Reply to Oron Shagrir
Oron Shagrir argues that computation and content are more interconnected than my account allows. Computational individuation, he claims, is subject to indeterminacy, and at least in some cases the indeterminacy is resolved by appeal to content. If content is a gloss, as I argue, then according to Shagrir computational characterization is partly a gloss too.
Shagrir deploys an example to make the case for computational indeterminacy. Figures 2 and 3 describe the behavior of system S under two different scenarios. In the scenario depicted in figure 2, S causes the arm to move if either or both sensors are stimulated, while there is no movement if neither sensor is stimulated. We would, he says, describe S in this first scenario as an OR gate. In the scenario depicted in figure 3, S causes the arm to move only if both sensors are stimulated; otherwise there is no movement. We would, he says, describe S in this second scenario as an AND gate. According to Shagrir it is the different interpretations – OR in scenario 1 and AND in scenario 2 – that explain the different movement capacities.
Let’s take a closer look at the example. We are told that S has the same causal organization in the two scenarios; it is the same physical system. The difference that determines the different behavior and hence the different interpretation in the two scenarios lies in how S responds to different input voltages. In the 1st scenario (figure 2), S receives high input voltages (h: between 5 and 10 V) and the stimulation of a single sensor is sufficient for the arm to move. In the 2nd scenario (figure 3), S receives low input voltages (l: <5) and both sensors must be stimulated for the arm to move.
Shagrir characterizes the difference that determines and explains S’s behavior in the 2 scenarios as a representational difference:
… specific content (carried by the voltages) determines the correct realization and interpretation functions. In our example, assume that S’s inputs represent pressure contact, or absence of contact, on the arm’s grasping surface. In the first scenario, input1’s h-values represent pressure on the surface’s left side, and input2’s h-values represent pressure on the surface’s right side (the low-values signify their absence). The output’s h-values represent pressure either on the left side or right side (or both). These representational contents would account for interpreting S as computing OR. In the second scenario, the l-values represent pressure on the surface’s left side (input1) and right side (input2), and the output’s l-values represent pressure both on the left side and right side. These representational contents would account for interpreting S as computing AND.
The difference in S’s behavior in the 2 scenario is determined by S’s response to the different input voltages it receives in the two scenarios. As Shagrir points out, the facts that determine how S reacts are external facts but crucially here, and contrary to what Shagrir says, (i) how S reacts in each scenario, given its inputs, is an internal, not an external, fact about S, determined by its causal organization, and (ii) the external facts that determine how S reacts are not semantic facts, i.e., they are not facts about what represents what. They are straightforwardly causal facts, facts about how a difference in voltage to S’s sensors affects its behavior. These external, causal facts may support a representational construal of the device (what I call a ‘gloss’), but any representational content that they support is not what explains the device’s behavior nor what determines its computational identity. Yes, S can be interpreted as an OR-gate under certain input voltages, and an AND-gate under others, but once one specifies S’s behavior under different voltage inputs, there is nothing indeterminate about S’s computational description, certainly nothing that an appeal to semantic facts would be needed to resolve.
In conclusion, there is no reason to think that content affects the identification of computational structures or that computational vehicles are subject to any indeterminacy that requires appeal to content to resolve.