In April 2006, Shane Legg and Marcus Hutter suggested a formal measure of machine intelligence based on the idea that intelligence basically amounts to achieving complex goals within complex environments and that this idea can be formalized within the framework of algorithmic information theory. While not aiming at characterizing intelligent systems, Daniel Dennett’s paper “Intentional Systems” suggests something similar: A way to tell intentional systems from non-intentional ones by thinking about the behavior of systems in a way that has a normative, logical basis rather than an empirical one. Dennett’s method of comparing the actual behavior of a system to the most rational things to do, given some goals, constraints and information about the present state of affairs, sounds very much like a recipe for a general test of intentionality, if not intelligence.

Dennett introduces three levels of abstraction we can use to describe intentional systems: Descriptions on the level of physics, a functional perspective and the intentional stance. In order to determine in how far these levels presuppose optimal design and rational behavior, I am first going to explain what Dennett means when he talks about intentional systems, then describe each of the three different levels of abstraction and finally analyze what role the notions of optimality and rationality play for each of them.

When Dennett talks about intentionality, what he refers to is a property of linguistic entities. An entity is intentional if it is not possible to replace part of its description with terms that describe the same physical object and not to lose part of what the entity means. Intentional systems are systems with states that “can be directed on something”, systems that have states that are “about something”.

Obviously, the most basic perspective of looking at a system in the physical world, be it intentional or not, is the physical stance. We base our description of a system solely on the actual physical state of a system and predict its behavior exclusively by applying our knowledge about the laws of nature to this state. Such a low-level stance makes it possible to achieve a level of perfection in our predictions that no other stance allows — only the physical point of view allows us to know beforehand when a system will break down and stop functioning as intended. There is no need to know anything about how close the behavior of the system is to optimality or whether it makes sense to even think about the system in terms of rationality. On the other hand, such a stance can’t tell us much about systems that are not really simple or simplified. Even in predicting breakdown, the causes have to be easily locatable, otherwise the sheer complexity of the system we want to explain will make it impossible to predict anything from a physical point of view.

As systems get more complex, it makes more and more sense to adapt the design stance. Thinking about how a system is designed and which functions it fulfills enables us to predict behavior in a way that is computationally much more economical. There is no need to know how the electrons within the central processing unit of my computer move in order for me to tell that typing a letter will make it appear on the screen. There is just one condition: My computer must not break down this very second, the system we are looking at needs to work as designed. Just like any other predictions on a non-physical level, design stance predictions can’t foresee malfunctions since everything we take into account when thinking about the behavior of the system is the functional design — there is no need to assume that the actual physical condition is part of our knowledge about the system. Every system can be divided into different smaller or larger functional parts. In order to make functional predictions, the first thing one needs to do is to establish the level of abstraction of the individual functional parts. How small are the functional elements that we need to look at in order to make meaningful predictions? For really complex systems, there is no clear-cut answer to this question. Since the human brain is said to be the most complex system in the known universe, we might need to adapt a different, more coarse stance.

Looking at a system from the intentional stance implies that we ascribe to the system the possession of certain information, assume that the system acts in the direction of certain goals and that it always follows the most reasonable action relative to the information and goals the system has. Call the information possessed by the system “beliefs”, call the goals “desires” and the level of abstraction of the intentional stance matches the level we use when we talk about behavior in our everyday life. There is no need to assume that a system that possesses information really believes, but it is nonetheless striking how well these two notions match. The intentional stance not only assumes that a system will function as designed as does the design stance, but also that the design is optimal and that the system therefore always chooses the most rational action relative to its goals. Therefore, whenever we can suppose that it is valid to assume that a system has optimal design, we are justified in assuming the intentional stance as it may well be impossible to apply either the design or the physical stance in a computationally feasible way. Natural selection is an optimization process that molds systems in the direction of optimal design. When we try to find out how close a system is to being optimally designed, we need to take into account the extent of the information the system has and how its actions are constrained.

If a system turns out to be unpredictable from an intentional stance, e.g. because it acts in a way that defies all rules of rationality and logic, we can still collect data about the response pattern of the system and adapt a design stance once we have collected enough data to draw some conclusions about the functional design of the system. Dennett himself makes clear that one can’t stay one the level of “belief” in order to explain the “actual, empirical” behavior of a believer — one needs to descend one level and look at the believer from a functional perspective.

According to Dennett, rationality is optimal design relative to a goal or to a hierarchy of goals and a set of constraints. When we predict intentional systems, what we are doing is no more than to predict what the most rational system would do, given the same goals, constraints and the same information about the state of affairs as the system we want to predict. The assumption of rationality entails a few other beliefs about the system that we suppose is rational. We can assume that, usually, it does not desire its own destruction and that it follows the truths of logic, since assuming that the system has some beliefs but acts contrary to what follows from these beliefs would violate basic rules of rationality. Systems in the real world are not perfectly rational, therefore not all logical truths appear in those systems, but, in order to make predictions from an intentional stance, we just assume perfect rationality and hope that the drawbacks are smaller than those associated with any other stance.

Dennett introduces the concept of an intentional system in order to connect the intentional domain to the non-intentional domain of the physical sciences. It is quite obvious that he achieves this goal. From a pragmatic point of view, our beliefs can be identified with the information we possess about the world, our desires are our goals and a large part of our behavior can be described by assuming that we approximate an optimal rational agent trying to achieve its goals within certain informational and physical constraints, a description that completely eliminates intentional wording. Dennett’s article does not explain intelligence, and neither does Dennett aim at a complete theory of mind with this article. What he does talk about is a way of thinking about systems, both intelligent and non-intelligent, that eliminates non-scientific vocabulary and that, in the end, might allow us to talk about and measure intelligence in a way that is both more formal and much more general than the Turing test or any known IQ test.