A growing body of literature suggests that making differentsampling assumptions about how data are generated can leadto qualitatively different patterns of inference based on thatdata. However, relatively little is known about how samplingassumptions are represented or when they are incorporated.We report the results of a single category generalisation exper-iment aimed at exploring these issues. By systematically vary-ing both the sampling cover story and whether it is given beforeor after the training stimuli we are able to determine whetherencoding or retrieval issues drive the impact of sampling as-sumptions. We find that the sampling cover story affects gen-eralisation when it is presented before the training stimuli, butnot after, which we interpret in favour of an encoding account.

How do cumulative cultural evolution and individual learningdiffer? In an abstract computational sense, both are optimisa-tion processes that search a space of possible explanations andprevious work has identified deep parallels in the mathematicalmodels used to describe them (Suchow, Bourgin, & Griffiths,2017). However, there are obvious differences as well: forexample, individual learning involves a single agent charac-terised by one set of prior beliefs, representational capabilities,and so forth, while cultural evolution involves multiple agentswho may vary along these factors. We argue that this differ-ence implies that the process of cumulative cultural evolutionshould involve searching a more restricted set of hypothesesand converge on simpler ones. In two iterated category learn-ing experiments, we test this prediction and find that transmis-sion chains composed of single individuals, who learn basedon their previous performance, consider both a wider varietyand more complex categorisation schemas than do chains in-volving multiple people.

A wealth of developmental evidence suggests that children es-sentialise natural kind but not artifact categories, and that bothadults and children use generic language less with artifacts aswell (Gelman, 2003). Here we further explore the latter resultusing a novel model for generic identification. We apply ourmodel to a much larger dataset than before, consisting of 26CHILDES corpora of naturalistic speech involving children ata variety of ages and in a variety of contexts. We found noconsistent preference for generic usage in animates over arti-facts. Follow-up analyses indicate that this result was probablydriven by our inclusion of a wider variety of nouns into ourdataset than previous work.

Many real-world situations involve learning entirely or mostlybased on the information provided by other people, which cre-ates a thorny epistemological problem: how does one deter-mine which of those people to trust? Previous work has shownthat even populations of rational Bayesian agents, faced withthis problem, polarise into “echo chambers” characterised bydifferent beliefs and low levels of between-group trust. In thisstudy we show that this general result holds even when thereasoners have a more complex meaning space and can com-municate about their beliefs in a more nuanced way. However,even a tiny amount of exposure to a mutually trusted “groundtruth” is sufficient to eliminate polarisation. Societal and psy-chological implications are discussed.