Source
Kalish, Griffiths & Lewandowsky (2007), “Iterated learning: Intergenerational knowledge transmission reveals inductive biases”, LIT-820, read in NOTE-616.
What was actually shown
The design. Four conditions (positive linear, negative linear, nonlinear and random starting functions) × 8 families × 9 generations = 288 participants. Each learner was trained on the previous learner’s test responses, with no filter between generations. The bias, toward linear functions with a positive slope, was stated before the experiment from Brehmer (1971, 1974), Busemeyer et al. (1997) and Kalish, Lewandowsky and Kruschke (2004).
The result (Figs. 3A–F). 28 of 32 families converged to a positive linear function. 11 families produced a negative linear function at least once (all 8 negative-linear families, 2 random, 1 nonlinear). 3 converged to it (2 negative, 1 random), and 1 had not converged after nine generations. The median correlation with y = x rose across generations in the three conditions not already at ceiling.
The simulation (Fig. 2). Bayesian learners with a prior on lines centred on slope 1, started from 20 points on y = 1 − x, produce a positive linear function within a few generations over 1,000 chains.
What could have come out otherwise. The endpoints could have depended on the starting function. Apart from the three negative endpoints, they did not.
What this does not say
- That the chains converged to the prior. The prior is a qualitative description (“favours positive linear”) taken from earlier studies, not a distribution measured on these learners. “Converges to the prior” is tested only as “converges to the prior’s mode”, and the authors say stationarity is unconfirmed (NOTE-616, C3, weak).
- That people are posterior samplers. A strong positive-linear mode would be reached by sampling and MAP learners alike, so the experiment cannot separate them. Participants were not assumed or shown to sample.
- That there is one attractor. Three families ended on the negative linear function, which the authors read as weight in the stationary distribution (C4, weak).
- That cultural forms are tailored to biases. The discussion’s claims about Bartlett and culture are not supported here; no stories or norms were transmitted (C5).
- That it generalises. One task, with one-dimensional input and output, undergraduates from one university, nine generations, and medians and representative chains with no test statistics. Slider responses within 5 units counted as correct in training, so motor and response noise is transmitted too, unmodelled.
Connections
- The theorem it is offered for. THEORY-155 states that
posterior-sampling chains converge to the shared prior (Griffiths and
Kalish, LIT-769). This is the nearest human test the record holds, and it
does not meet THEORY-155’s
promote_when: the prior was neither measured nor manipulated, and the long-run distribution was not compared with it (NOTE-616, Bearing on the record). - The filtered chain. Kirby, Cornish and Smith (LIT-771) measured no bias and used an experimenter’s filter. This experiment has no filter and a bias known in advance; its weakness is the opposite one, a task far from language. CLAIM-066 makes that contrast.
- The record’s own argument. CLAIM-090 cites this result as its evidence from human chains, partial counter-evidence included, and CLAIM-234 lists it among the iterated-learning readings, none of which measures stance or speech-act function along a chain.
- Bartlett’s serial reproduction (LIT-768) is the experimental lineage, filed unread.