- AI model collapse is already underway as models increasingly train on AI-generated content rather than original human writing.
- AI model collapse isn’t a future risk — it’s a structural problem quietly compressing intelligence across every major LLM today.
- Labs like Reddit, publishers, and forum networks are now selling human-generated data as premium AI training infrastructure.
- Scaling compute cannot fix a poisoned data distribution — more GPUs just produce faster, more confident averages.
- AI model collapse is already underway as models increasingly train on AI-generated content rather than original human writing.
- AI model collapse isn’t a future risk — it’s a structural problem quietly compressing intelligence across every major LLM today.
- Labs like Reddit, publishers, and forum networks are now selling human-generated data as premium AI training infrastructure.
- Scaling compute cannot fix a poisoned data distribution — more GPUs just produce faster, more confident averages.
AI Model Collapse and the Data Problem Everyone Is Ignoring
The AI industry has spent the last three years obsessing over compute. More GPUs. Bigger clusters. Faster training runs. And for a while, that obsession was justified — scaling worked, almost embarrassingly well. But AI model collapse is now emerging as the constraint that raw compute was never designed to solve, and it’s arriving quietly, structurally, without a dramatic announcement from any lab.
The problem isn’t that we’re running out of processing power. It’s that we’re running out of something far harder to manufacture: high-quality, genuinely human-generated data. And in the vacuum left behind, we’ve started filling training sets with the one thing we had plenty of — output from the very models we’re trying to improve.
What the Early Internet Actually Gave AI
It’s easy to romanticise the early web, but there’s something real worth acknowledging here. The data that trained the first generation of serious foundation models wasn’t clean. It wasn’t structured or optimised for machine readability. It was messy in exactly the right ways.
Think about what that actually looked like in practice. Stack Overflow answers typed out under pressure at 2am by engineers who’d just spent four hours debugging a race condition. Reddit threads where someone’s confident wrong answer got publicly dismantled in the replies. GitHub repositories with half-finished documentation and commit messages that told the real story of how software actually gets built. Research papers with genuine uncertainty in the methodology sections, not just hedged conclusions added by reviewers.
That wasn’t just content. It was compressed human reasoning under real constraints — disagreement, failure, correction, and the occasional unexpected leap that nobody could have predicted. The chaos had signal in it. The contradiction density was high. And that variation, it turns out, is exactly what intelligence needs to keep compounding.
We’re losing it, and we’re losing it fast.
The Recursive Loop Nobody Wants to Name
Here’s what’s actually happening across the web right now. A substantial and growing share of online content is AI-written blog posts, SEO pages generated at scale, code snippets that have been rewritten by multiple LLMs before landing anywhere, and summaries of summaries of summaries. Individually, each piece looks fine. Collectively, they’re reshaping the training distribution in ways that are structurally dangerous.
The loop looks like this: human data trains a model, that model generates content, that content enters the web, and the next generation of models trains on a dataset that’s partially its own ancestor’s output. Repeat the cycle. Each iteration quietly reduces variance, originality, contradiction density, and the weird edge cases that force a model to actually reason rather than pattern-match. Each iteration increases stylistic convergence, templated explanation, and what you might call safe average reasoning — the kind of output that offends nobody and surprises nobody.
This is the mechanism behind AI model collapse, and it’s not theoretical. Research published in Nature in 2024 by Ilia Shumailov and colleagues at the University of Oxford formally demonstrated model collapse in controlled settings, showing that models trained on recursively generated data experience
