nucleation

LIT-771: Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language

2 min read

Properties6
StatusActive — skimmed 2026-10-09 (NOTE-596): every section of the main text was covered, but through a fetch tool's section-by-section account of the PubMed Central page, not a verbatim copy, and the SI was not reached; worth reading as the laboratory demonstration of iterated learning: four chains of ten adults, each learning a 27-item artificial language from 14 items and passing on its test output, made the languages steadily easier to learn and more structured while believing they were copying. Unfiltered, the languages collapsed to 2–5 words that systematically underspecified meaning (colour lost in every chain); with homonyms filtered out of training, they became compositional. Consistent with THEORY-155, not a test of it: no prior was measured, and the filter is a selection the theory excludes.
Tagslinguistics, compositionality, social-science, cognition
SupportsCLAIM-066 — Kirby, Cornish and Smith show structure emerging under transmission with an experimenter's filter, not filtering by measured inductive biases, and Griffiths and Kalish's convergence to the prior holds only for sampling learners, 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, CLAIM-234 — The record's iterated-learning readings bear on lexical, structural and functional transmission, and none measures stance, footing or speech-act function along a chain
Cited as “grounds” byARG-027 — The record's iterated-learning readings show that pragmatic stances and footing are stabilized by transmission, ARG-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
Cited as “paper” byNOTE-596 — Cumulative cultural evolution in the laboratory
Cited as “source” byTHEORY-155 — Iterated learning by Bayesian agents who sample from the posterior converges, in an ergodic chain, to the shared prior

Graph View

Simon Kirby, Hannah Cornish and Kenny Smith (2008), PNAS 105(31):10681–10686 — DOI-10.1073/pnas.0707835105

Key takeaways

  • Design. Diffusion chains: 27 pictures (three shapes × three colours × three motions), each labelled by a written string; each adult saw 14 string–picture pairs, was tested on all 27, and that output, re-split at random, trained the next. Four chains of ten generations per experiment, 80 participants in all, starting from random 2–4-syllable labels. Participants were told to reproduce the language, not that they were in a chain.
  • Experiment 1: learnable by underspecification. Transmission error (normalized edit distance to the previous generation) fell and structure (correlation of string distance with meaning distance, as a Monte Carlo z-score) rose. Distinct words fell from 27 to 2, 4, 5 and 4. The ambiguity was systematic, one word for a class of similar meanings, so a learner could rebuild the whole from a fragment. Colour distinctions were lost in all four chains; shape and motion survived.
  • Experiment 2: compositional under a filter. When strings labelling more than one picture were cut to one before training, as an analogue of a pressure to be expressive, error still fell and structure still rose, words stayed between 10 and 27, and the strings came to split into morphemes for colour, shape and motion, with irregular items.
  • Claim. The adaptation is cumulative in learnability and structure, not in expressivity, and is offered as the first experimental validation that cultural transmission yields the appearance of design without a designer.

The paper names Bartlett (LIT-768) as the first user of the diffusion-chain method. It cites Griffiths & Kalish (LIT-769, read in NOTE-599) only in a block of model references and does not use priors; read against that analysis, the lost colour distinctions look like a learner bias surfacing through transmission, and the Experiment 2 filter is a selection pressure that the analysis assumes away.

Standing in the record

Filed on 2026-10-09 at the owner’s request, as one of the works in the reference list of the owner’s working manuscript (October 2026) that the record did not yet hold. See the curation entry of that day.

Backlinks