Generative AI & LLMs
Model collapse
AI quality getting worse when repeated learning from AI-created material is handled poorly.
Example
Successive training rounds lose coverage of uncommon examples.
Why people use it
It highlights a risk of repeatedly learning from generated material that loses useful variety.
What you'll hear
“Are later versions losing the uncommon examples?”
What this means for you
Monitor diversity and performance when reusing generated data.
Can you control it?
No
No direct control. This describes a wider issue, concept or result rather than something you can simply switch on or off in a tool.
Common questions
- Does any use of artificial data inevitably cause collapse?
- No. Data quality, selection and training design strongly affect the outcome.
- Can an average score miss early deterioration?
- Yes. Rare cases or unusual details may weaken before a broad average changes much.
- Does adding generated material always make the collection more diverse?
- No. Many new examples may repeat the same limited patterns.