AI Training on AI: How Model Collapse Is Destroying Content Readability

When Models Start Eating Their Own Output

The AI industry is facing an increasingly thorny problem: when large language models (LLMs) are trained on text generated by other AI systems, the readability of their output drops sharply. This phenomenon has sparked widespread discussion on Reddit and once again pushed Model Collapse into the spotlight.

The proportion of AI-generated content on the internet is rising fast. By some estimates, a growing share of web pages, social media posts, comments, and even press releases are now produced by AI. This content then gets crawled, cleaned, and fed back as training data for the next generation of models. The result is a dangerous feedback loop: AI learning from AI, rather than from authentic human expression.

What Is Model Collapse

The Vicious Cycle of Recursive Training

Model collapse refers to the progressive degradation in model quality that occurs when a generative model is repeatedly trained on the outputs of its predecessors. Researchers have warned about this risk in multiple papers: models gradually lose the "long tail" of the data distribution — the rare but important expressions and knowledge points — eventually converging toward monotonous, repetitive, low-diversity output.

Declining readability is one of the most visible symptoms of this degradation. When a model learns from text that already carries a heavy "AI flavor" — overly templated sentence structures, stacked connective words, mechanical formatting — it amplifies those characteristics further, ultimately producing content that is even more opaque, stilted, and devoid of human warmth.

From Diversity Loss to Linguistic Rigidity

The richness of human language lies in its variety: spoken versus written, formal versus casual, concise versus elaborate. AI-generated text, by contrast, tends toward a kind of "average state" — grammatically correct but soulless.

When that average state becomes the primary training material, new models lose their grasp of linguistic nuance. The output increasingly sounds like "a model talking to a model," and ordinary readers find it harder and harder to follow.

Why Readability Is the First Casualty

Readability collapses first because it is highly sensitive to the "human-ness" of training data.

Real human writing naturally adjusts its complexity to its audience, deploying apt metaphors, deliberate pauses, and emotional texture. AI text, in its pursuit of "informational completeness" and "logical rigor," tends to sacrifice exactly these elements that make writing easy to understand. As models keep learning from such text, they come to mistake verbosity and complexity for "high quality," packing in ever-greater information density and degrading the reading experience in the process.

This creates a paradox: a model may look more "professional" on certain quantitative benchmarks, yet become increasingly difficult to read and use for real users.

Data Contamination: An Industry-Wide Challenge

Clean Data Is Becoming Scarce

As AI-generated content floods the internet, obtaining truly human-produced corpora is becoming increasingly difficult. Pre-2021 internet data is regarded by some researchers as an invaluable "uncontaminated" resource — comparable to "low-background steel," the metal produced before nuclear testing that is free of radioactive contamination. Data produced after that point inevitably carries AI traces.

This means the barrier to training high-quality future models may lie not in compute or parameter scale, but in whether you can access sufficiently clean and diverse human-generated data.

Approaches the Industry Is Exploring

In response to the data contamination challenge, the industry is experimenting with several strategies:

• Data provenance and labeling: Watermarking or embedding metadata in AI-generated content so it can be filtered out before training.
• Preserving human data: Prioritizing verified human-authored corpora and compensating creators to secure high-quality original content.
• Mixed training strategies: Strictly controlling the proportion of synthetic data in training sets to prevent it from dominating.
• Better evaluation metrics: Introducing readability and diversity measures that more closely reflect human perception, rather than relying solely on traditional perplexity scores.

What This Means for Content Creators

For the content industry, this phenomenon actually highlights the growing value of authentic human creation. As machine-generated text spirals into self-referential decay, human content with distinctive perspective, genuine experience, and clear expression will become a scarce resource.

In other words, the more ubiquitous AI becomes, the more valuable writing with a human touch will be. This is not only a form of protection for creators — it also reminds the entire industry that AI progress must ultimately be built on the richness of the real world, not on the illusion of self-replication.

The More AI Spreads, the More Human Writing Is Worth

The topic of "readability loss when models train on AI text" may look like a technical footnote, but it cuts to a fundamental anxiety at the heart of generative AI development. If AI-generated content is allowed to contaminate the data pool without restraint, the next generation of models may not become smarter — they may instead sink into output that is increasingly incomprehensible and homogeneous.

How to strike the right balance between synthetic and real data, and how to safeguard linguistic diversity and readability, will be core questions that AI research and industry practice must take seriously going forward.