nucleation

CLAIM-090: Serial transmission is reconstructive and filtered by the transmitters' inductive biases, so a stable endpoint can be far from the ancestral utterance while remaining interpretable

2 min read

Properties12
StatusProposed
Version3
Rolethesis
Warrantempirical
Tagslinguistics, cognition, compositionality
Linepragmatic-transport
Workswhat-survives-translation
GroundsTHEORY-155 — Iterated learning by Bayesian agents who sample from the posterior converges, in an ergodic chain, to the shared prior, LIT-769 — Language Evolution by Iterated Learning With Bayesian Agents, LIT-771 — Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language, THEORY-489 — Chains of human function learners, each trained on the previous learner's responses, forget their starting function and in 28 of 32 families converge within a few generations to a positive linear function, the bias earlier studies report, whatever function trained the first learner; this is convergence to the learners' favoured function, not shown to be convergence to their prior
SupportsCLAIM-068 — Repeated reconstruction is attraction toward conventional regions of communicative possibility: pragmatic stances, not only lexicon and grammar, can be stabilized by transmission
Illustrated byCASE-068 — Risk messages passed down diffusion chains: the evaluative signal kept and amplified while the content decays (Moussaïd, Brighton and Gaissmaier; Jagiello and Hills)
Cited as “concludes” byARG-035 — Griffiths and Kalish's convergence theorem and Kirby, Cornish and Smith's chains show that serial transmission is filtered by the transmitters' inductive biases
Rendered intodocs/arguments-by-conclusion.md

Graph View

The claim

§8: “Bartlett’s serial reproduction experiments and Bayesian iterated-learning models both show that transmission is reconstructive and filtered by inductive biases rather than neutral copying.” The manuscript adds the data-processing inequality: information about the ancestor cannot increase along a Markov chain, even as later retellings become easier to process.

What it does not say

It does not say messages get less comprehensible along a chain; the manuscript is explicit that conventionalization can make them easier. Bartlett is unread here (LIT-768).

Evidence from human chains

Kalish, Griffiths and Lewandowsky (LIT-820) report human function-learning chains: 28 of 32 converged on a positive linear function, a bias stated before the experiment, mostly whatever the starting function. Two of the eight negative-start families ended on the negative linear function, which is partial counter-evidence. The prior was taken from earlier studies rather than measured on these learners.

Note of 2026-10-29: what holds it back

The claim’s general form, that transmission is filtered by the transmitters’ inductive biases, waits on a chain whose endpoint matches a prior measured on its own transmitters before transmission (QUESTION-137). THEORY-489’s chains converge on a bias taken from earlier studies and do not tell sampling learners from MAP learners. CASE-068’s priors were measured after the experiment (Moussaïd), or predict less than the received message (Jagiello and Hills). ARG-035, from the two readings the manuscript cites, is Proposed because it holds only for sampling learners.

Backlinks