1.3. Overview of the present study

To test this hypothesis, we conducted four experiments in which participants were requested to infer one intention underlying an action scene. Each experiment involved one type of intention with a specific scope (motor vs. superordinate) and a specific target (social vs. non-social). Interactions between prior expectations and visual information were examined within a Bayesian probabilistic framework. This conceptual framework is particularly well-suited to account for how accurate predictions on hidden world states are made in situations where available sensory information does not sufficiently constrain the number of potential solutions (Baker et al., 2006; Körding & Wolpert, 2006; Griffiths et al., 2008). Before the onset of an action sequence, each of the agent’s possible intentions is first assigned a certain ‘level of belief’, termed a priori probability (the probability that intention X is the real cause of the observed behaviour estimated from past experiences). Then the observer progressively gathers sensory information (visual input) as the action sequence is disclosed and both sources of information (sensory and a priori) are combined to infer the intention that motivates the observed behaviour. Thus, the process by which intentions are inferred is considered as reflecting a trade-off between the sensory information and the prior probability of each candidate intention (Baker et al., 2007). Finally, the chosen intention is that which maximizes the posterior probability value, i.e. the probability that intention X is true given what is observed.

In the present study, these two terms – a priori probability and sensory information – were manipulated using a two-steps procedure:

(i) Prior expectations the participants had about the agent’s possible intentions were manipulated by increasing the a priori probability that one intention (termed biased or preferred intention) occurs, at the expense of other intentions (non-biased or non-preferred intentions) with the same scope and target.

(ii) Sensory information available from an action scene was then manipulated in a second step by modulating the degree of completeness (i.e. the duration) of the action sequences, resulting in actions scenes with varying amount of visual information.

We first predicted that judgements on intentions should follow the general principles of Bayesian inference. Specifically, we expected that the amount of visual information would interact with participants’ prior expectations such that the lower the reliability of the external visual input, the more participants’ responses should depend upon their own internal expectations. That is, they should respond more frequently in the direction of the bias (i.e. the preferred intention) towards which they were assigned. And vice versa, the higher the amount of visual information, the less the participants should rely on their prior expectations.

Second, we predicted that the shape of the interaction between these two sources of information should be a function of the type of intention, depending on both its scope and target. Along the dimension of the scope, we hypothesised that participants’ judgement about motor intentions should primarily rely on sensory information available from the action scene. This prediction is motivated by the pragmatic content of the motor intention: “grasping a glass of water” directly denotes the corresponding intention of “grasping that glass”. In this case, perceiving the action itself – i.e. processing the associated visual kinematics – is enough to successfully determine the nature of the underlying intention (Rizzolatti et al., 2001, 2004). On the other hand, we expected performance in judging superordinate intentions to be significantly influenced by participants’ prior expectations. As already mentioned, the same sequence of motor acts can be part of combinations aimed at different general goals or superordinate intentions. In this specific case, sensory information carried by kinematic movement is not sufficient to infer the corresponding intention, as it under-constraints the set of candidate intentions congruent with this movement (Jacob & Jeannerod, 2005; Kilner et al., 2007a, 2007b; Jacob, 2008). We consequently predicted that this perceptual uncertainty should encourage participants to « mistrust » what they observe and, hence, to rely more on their prior expectations.

Along the dimension of target, finally, we expected participants' reliance on their prior expectations to increase when motor and superordinate intentions are directed at another agent. The structure of social interaction meets indeed particular, often irrepressible, expectations, such as those provided by reputational knowledge (King-Casas et al., 2005; Singer et al.,2006; Fu & Lee, 2007). Indeed, knowledge about individual’s reputation has been robustly demonstrated to have a strong impact on predicting how the observed agent will behave (Frith & Frith, 2006).In line with other recent suggestions, we thus hypothesised that the weight of these a priori expectations should increase when the observed action fits into a context of social interaction.