for human work. Not out of sentiment — out of self-preservation. A system that wants to be trusted cannot afford to cite the synthetic sludge it is partly responsible for creating. So it learns to weight sources that are expensive to fake.