Is model collapse inevitable when most online content becomes AI-generated?
kreschnav · reddit · 2026-09-15
An ambivalent Reddit user raises a long-horizon objection: if model collapse — degradation from training AI on AI-generated content — is a real concern, how do you avoid it 10-20 years out, when AI is pushed into everything, most people use it for all content creation, and the majority of online text is AI-generated? He concedes some human content will always exist, but argues that under near-universal adoption, training on AI content seems unavoidable, making collapse look inevitable. The thread invites debate over synthetic-data training strategies, data provenance, and whether collapse concerns are overstated.
More from AGI Musings
- Full video: Terence Tao on why LLM capabilities remain unpredictable — FinanceYF5 · 2026-09-15
- Terence Tao: LLM math is easy; unpredictability of capabilities is the hard part — FinanceYF5 · 2026-09-15
- Halvar Flake: AI risk discourse is measured in ego, not fear — basedjensen · 2026-09-15
- New models are more obedient, not less: why intentionality, not intelligence, is the real risk question — LucaAmb · 2026-09-15
- antirez: The core AI misunderstanding is tool vs. liberation from human limits — antirez · 2026-09-15
- AI psychosis is self-reinforcing, and benhylak says those jokes are cries for reassurance — teodorio · 2026-09-15