A 2025 position paper argues 'model collapse' research encompasses eight distinct and at times conflicting definitions

Schaeffer, Kazdan, Arulandu, and Koyejo, in the 2025 arXiv position paper "Model Collapse Does Not Mean What You Think" (arXiv:2503.03150), survey prior model-collapse research and report a terminological problem at its centre: "We highlight that research on model collapse actually encompasses eight distinct and at times conflicting definitions of model collapse, and argue that inconsistent terminology within and between papers has hindered building a comprehensive understanding of model collapse." On this account, the AI field's working typology of collapse failure modes was assembled bottom-up from statistical and technical considerations — collapsing variance, disappearance of distributional tails, mode disappearance or entanglement, appearance of hallucinated data — rather than imported from, or cross-walked against, any prior psychological taxonomy such as Bartlett's.

The abstract, read directly via WebFetch, does not mention Bartlett, serial reproduction, or assimilation/leveling/sharpening in the cognitive-psychology sense; the full paper body was not read this session, so an absence of any Bartlett citation in the full text is not independently confirmed beyond the abstract.

[unverified-mechanism -- needs primary] The specific enumeration of the eight named definitions is not verified. Only the top-line count ("eight distinct definitions") is quote-grounded in the abstract; the individual sub-definitions above were reconstructed from an aggregated search summary, not a directly quoted full-text passage. The verification is routed to question-verify-schaeffer-2025-eight-model-collapse-definitions.

This claim is the AI-side half of a would-be bridge: the independently-built, Bartlett-free collapse taxonomy that Bartlett's three-category typology would have to be mapped against. The other half — the one paper that invokes Bartlett as analogy — does not attempt the mapping, and that no located source does is recorded in observation-bartlett-typology-not-yet-operationalized-as-model-collapse-metrics. It also underlies the iterated-learning reframing, which argues the statistical account "lacks a linguistic explanation for which structures degrade."