Feedback loop and baseline contamination
Parent: Human adoption of LLM vocabulary and its effect on tic baselines · Published reference · snapshot 2026-09-23
↓ Facts as markdownall context files
Source-anchored research for Feedback loop and baseline contamination. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Problems, failure modes and limitations
- Model collapse under a replace-data regime (each generation's synthetic data replaces the prior data) degrades models, but if synthetic data accumulate alongside the original real data, test error has a finite upper bound independent of iteration count; this held across language models, diffusion models and VAEs. (medium confidence) [source]
Facts and statements
- A Max Planck study of 737,083 hours of podcast speech (824,634 episodes) found abrupt post-ChatGPT increases in words such as delve, showcase, boast, intricacies and meticulous; synthetic-control analysis supports a causal link, and the authors frame it as machines trained on human data feeding their traits back into human language. (low confidence) (verified as of 2026-07-16) [source]
- A preregistered experiment (N=496) found brief chatbot interaction led participants to adopt the chatbot's words, persisting through a distractor task and in forced lexical choice. (low confidence) [source]
- The term 'model collapse' is used inconsistently in the literature and is conflated with generic performance loss or any synthetic-data training; critics argue claims of inevitable collapse need scoping to the specific definition and data regime. (low confidence) [source]
Children
- No children recorded.