AI Content Collapse (Model Collapse)

Why AI-generated content fails—and what happens when AI trains on its own output

Ethan Smith
The ultimate guide to AEO: How to get ChatGPT to recommend your product

AI Content Collapse (Model Collapse)

"If you feed in derivatives of derivatives into the model, you'll basically take the wisdom of the crowd and that will shrink and you'll have a single opinion on everything, which is really bad." - Ethan Smith

What It Is

AI Content Collapse explains why 100% AI-generated content (without human editing) doesn't work for SEO/AEO and predicts a dystopian outcome if it ever did work. It's both a practical warning about current content strategy AND a theoretical framework for understanding AI content economics.

The core insight: There's now more AI-generated content on the internet than human-created content. Google and LLMs are specifically filtering it out because if they didn't, search would become a search engine for ChatGPT responses—making both Google and ChatGPT useless.

How It Works

The Historical Pattern

Ethan experienced this cycle firsthand in 2007 with shopping comparison sites:

  1. Discovery: Mass auto-generated landing pages worked for SEO
  2. Arms race: Everyone scraped and rewrote each other's content
  3. Spam explosion: 100 million auto-generated pages per site
  4. Google response: Panda algorithm eliminated the entire category

2020s AI content follows the same pattern:

  1. ChatGPT makes content generation trivial
  2. "AI SEO" industry emerges promising automated content
  3. More AI content than human content now exists
  4. Google and LLMs filter it out

The Evidence

Graphite's research study:

Finding Data
AI detection accuracy ~92% (8% false positive rate)
AI content in Google results 10-12%
AI content in ChatGPT citations 10-12%
AI content on the internet overall >50% (majority)

Despite AI content being the majority of content online, only 10-12% appears in search/LLM results. The filter is working.

The Wisdom of the Crowd Problem

If AI content DID work at scale, a worse problem emerges:

Normal operation: LLMs summarize opinions from many diverse sources—the "wisdom of the crowd." The average of many opinions is better than any single opinion.

AI content loop:

  1. Generate AI content → post it
  2. That content gets cited by LLMs
  3. LLMs summarize AI-generated summaries
  4. Repeat: "derivatives of derivatives"

Result: Convergence to a single opinion

"If you ask 'What's the best flavor of ice cream?' It will eventually say 'It's vanilla and it's only vanilla, and there's no other flavor of ice cream.'"

This is called model collapse—documented in academic research for both core model training and RAG.

Why This Matters for Strategy

AI-assisted content DOES work:

  • Human writer uses AI for drafts, research, editing
  • Human applies judgment, expertise, original thinking
  • Final output contains information gain

100% AI-generated content does NOT work:

  • No human-in-the-loop
  • Derivative of existing content by definition
  • Detected and filtered by search algorithms

How to Apply It

Content Strategy Implications

  1. Don't pursue fully automated content

    • It won't rank
    • It wastes resources
    • Algorithm changes will eliminate it anyway

  2. DO use AI-assisted content creation

    • AI for research, outlining, drafting
    • Human for judgment, expertise, editing
    • This is the future of content

  3. Focus on information gain

    • Did you say something others didn't?
    • Do you have original research or data?
    • Are you a domain expert?

Investment Decisions

When evaluating AI content tools, ask:

  • Does this tool help humans write better? → Good investment
  • Does this tool replace humans entirely? → Bad investment (short-term arbitrage at best)

Competitive Strategy

  • Competitors using 100% AI content will eventually get filtered
  • Investment in human expertise and original research compounds
  • "Information gain" becomes more valuable as AI content floods the internet

When to Use It

  • When evaluating AI content generation tools
  • When deciding content investment strategy
  • When competitors are using mass AI content
  • When explaining to leadership why "just use AI" won't work

The counter-intuitive insight: As AI makes content generation cheaper, human expertise and original research become MORE valuable, not less.

Source

  • Guest: Ethan Smith
  • Episode: "The ultimate guide to AEO: How to get ChatGPT to recommend your product"
  • Key Discussion: (00:51:59 - 00:58:25) - AI content research and model collapse
  • YouTube: Watch on YouTube

Related Frameworks

Back to Decision Making