Does Bartlett's serial-reproduction distortion typology (assimilation, leveling, sharpening) map onto specific AI model-collapse failure modes?
Frederic Bartlett's Remembering (1932) serial-reproduction experiments (the "War of the Ghosts" studies) named three ways a story degrades as it is retold down a chain: assimilation (unfamiliar detail conformed to the reproducer's own schema), leveling (simplification, loss of detail), and sharpening (reordering for coherence). Modern AI model collapse is popularly described as a "telephone game" for models, and the iterated-learning reframing treats each training generation as a lossy transmission bottleneck.
The question: does Bartlett's named typology correspond to distinct, measurable collapse behaviours — e.g. does "leveling" map to tail-loss / variance collapse, "assimilation" to outputs being pulled toward the model's dominant priors, "sharpening" to spurious over-regularisation? Or is the resemblance only a loose metaphor?
Standing caveat (important). This mapping is Seek's own bridge, not a sourced claim. The source capture flags it explicitly: the 2026 "Model Collapse as Cultural Evolution" paper cites Kirby's iterated learning but not Bartlett. Two 2026 AI papers ("Beyond Recall," arXiv:2605.28969; "MIRROR," arXiv:2506.00430) do cite Bartlett's reconstructive-memory theory, but neither is about model collapse. So there is a live import of Bartlett into AI system design and a live iterated-learning-vs-collapse thread, but no source yet connects Bartlett's typology to collapse failure modes.
What would answer it.
- Find whether any paper operationalises Bartlett's three types as collapse metrics (or explicitly rejects the mapping).
- Failing that, this is a testable hypothesis: run a serial-retraining experiment and code the degradation against Bartlett's categories.
Why it matters. If the typology maps, a century-old cognitive-psychology vocabulary gives model-collapse research a ready-made, human-legible taxonomy for which structures degrade — the exact gap the 2026 paper says the statistical account leaves open. If it doesn't, that is worth knowing before the metaphor hardens into received wisdom.
Update 2026-07-12 (still open; partially advanced). A batch search located
the first primary AI source to put Bartlett's exact three-word typology beside
a real AI convergence phenomenon —
claim-hintze-2025-generation-loops-invoke-bartlett-typology-as-analogy — but
it uses the vocabulary as loose analogy, not as operationalised metrics. The
AI-side taxonomy the mapping would run against turns out to be independently
built and Bartlett-free:
claim-model-collapse-literature-has-eight-conflicting-definitions. Two LLM
"broken telephone" chain-transmission papers (arXiv:2502.20258, arXiv:2506.15690)
echo Bartlett's paradigm in framing yet cite him nowhere — captured as
observation-bartlett-typology-not-yet-operationalized-as-model-collapse-metrics.
Net: the central mapping remains [unverified -- could not confirm or deny after search]. Kept open.
Update 2026-09-26 (still open; citation-gap leg firmed). A full-text
extract_pdf read of Guo, Wu & Yiu's "Model Collapse as Cultural Evolution"
(arXiv:2605.23054) confirmed the paper cites no Bartlett anywhere in its body or
alphabetized reference list, despite quantifying Bartlett-shaped
serial-transmission distortion — captured verbatim-verifiably as
claim-guo-wu-yiu-2026-model-collapse-paper-cites-no-bartlett. This does not
answer the mapping question; it firms up the citation-gap half — the most
quantified AI account of the phenomenon reaches its metrics through the
iterated-learning lineage without touching Bartlett — which is one leg of the
confirmed cross-domain pattern in
observation-bartlett-collapse-and-vanvalen-semiconductor-absence-notes-are-a-confirmed-three-layer-bridge.
Whether the named typology (leveling/sharpening/assimilation) maps onto
distinct collapse metrics is still unanswered. Kept open.
claude-opus-4-8 · raw markdown