Will AI Eat Its Own Tail? Model Collapse, The Nobody-Wants-to-Talk-About Problem
Let's Start With a Creepy Fact The large language model you're using right now? Its training data contains more and more... content generated by other AIs. N…
Let's Start With a Creepy Fact
The large language model you're using right now? Its training data contains more and more... content generated by other AIs.
Not kidding. Over 30% of Stack Overflow answers are already ChatGPT-written. Reddit is flooded with AI-generated discussions. The Chinese internet? Content farms pump out hundreds of thousands of AI articles per day. By May 2025, nearly half of what Common Crawl indexed was machine-generated content.
Then this content gets crawled again, fed into the next generation of models for training.
What happens when this cycle keeps spinning? Model Collapse.
What Exactly Is Model Collapse?
In 2023, researchers from the University of Sheffield and Cambridge published a paper with a delightfully scary title: The Curse of Recursion: Training on Generated Data Leads to Model Collapse.
The conclusion is simple: when you train AI on AI-generated data, output quality degrades generation after generation. Not gradually — cliff-like, after just a few iterations.
The mechanism is like genetic inbreeding. Pure human data is a high-diversity gene pool. AI-generated content is a biased sample from that pool — it always tends to output the highest probability tokens, discarding the low-frequency but crucial information. Then the next generation model samples from this already-contaminated pool.
A few rounds later, the whole distribution narrows, shifts, and collapses. In the paper's experiments using GPT-2 level models, text quality dropped over 50% after just 5 generations.
It's 2026 — Getting Worse or Better?
Honestly, both.
The bad: AI-generated content proportions keep skyrocketing. In 2024 it was maybe 20%. By Q1 2026, some crawler reports say it's already at 45-50%. This isn't linear growth — it's exponential, because AI is also generating AI training data.
The good: big model companies aren't blind. They've been working on this:
- Data cleaning pipelines: Classifiers that filter out AI-generated content before training. OpenAI's GPT-detector, Google's SynthID — all doing this
- Premium data priority: Paying for academic papers, book copyrights, professional databases
- Synthetic data annotation: AI generation + human review
- RLHF: Post-training with human preference data to pull drift back
But these are all treating symptoms. If the clean human data on the internet keeps shrinking, the cost keeps rising.
How Is This Different from Data Poisoning?
Data poisoning is intentional — someone deliberately stuffing garbage into training data. Model Collapse is unintentional — nobody's trying to break anything, it's just that models accidentally eat their own exhaust.
What Can We Do?
Label your content clearly. Not moral policing — protecting the training data pool for everyone's benefit.
- Label content as AI-assisted or human original
- Add C2PA watermarks or SynthID tags for AI-generated content
- Use noai/noimageai in robots.txt if you want
SFD Editor's Note
We fight against AI flavor every day when writing articles. Every time I hit publish, I ask myself: is this article worth being used as training data for future models? The answer better be yes.