Block Head Argument

          The Blockhead is a theoretical computer program that simulates intelligence by being programmed to respond to a vast but finite array of inputs (in the form of a conversation, in the case of the Turing Test) with an appropriate preprogrammed response. Any conversation must fall within a finite (although huge) possibility of logically, and grammatically formed sentences based on a finite number of words, thus the program has a complex look-up tree of responses to follow for any input it could possibly receive in an allotted time, simulating a seemingly intelligent conversation.

             The purpose of the Blockhead thought-experiment is to show that the functionalist conception of intelligence being based on behavior is not plausible. That is, there is more to intelligence than merely acting intelligently. The Blockhead may pass the Turing Test for behavioral intelligence, but it has the intelligence of a toaster, argues Ned Block (albeit a very complex toaster). Although the Blockhead replies to input with seemingly intelligible responses, it lacks understanding, which appears to be a fundamental constituent of intelligence. The machine is merely computing its response by means of a complex response-tree that it has pre-programmed. Computation is not understanding. Therefore, the Blockhead neither understands the input nor its own output. This cannot be intelligence, or at least in the commonsensical way we use the term intelligence.

            The basis of the Blockhead argument lies in the claim that intelligence does not supervene on behavior, behavioral disposition, or behavioral capacity. Therefore, two beings may behave exactly alike physically, yet intellectually dissimilar. One may be highly intelligent, while the other has no intelligence whatsoever. By programming a machine with the Blockhead program, a computer is designed only to behave intelligently, perhaps fooling onlookers into thinking there lays some genuine intelligence behind the behavior. However, the program is based on simple algorithms and lacks any genuine cognition. Some functionalists may argue that although the program and machine may not independently be intelligent or understand what it is doing, the system as a whole is intelligent, as exemplified through its seemingly intelligent behavior. However, being a mere conduit of information is in no way analogous to thinking. All of the Blockhead’s possible responses are preprogrammed, or thought out in advanced by computer programmers. So the system as a whole is never doing any real thinking, but merely retrieving preformed responses. It is crucial to understand that the Blockhead lacks any real cognitive capacity and is merely computing based on a simple input-output response tree.

            The Blockhead thought-experiment confronts another issue with the behaviorist conception of intelligence, namely the overlooking of causality on the behaviorist’s part. Being intelligent depends on having thoughts that evolve or are caused in the right way, not simply behaving intelligently. Present thoughts should be caused by past thoughts, and should later cause future thoughts. However, the Blockhead merely processes each stimuli input independently, retrieving an appropriate output from its look-up tree of preformed responses. The Blockhead appears to be static, with no causal dependency on its earlier mental states (for it has none). It simply depends on a stock of responses that have been written down in advanced, to generate the appearance of intelligent behavior. Braddon-Mitchell and Jackson give a more complex version of the Blockhead scenario to illustrate this problem with causality. We are to imagine that a Blockhead clone of Jones is created – a physical duplicate – with a complex look-up tree based on Jones’s mental states necessitating all the output responses to the possible physical inputs one could receive in a lifetime (a larger but still finite number). Because the clone’s look-up tree is based off of Jones’s mentality, they behave exactly alike in every situation they find themselves in. (This intuition pump assumes that it is possible to create a look-up tree modeled perfectly after someone’s mental states.) Moreover, they also would respond exactly the same to every situation they could possibly find themselves in. So their behavior is identical, however, their intelligence is not. The Blockhead twin has no intelligence and is merely an automaton following a preset path based on computing input with the appropriate output on his look-up tree. While Jones’ behavior is based on earlier mental states, perhaps being caused by earlier physical events, the Blockhead is merely processing each input stimuli independently and retrieving its appropriate response with no mentality or choice in the matter. We could say Jones’s behavior is caused in the right way, while the Blockhead is merely caused by mindless computational algorithms. But genuine thinking is much different; what we think depends on previous thoughts in complex ways for it to count as rational thought.

             Although it may not be physically possible to construct a Blockhead machine, this does not discount the practicality in conceiving what such a machine says about intelligence. The string of responses the Blockhead would have to choose from for output is too vast to exist, and of course could never be processed in real time. But such a machine is certainly logically possible. Because the argument is attacking a conception (namely the behaviorist conception) of what intelligence is; a conception of the Blockhead machine will suffice as a response. The Blockhead argument is merely a thought-experiment or intuition pump that serves to challenge the behaviorist conception of intelligence. It need not be manifested to be valid. In all practicality, it makes no difference whether such a machine exists insofar as it would not be intelligent even if it did. It would simply be like any other computer, just running an extremely expansive program. But the sheer volume of the response-tree, if the Blockhead somehow were constructed, would not be enough to generate real thinking. So what matters is that the thought experiment is logically feasible.

            The relevance of the impossibility of a Blockhead machine being constructed perhaps even strengthens the premise that intelligence does not supervene on behavior. Human beings clearly behave intelligently without having all of their outputs preformed in response-trees. It is also physically impossible for a human to be implementing such a ‘program’. Yet somehow the human capacity does appear to be as seemingly infinite as the Blockhead would be. This is because humans go through logical cognition. That is, humans process stimulus input based on past thoughts and experiences, and then generate output that is causally determined from a complex network of these sources. Human intelligence then appears to be more than mere computation of input-output like the Blockhead. Therein lies the distinction between a behaviorist conception of thinking and what we generally think of when we use the word thinking. Behavior (output) cannot be analyzed independently with respect to its input and deemed intelligent. Intelligence lies in the ability to relate early input and mental states. The vast impossibility of creating a machine that can even come close to mimicking intelligent human behavior may be indicative of the infinitely vast impossibility of creating a machine that genuinely thinks in the right way.

            I think most importantly, the Blockhead argument shows that a non-intelligent being can pass Turing’s Test. Thus it follows that Turing’s attempt to redefine intelligence in non-mentalistic terms was a failed project. This is not to say that mentality is necessary for intelligence, but rather that the attempt to separate the two by means of an imitation game was futile. There is more to intelligence than just observable behavior.

             Sophisticated behavior is not necessarily indicative of a sophisticated mind. The Blockhead could clearly pass the Turing Test, but no one is willing to redefine intelligence based on a mere computational program. The thinking humans do is far more complex than just retrieving information from a look-up tree of preformed stimuli response. If we had previously intuited that this is what it is to think, there would have been no need for a Turing Test in the first place. If humans were thought to merely compute input/output responses, it would have been agreed that all computers are thinkers.

            The Blockhead intuition pump also seems to indirectly say something about the importance of the internal mechanism of a thinking agent. A preprogrammed computer chip is not sufficient for thinking. A human brain is sufficient for thinking, however. It is not clear that a brain is necessary for a thinking being, but it will certainly suffice. A computer chip lacks the complex causal dependencies found in the brain. Braddon-Mitchell and Jackson define a node as the combination of input and output response at a given moment. The Blockhead, when stimulated with input responds with the one preformed response it has been programmed to respond with. Dealing with each input independently, machine only ever has one active node. This is what might be called a cheap trick, clearly not thinking. Humans, when stimulated with input, rely on various active nodes – past beliefs, thoughts, inclinations – all of which contribute to the output. Perhaps this causal dependency from complex nodal activity is precisely what gives rise to genuine thinking. So while we may just be a collection of tricks, the Blockhead is just one – and not a thinking one.

(NYU year 2 Minds and Machines)

Advertisement