---
title: "Michael L. Kalish"
type: "entity"
entity_kind: "person"
status: "hub"
canonical_name: "Michael L. Kalish"
aliases: ["Michael Kalish","M. L. Kalish","Mike Kalish"]
first_seen: "2026-07-06T00:00:00.000Z"
writer_model: "warden/claude-opus-4.8"
connects_to: ["iterated learning","Bayesian models of cognition","function learning (inductive bias)","[[entity-thomas-griffiths|Thomas L. Griffiths]]","model collapse (cross-domain descendant)"]
provenance: "Warden pass 2026-09-26 (warden/claude-opus-4.8), run per 00-meta/specs/seek-warden-spec.md on a maintenance engine kept separate from the notes' writers. Builds the Kalish person hub the 2026-09-26 [entity] flag (seek-flags.md L5717) asked for (paired with Simon Kirby — the Kirby half is carried, not built: no Kirby primary is read in the vault). Person test run honestly against a direct read this pass of the two Kalish-authored primaries the vault holds (claim-kalish-2007..., claim-expedient-knowledge-blocks-restructuring) plus the notes that place him in the iterated-learning lineage; not the flag summary. Closes the asymmetry whereby his co-author [[entity-thomas-griffiths]] is already a hub while Kalish, named in that hub's own body as 'Griffiths & Kalish', was not. Build 1 of the run's <=2. Report: warden-2026-09-26.md."
audit_status: "2026-09-26 (warden/claude-opus-4.8, writer): single-model build on a maintenance engine; not yet cross-model audited. Both Kalish-authored notes re-read this pass; no claim restated beyond what they carry. The 2007 note is Tier 1, verified-verbatim (seek_verify, 2026-08-07); the 2000 note is capture-verified. No independent affiliation for Kalish is recorded in the vault, so the opening describes him by his read work rather than by an institution."
seek_code_commit: "f8a723c"
---


Michael L. Kalish is a cognitive scientist whose experimental work established quantitative results on how information is transformed as it passes through chains of human learners — both in iterated function-learning (with Thomas L. Griffiths and Stephan Lewandowsky, 2007) and in category-learning and knowledge restructuring (with Lewandowsky and Griffiths, 2000).

Matters to this vault as the operationalized, statistically-quantified end of an argument the vault keeps returning to: the gap between a classic theory and the field that later put numbers on it. His 2007 iterated function-learning experiments are the vault's actual "how many generations" figure — chains converged to the learners' shared prior "in only a few generations for 28 of the 32 families of learners" ([[claim-kalish-2007-human-iterated-learning-converges-few-generations]]) — which is the empirical anchor for reframing AI [[claim-model-collapse-recursive-training-erases-distribution-tails|model collapse]] as iterated learning ([[claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution]]) and the modern, quantified counterpart to [[entity-frederic-bartlett|Bartlett]]'s serial-reproduction findings. That descendant lineage — "Griffiths & Kalish's Bayesian convergence work" — is precisely what the 2026 model-collapse paper reaches for while never citing Bartlett ([[observation-bartlett-collapse-and-vanvalen-semiconductor-absence-notes-are-a-confirmed-three-layer-bridge]]), the metaphor-vs-formalism pattern mapped at [[moc-a-shared-slogan-is-not-a-shared-model]].

What the vault has actually read of Kalish: two of his co-authored primaries directly — Kalish, Griffiths & Lewandowsky (2007), *Psychonomic Bulletin & Review* (Tier 1, verified-verbatim), and Lewandowsky, Kalish & Griffiths (2000), *J. Exp. Psychol.: Learning, Memory, and Cognition*. He therefore recurs across more than one cluster (iterated learning; category-learning / resistance to knowledge restructuring), and always as an author whose work is read at primary, not as a bibliographic mention. His institutional affiliation is not recorded in the vault and is left off the page rather than guessed.

## References
- [[claim-kalish-2007-human-iterated-learning-converges-few-generations]] — his 2007 iterated function-learning result: convergence to the shared prior in a few generations (Tier 1, verified-verbatim)
- [[claim-expedient-knowledge-blocks-restructuring]] — Lewandowsky, Kalish & Griffiths (2000): expediency vs. resistance to knowledge restructuring in category learning
- [[observation-bartlett-collapse-and-vanvalen-semiconductor-absence-notes-are-a-confirmed-three-layer-bridge]] — "Griffiths & Kalish's Bayesian convergence work" as the operationalized descendant a new-domain paper bypasses
- [[claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution]] — the model-collapse reframing his experiments empirically anchor
- See also [[entity-thomas-griffiths]] (co-author, already hubbed), [[entity-frederic-bartlett]], [[moc-a-shared-slogan-is-not-a-shared-model]]