3.1. Interaction between perceptual and prior information

In the 4 experiments of the present study, the degree of prior expectations contribution strongly depended on the reliability of the visual information conveyed by the video scenes. In conditions with low amount of visual information, whatever the type of intention, participants tended to give priority to preferred intentions at the expense of non-preferred intentions ; that is, to rely mostly on the intention they estimated to be the most likely cause of what was observed. This tendency to favor prior knowledge over perceptual information may further be accounted for byconsidering intentional inferenceas an inverse problem (Kilner et al., 2007a, 2007b; Baker et al.,2007; see also Csibra & Gergely, 2007). Inverse problems characterise situations in which the same sensory input can have different causes. This type of problem is commonly encountered in ambiguous perceptual tasks – such as those using bi-stable or degraded stimuli – the resolution of which requires appealing to prior knowledge or making further assumptions about the nature of the observed phenomenon (Mamassian & Goutcher 2001). The significant contribution of prior expectations in conditions of high visual ambiguity precisely suggests that when sensory information was not sufficient to unambiguously infer one intention, participants compensated by massively appealing to their prior knowledge (i.e. about the space of the agent’s possible intentions). This strategy resulted in a significant shift towards the intention with the highest probability.

Overall, this result comforts the idea that, in situations of sparse or incomplete data, a successful inference depends on an adaptive integration between bottom-up information (from the observation of behaviour) and top-down prior knowledge about goals or intentions (Baker et al., 2006). This integration is consistent with a mechanism complementing the available perceptual information when it does not sufficiently constrain the number of potential solutions, namely, the many competing intentions congruent with what is observed. In line with this assumption, some authors have suggested that inferring another person’s intention necessarily requires sensorimotor information to be complemented with information about mental states and attitudes (de Lange et al., 2008). Besides, it has been demonstrated that prior expectations are frequently used by children, even at a very early age. This tendency combines with a tendency to interpret actions as being directed towards a goal (‘teleological obsession’, Csibra & Gergely, 2007). When the visual information is not sufficient for interpreting the action as a goal directed one (Csibra, Gergely, Bíró, Koós, & Brockbank, 1999), or when the action is incomplete (Onishi, Baillargeon, & Leslie, 2007),children posit states of the world occasionally counterfactual to the perceptual evidence (such as the presence of occluded physical objects). The results of the present study are consistent with the existence of such a mechanism of data completion/correction operating through the default use of prior expectations. Crucially, however, we further show that reliance on this mechanism also depends on the type of the intention to be inferred, according to its scope (motor vs. superordinate) or its target (non-social vs. social).