Human iterated-learning chains converge to the learners' shared prior within a few generations, regardless of the seed input

In Kalish, Griffiths & Lewandowsky's (2007) iterated function-learning experiments, participants were shown a scatter of (x, y) points generated by the previous participant's guesses and asked to predict y for new x values; their predictions became the training data for the next participant in the chain. Across 32 independently-seeded chains ("families of learners"), regardless of the function the chain started from, "iterated learning converged to a linear function with positive slope in only a few generations for 28 of the 32 families of learners." The chains did not preserve or gradually erode the seed function — they snapped to a shared prior (a positive-linear relationship) within roughly one to four transmission steps.

This is the empirical anchor for the claim that iterated-learning theory applies to AI model collapse: both are transmission chains whose fixed point is the learner population's own inductive bias, not the transmitted content. It is also a modern, quantified counterpart to Bartlett's serial-reproduction findings that a durable "deep structure" — see claim-gersick-1991-punctuated-equilibrium-deep-structure for the organizational-theory cognate of that phrase — survives repeated retransmission while surface content does not; see question-bartlett-distortion-typology-maps-to-model-collapse-failure-modes for the open question of whether Bartlett's own distortion typology maps onto collapse failure modes specifically.