Learning from reliable and unreliable speakers

Abstract

As a way to bypass the need for the many iterations of individual experiences required for trial and error learning, humans and several non-human animal species have evolved a capacity to rely on information provided by others. Non-human animals typically learn from conspecifics through stimulus enhancement or emulation learning, using the same tools a conspecific used to obtain the same goal or trying to replicate the outcome of certain behavior. Human children on the other hand do not just focus on the outcome when learning from others. They pay a lot of attention to the teacher and the teaching process, draw inferences from ostensive cues, and try to replicate actions faithfully with a focus not just on what is demonstrated but also how. This high-fidelity in reliance on socially conveyed information gives rise to human cumulative culture and appears to be unique in the animal kingdom. However, it can also lead to biases, as can be seen in children’s tendency to overimitate demonstrated behavior, even if it is not instrumental in obtaining a certain result. In the domain of testimony, we also find these apparent artefacts of children’s credulity in their reliance on others. Children will sometimes reproduce testimony by an adult even if the reliability of this testimony is questionable. However, children are not always this gullible. In fact, much research has demonstrated preschoolers’ selective trust in some sources over others, based on several distinct features of these sources. An area that has received considerably coverage in recent developmental research is children’s inferences from past verbal utterances as an estimate of future reliability of the same source of testimony. While the phenomenon is well documented, less is known about its cognitive underpinnings and early development. How do children generally treat testimony from an unreliable informant? Do they always encode testimony from reliable and unreliable sources differently? Are differences in learning from reliable and unreliable speakers influenced by the extent to which these speakers cohere or contradict each other? And to what extent do toddlers rely on the same strategies as preschoolers? Or do they simply learn less in unusual or confusing situations? In this thesis, I try to discern under what circumstances and to what degree children will continue to rely on testimony from unreliable sources, what mechanisms underlie their selective trust in testimony and examine the early development of this ability. To address these issues, I conducted several experiments with 2- and 5-year-old children examining their trust in testimony using both eye-tracking and interactive measures. Whereas 5-year-olds demonstrated selective trust across a range of tasks and measures, 2-year-olds showed no evidence of selectively learning novel labels from the more reliable source. If the two speakers provided conflicting information, toddlers performed at chance level, rather than selectively endorsing the information by the more reliable source. However, in a follow-up study that presented the same word-learning demands, but not necessarily the need to consider the speakers’ respective reliability, 2-year-olds showed successful learning of novel labels. Thus, toddlers seemed to have struggled with making person-specific attribution of reliability rather than with the tasks’ word learning demands. 5-year-olds on the other hand are able to make person-specific inferences about the reliability of an informant and adjust their reliance according to the pragmatics of the situation. It appears that children are willing to learn even from a previously unreliable speaker, given that the information offered by her is not contradicted by a more reliable source. However, they do not simply discount accuracy information. Rather, in case contradictory testimony from a less unreliable source surfaces later on, 5-year-olds can still adjust their reliance and discount information from a demonstrably unreliable source. Taken together, I find that children do not generally encode information from reliable and unreliable sources differently, but they are also not gullible. My findings suggest that children’s selective learning of novel labels is not simply based on inattention towards the information provided by unreliable source, but rather on the selective encoding and consolidation of the semantic information provided by a more reliable source. At least at five years of age, human children make rational inferences about the reliability of sources. Even if they initially rely on questionable testimony with apparent credulity, these inferences let children adjust their reliance before deciding whether to reproduce such testimony. Infants likely rely on less demanding strategies than preschoolers to adjust their reliance on testimony from others, as they are not yet able to entertain person-specific inferences. In conclusion, children’s trust is testimony appears to be the result of opposing forces in form of attribution of reliability to particular sources of testimony on one hand and a credulity in reliance on socially conveyed information on the other hand.