Bratman (1987) considers future-directed intentions as the central case of intending to act and contrasts this approach to intention with an alternative approach that gives priority to immediate intentions or intentions in action. He notes that this second approach naturally leads to the idea that intentions in action reduce to complexes of beliefs and desires, i.e., that what makes it the case that an agent acts with a certain intentions are simply facts about the relation between the agent’s actions and his beliefs and desires, and that this in turn tempts us into thinking that the same reductive strategy can be extended to future-directed intentions..[2] Focusing instead on future-directed intentions as the central case of intending allows us to identify functions of intentions that cannot easily be accommodated within a belief-desire model and thus makes the reductive strategy much less appealing. This would account for Bratman’s emphasis on the deliberative and coordination functions of intentions. The flip side of the coin, however, is that present-directed intentions are then seen as little more than transmission belts needed to convey the motivational force of future-directed intentions. As noted by Velleman, this leaves us with a potentially large class of actions where present-directed intentions appear to have no role to play, namely all these actions that are intentional yet not preceded by future-directed intentions. What belief-desire reductive approaches, Bratman’s account and Velleman’s account all seem to overlook is a specific pragmatic function of intentions in action or present-directed intentions, namely their role in the guidance, control and monitoring of action execution.
Harry Frankfurt (1978) was one of the first philosophers to criticize this oversight and insist on the importance of this pragmatic function of intentions. He emphasized that “a person must be in some particular relation to the movements of his body during the period of time in which he is presumed to be performing an action” (Frankfurt 1978, p. 157) and characterized this relation as one of guidance. Other philosophers have since shared his insight. For instance, Brand (1984), Bishop (1989) and Mele (1992) all insist that an adequate account of intentions should incorporate the guiding and monitoring roles of intentions in order to properly capture the close and continuous connection between intention and ongoing action.
The main reason why this connection between intention and ongoing action is needed is that human agents are neither infallible nor omniscient. Their expectations about the circumstances in which the action is to take place may not always be correct and they may fail to anticipate some of the relevant aspects of the situation of action. In other words, their situational beliefs may be incorrect or incomplete. The same goes for their instrumental beliefs. Suppose, for instance, that I intend to visit a colleague in her office. I may be wrong in thinking that this is the door to her office (incorrect situational belief) or unsure which door is her office door (lack of relevant situational belief). Similarly, I may also be wrong in thinking that I should pull the door to open it (incorrect instrumental belief) or unsure whether to push or pull (lack of relevant instrumental belief). If intentions are to reliably produce behavior matching their representational content (e.g., visiting my colleague in her office), they should have some flexibility and incorporate monitoring processes to detect deviations that jeopardize the success of the action and correction processes to trigger compensatory activity.
This emphasis on control finds a strong echo in the literature on motor cognition (see, e.g., Jeannerod 1997, 2006). Indeed, it is in this literature that we can find the most precise characterization of the monitoring and control functions of intentions and of the mechanisms that support them. According to the very influential internal model theory of motor control, motor control strategies are based on the coupling of two types of internal models: inverse models and forward models (Frith et al. 2000; Jordan & Wolpert 1999; Wolpert 1997). Inverse models compute the motor commands needed for achieving a desired state given the current state of the system and of the environment. An efference copy of these commands is fed to forward models, whose role is to make predictions about the consequences of the execution of these commands. The control of action is thought to depend on the coupling of inverse and forward models through a series of comparators: error signals arising from the comparison of desired, predicted, and actual states (monitoring) are used for various kinds of regulation (control). In particular, they can be used to correct and adjust the ongoing action in the face of perturbations, as well as to update both inverse and forward models to improve their future functioning.
Recent experimental work in motor cognition also suggests, however, that much of action control is automatic and proceeds independently of conscious awareness. For instance, in an experiment (Castiello et al. 1991) participants were asked to reach for and grasp a target as quickly as possible and their hand trajectories were recorded. On some trials, though, the target shifted position after the movement had started. When this happened, participants were instructed to correct their movement in order to reach accurately for the target and to signal the time at which they became aware of its displacement by shouting "Tah!". The experiment showed that the participants started correcting their movements more than 300ms before they signaled awareness of the target displacement. A subsequent study (Pisella et al. 2000) was especially instructive. In a first experiment they used a similar paradigm but introduced a condition where participants were requested to interrupt their movement when the target changed location. Despite the instruction, the participants could not prevent themselves from correcting their movements instead of stopping for a good 200 ms. In contrast, however, in a second experiment green and red targets were presented simultaneously in the two positions and the participants' task was to point at the green one. On some trials, the color of the two targets could be unexpectedly interchanged at movement onset. When this happened, one group of participants was instructed to interrupt their ongoing movement and the other group to correct it. In contrast to what happened in the first experiment, no automatic corrective movements were observed in the group instructed to interrupt their movement and in the other group corrections involved a significant increase in movement time. Thus, these results suggest that while corrections made in response to spatial perturbations are under automatic control, corrections in response to chromatic perturbations require intentional control.
On the one hand, the mere fact that some or much of action control can be automatic is not a sufficient reason to deny a control function to intentions. The experimental studies presented in the previous paragraph suggest that action control can indeed operate automatically and outside of conscious awareness and that when there is a conflict between automatic and intentional control, automatic control may take precedence over intentional control. Yet, they also provide evidence that some corrections cannot be carried out automatically but depend on intentional control. On the other hand, the mere fact that intentional control seems needed to compensate for chromatic perturbations may not provide sufficient ground for considering that the intentional control of action execution is a central function of intentions. One would want a more systematic account of the respective roles of automatic and intentional control. Recent developments of the internal model approach to motor control may constitute a useful guide.
While the internal model approach to motor control was initially introduced to account for fine-grained aspects of motor control, more recent versions of this approach emphasize the hierarchical nature of motor control (Hamilton & Grafton 2007; Jeannerod 1997; Kilner et al. 2007). They propose that internal inverse and forward models are arranged in a hierarchy and that error signals generated at one level of the hierarchy can propagate to the next level when correction mechanisms at this level are not able to make the necessary compensations. I have suggested elsewhere (Pacherie 2008) that one can distinguish three broad levels in an action specification hierarchy. At the highest level, action representations represent the whole action as a unit, in terms of its overarching goal and of the sequence of steps or subgoals needed to achieve that goal. At this level, the action may still be represented in a rather abstract format. The second level is concerned with the implementation of each step in the action plan and involves selecting an appropriate motor program given the immediate goal and contextual information about the current state of the agent and the current state of its environment. In other words, processes at this level are in charge of anchoring the successive steps of the action plan in the current situation and of selecting appropriate motor programs. Finally, once a motor program has been selected, the exact values of its parameters must still be set. This is done at the third level, where incoming sensory information about external constraints is used to specify these values.
Acknowledging the existence of different levels of action control corresponding to these different levels in the action specification hierarchy may allow us to accommodate both automatic and intentional action control processes. As long as error signals can be reduced by automatic corrections made at lower levels in the hierarchy, there is no need for the intervention of intentional control. However, there are two classes of cases where automatic corrections may not be sufficient to put an action back on track. First, important external perturbations can lead to discrepancies that are too large to be automatically compensated. In such a case, error signals would propagate upwards, we would become aware of them and shift to a conscious, intentional compensation strategy. Second, in some instances there may also be discrepancies in the ways the action is or can be specified at different levels of the action representation hierarchy (inter-level representational misalignment). Thus, the study by Pisella and colleagues (Pisella et al. 2000) suggests that action specification at the sensorimotor level does not encode chromatic information and uses spatial information as a proxy for it. When chromatic information and spatial information vary independently, as they do in one of the conditions of the experiment, representations at different levels of the action representation hierarchy become misaligned and the intervention of conscious control becomes necessary to realign them.
Importantly, on this conception of intentional control and as Frankfurt had already noted, what is essential for actions to be intentionally controlled is not that intentional control processes actually affect their course, but that these control mechanisms would have intervened to adjust the action had the need arisen. In other words, an action may be intentionally controlled even though automatic rather than voluntary control mechanisms intervene to compensate for deviations, provided these voluntary control mechanisms would have kicked in, had automatic corrections proved insufficient.
Even more importantly, if action control is an essential function of intentions, then we should stop thinking of intentions as simply mental representations of goals somehow triggering motor processes that, if everything goes well, will yield the desired outcome. Rather, we should think of monitoring and control processes as intrinsic to intentions, that is, of intentions as encompassing not just representations of goals but also a specific set of monitoring and control processes organizing and structuring the motor processes that themselves generate movements.
In this section, I argued for the idea that the control of action execution is an important pragmatic function of intentions. Acknowledging the existence and importance of this function allows us to plug gaps in the creation myths considered earlier. First, it allows us to attribute a specific pragmatic function to present-directed intentions rather than considering them as mere transmission belts in charge of conveying the motivational force of future-directed intentions. We can thus assuage one of the main worries raised by Velleman against Bratman’s pragmatic account of intentions and the pragmatic creation myth derived from it. Second, Anscombe’s and Velleman’s accounts of intentions both assume that intentions reliably cause behavior that matches their representational content. Human agents, however, are neither infallible nor omniscient. Their situational and instrumental beliefs can be incorrect or they can lack situational and instrumental beliefs that are relevant to the successful execution of their intentions. Thus, the reliability demanded by Anscombe’s and Velleman’s accounts largely depends on our having powerful and flexible control processes allowing us to put our actions back on track when perturbations deviate their course.
One may agree that the conscious control of individual action is a function of intention in the sense that intentions have this causal role, but still be skeptical that this is the role intentions are designed for, or to put it in other words, that it is a teleofunction of intentions. Thus, one could argue that very large external perturbations are rare and that inter-level representational misalignment is the exception rather than the rule. If so, most of action control would be automatic anyway and intentional action control would play at best a marginal role. It would therefore be unlikely to confer on intention-forming creatures benefits important enough to warrant the claim that intentions are designed for action control. As I have tried to argue in this section, the benefits conferred by online conscious control over actions are not as negligible as this deflationary view implies. In addition, I think we can build a very strong case that conscious action control confers important benefits if we consider joint activities rather than just individual actions. Acting jointly demands that we solve coordination problems that do not arise (or arise only in a very attenuated form) in individual action. In what follows, I will argue that online conscious control plays a crucial role in solving these coordination problems. I will further speculate that conscious online control over actions might indeed have become established as the primary function of intentions because of the role it served in solving these coordination problems and because of the benefit this conferred on creatures capable of solving these coordination problems and thus of acting jointly in an efficient and flexible way.