Prinz (2007) supplies a concise statement of his view, captured here in the following quotation:
The feeling of agency could be explained by a kind of prediction that the brain makes when we are about to act. If you elect to move your arm, you will be able to anticipate its movement. According to some leading neurobiological theories, when a plan is generated in the premotor cortex, a representation is sent to the somatosensory cortex corresponding to what the bodily senses should perceive when that action is executed. That representation is called a “forward model.” A forward model is an anticipatory somatosensory image. When our bodies carry out motor plans, the forward model is compared with the actual changes that take place in our body as we move. The feeling of agency may arise from this matching process. If a match occurs, we feel we are in control. If a match doesn’t occur, it’s because our bodies didn’t move as we predicted they would, and that results in an experience of being passively moved by an external source. (p. 342).
One way of appreciating a problem with Prinz’s view involves the way it combines a concept from control theory, that of a forward model, with the concept of a sensory image. That forward models are involved in the control of bodily movement is a highly plausible suggestion. That they be regarded as sensory images is somewhat less plausible. Before further fleshing out the problem, a bit more needs to be said about the distinct notions of a forward model and a sensory image.
Many philosophers are aware of control theory via the work of Rick Grush (e.g. (Grush, 2001)) and I here rely on his exposition of its basic ideas. In the simplest kind of control system, open-loop control, a desired goal signal is fed into a controller, which sends control signals to a target system or plant. Applying these concepts to motor control involves viewing parts of the musculo-skelatal system as plants and neural systems generating motor commands as controllers. The controller implements a mapping, the inverse mapping, of goal states onto command sequences. The plant implements a mapping, the forward mapping, of command sequences onto goal states (Grush, 2001, pp. 352-353). A slightly more complex control system, closed-loop control, has all of the components as in open-loop control plus the addition of feedback signals from the plant to the controller. While for many control purposes, closed-loop control is superior to open-loop control, closed-loop control is not without certain problems. If, for example, there are significant delays in the receipt of the feedback signal due to slow signal speeds and/or a relatively distant plant, then the system can oscillate wildly through potentially destructive cycles of overshooting and overcompensation. A slightly more complex control system that potentially overcomes such problems is pseudo-closed-loop control. One way of conceiving of pseudo-closed loop control is by thinking of it as built by adding features to open-loop control. The first addition involves a second signal being sent by the controller, an efferent copy, which is a duplicate of the signal sent to the plant. This duplicate signal, however, is not sent to the plant, but instead to an emulator or forward model, which, in turn sends signals back to the controller.
Now, it is tempting to follow Grush in calling the signal from the forward model “mock sensory information about what the real target system would do under various conditions” (p. 356 emphasis added), but I will want to resist such temptation.
It is useful here to consider the following two questions. First, what is involved in something’s being sensory in the sense of the term relevant to the current discussion? Second, do we have adequate reason for thinking that a forward model is relevantly sensory?
Starting with the first question, it is useful to look at Prinz’s own account of what makes something sensory. Prinz writes:
I will define a perceptually conscious mental state as a mental state that is couched in a perceptual format. A perceptual format is a representational system that is proprietary to a sense modality. To say that phenomenal states are perceptual is to say that their representational vehicles always belong to one of the senses: touch, vision, audition, olfaction, and so on. (J. Prinz, 2007, p. 336)
Further elaboration comes from what Prinz takes the negative aspects of his key thesis to be: “We do not have conscious states couched in non-perceptual formats. If I am right, we never have conscious states in our motor systems, and no conscious experiences are constituted by amodal representations…” (J. Prinz, 2007, p. 337).
In an earlier work dedicated to elaborating Prinz’s brand of empiricism, he spells out his view that “the senses are dedicated input systems” (J. J. Prinz, 2002, p. 115). Crucial to Prinz’s characterization is that each sense has both a proprietary class of inputs (physical magnitudes) and a proprietary representational format (thus denying that separate senses share a ‘common code’ (J. J. Prinz, 2002, p. 117).)
It is worth noting that in this earlier work Prinz endorses a view of imagery whereby “we can form mental images by willfully reactivating our input systems” (J. J. Prinz, 2002, p. 115). It seems natural to suppose that what is responsible for these reactivations counting as sensory imagery is that it is input systems that are reactivated.
With these remarks about what the “sensory” in “sensory imagery” consists in, let us return to the question of whether forward models need be conceived of as sensory imagery. In the basic outlines of pseudo-closed loop control, there is nothing that makes compulsory a sensory-imagery interpretation of the forward model. The forward model is not receiving sensory inputs and thus cannot count as a sensory system as characterized by Prinz. A fortiori, it cannot count as sensory imagery since it does not count as the reactivation of an input system.
Of course, it should be noted that there may be alternate architectures that incorporate forward models satisfying criteria for being sensory. However, the core idea of a forward model does not alone satisfy such criteria. It is also worth noting that the characterization of imagery as the willful reactivation of input systems threatens to make the imagery account collapse into a kind of non-sensory view. This is so if a crucial part of a state’s being imagery is its activation of a control signal.