Blocking the Exits

Blocking the Exits

Imagine two black boxes containing independently built automated encyclopedias. Their exchanges produce a reliable pattern of colored lights, yet investigators cannot explain the pattern from the code and signals they inspect. Dennett’s proposed explanation is that both systems represent the same world. Here he examines the details intended to rule out rival explanations—the exits a critic might take from his argument.

(1) The devices in boxes A and B are nothing but automated encyclopedias, not even “walking encyclopedias” but just “boxes of truths.” Nothing in the story presupposes or implies that these devices are conscious, or thinking things, or even agents, except in the same minimal sense that a thermostat is an agent. They are utterly boring intentional systems, rigidly fixed to fulfilling a single, simple goal. (The same is true of IBM’S Watson, of course.) They contain large numbers of true propositions and the inferential machinery necessary to generate more truths, and to test for “truth” by testing a candidate proposition against their existing databases.

(2) Since the two systems were created independently, they cannot plausibly be supposed to contain (actually or even virtually) exactly the same truths, but for the prank to work as well as I claim it did in the story, we must suppose a very large overlap, so that it was highly unlikely that a truth generated by A would not be recognized as such by B. Two considerations, I claim, make this plausible: (i) Al and Bo may live in different countries and have different native languages, but they inhabit the same world, and (ii) although there are kazillions of true propositions about that world (our world), the fact that both Al and Bo set out to create useful databases would guarantee a high degree of overlap between the two independently created systems. Although Al might know that on his twentieth birthday his left foot was closer to the North Pole than the South Pole, and Bo had not forgotten that his first French teacher was named Dupont, these would not be truths that either one would likely put in their respective databases. If you doubt that the mere fact that they were each intent on creating an internationally useful encyclopedia would ensure such a close correspondence between their respective databases, just add, as an inelegant detail, the convenient fact that during their years of hacking they compared notes as to topics to be covered.

(3) Why not have Al and Bob (a fellow American), or for that matter, why not simply have a duplicate of Al’s system in box B? Because it is critical to my story that no simple, feasibly discoverable syntactic matching-up could explain the regularity. That is why Bo’s system is in Swedish Lisp—to conceal from prying eyes the underlying semantic commonalities between the data structures consulted during A’s sentence-generation task and B’s sentence-translation-and-truth-testing task. Computers, as physical systems, must be, at best, syntactic engines, responding directly to physically transducible differences, not meanings. But both A and B have been designed to mirror as closely as possible the imaginary know-it-all, a semantic engine full of understood truths. When two different syntactic systems, A and B, have been designed to mirror the same semantic engine, the only way of accounting for the remarkable regularity they reveal is to ascend to the semantic-engine level, where truths are believed and assertions are intended. The idea, then, was to create two systems that exhibited the fascinating regularity of external behavior described but that were internally as different as possible, so that only the fact that their respective innards were systematic representations of a common world could explain the regularity. (Recall that this was the theme in chapter 13.)

We might pause to ask whether or not two such systems could ever be so inscrutable as to be invulnerable to reverse engineering. Or in other words, could the scientists have remained baffled for so long? Cryptography has moved into such rarefied and arcane regions that one should think thrice at least before declaring either way. I have no idea whether anybody can make a sound argument that there are unbreakable encryption schemes or that there aren’t. But encryption aside, hackers will appreciate that all the convenient comments and other signposts one places in the source code when composing a program vanish when the source code is compiled, leaving behind an almost impossible-to-decipher tangle of machine instructions. “Decompiling”—reverse engineering the object code and recovering the source code—is sometimes possible in practice (is it always possible in principle?), though it won’t restore the comments but just render salient the structures in the higher-level language. My assumption that the scientists’ efforts at decompiling the program and deciphering the databases came to naught could be strengthened by postulating encryption, if need be.

In the story as told, we can agree that it is bizarre that the scientists never thought of checking to see if there was an ASCII translation of the bit streams running through the wire. How could they be so dense? Fair enough: you can fix this flaw in the thought experiment by sending the whole gadget (boxes A and B, and the connecting wire) to “Mars,” and let the alien scientists there try to figure out the regularity. The fact that all α’s cause red lights, all β’s cause green lights, and random-bit strings cause amber lights will be just as visible to them as to us, but they will be clueless about ASCII. To them, this gift from outer space will exhibit an utterly mysterious regularity, totally beyond all analytic probes unless they hit on the idea that each box contains a description of a world, and that the descriptions are of the same world.[4] It is the fact that each box bears multifarious semantic relationships to the same things, though expressed in different “terminology” and differently axiomatized, that grounds the regularity.

When I tried this thought experiment out on Danny Hillis, the creator of the Connection Machine, a pioneering massively parallel computer built by his company, Thinking Machines, in the early 1980s, he thought immediately of a cryptographic “solution” to the puzzle, and then granted that my solution could be profitably viewed as a special case of his solution: “Al and Bo were using the world as a ‘one-time pad!’ ”—an apt allusion to a standard technique of encryption. You can see the point by imagining a variation. You and your best friend are about to be captured by hostile forces (space pirates, let’s say) who may know English but not much about your world. You both know Morse code, and hit on the following impromptu encryption scheme: for a dash, speak a true sentence; for a dot, speak a false sentence. Your captors are permitted to listen to you two speak: “Birds lay eggs, and toads fly. Chicago is a city, and my feet are not made of tin, and baseball is played in August,” you say, answering “no” (dash-dot; dash-dash-dash) to whatever your friend has just asked. The next time you need to say no, you use different sentences. Even if your captors know Morse code, unless they also can determine the truth and falsity of these sentences, they cannot detect the properties that stand for dot and dash. This variation could be added to our fable, for spice, as follows: Instead of shipping the computer systems in the boxes to Mars, we put Al and Bo in the boxes and ship them to Mars. The Martians will be as puzzled by Al and Bo, if they play the Morse code prank, as by the computers, unless they draw the conclusion (obvious to us, but we’re not Martians) that these things in the boxes are to be semantically interpreted.

The point of the fable is simple. There is no substitute for the intentional stance; either you adopt it, and explain the pattern by finding the semantic-level facts, or you will forever be baffled by the regularity—the causal regularity—that is manifestly there.[5]

At this juncture, if you are like many philosophers, you may once again be attracted by the claim that this intuition pump “works” only because boxes A and B are artifacts whose intentionality, such as it is, is entirely derived and artifactual. The data structures in their memories get their reference (if they get any at all) by indirect reliance on the sense organs, life histories, and purposes of their creators, Al and Bo. The real source of the meaning or truth or semanticity in the artifacts lies in these human artificers. Al and Bo have original intentionality, and A and B have only derived intentionality. (That was, of course, the point of the suggestion that in a certain sense Al and Bo were in their respective boxes.) I might have told the story differently: inside the boxes were two robots, Al and Bo, each of which had spent a longish “lifetime” scurrying around the world gathering facts before getting in their respective boxes. I chose a simpler route, to forestall all the questions about whether box A or box B was “really thinking,” but if you want to reconsider the thought experiment with this complication, note that the intuition pump about the giant robot survival machine has already cast doubt on the otherwise compelling idea that genuine intentionality could not arise in any artifact.

[1] IBM’s Watson came along just in time to turn my science fiction into science fact, close enough. You could consider Watson to be in box A if you like, and box B to contain a Swedish alternative to Watson, independently developed by Bo. When I first published this thought experiment (Dennett, 1995a), the best I could say was this: “For a real-world example of such a project, see Douglas Lenat’s enormous CYC (short for ‘encyclopedia’) project at MCC (Lenat and Guha, 1990).” See also the footnote on CYC in chapter 28. The approach to AI that Watson exemplifies was largely unimagined back in 1995, but it has been making great strides. Unlike CYC, which is largely hand-coded, Watson can graze unattended on the Internet for its own supply of truths, and make powerful use of statistical properties of the data available there. Both Watson and CYC in their different ways sorta understand the data in their data banks—much, much more than other computers with large databases.

[2] When I first wrote this intuition pump, ASCII (American Standard Code for Information Interchange) code was the standard format for almost all word-processing, e-mail, and Internet language. It has since been supplanted by an extended, backward-compatible format, UTF-8, which stands for Universal [Character-set] Transformation Format—8-bit. That is, ASCII is still a part of UTF-8.

[3] For philosophers only: Some have argued that my account of patterns in “Real Patterns” (1991b) is epiphenomenalism about content. This is my reply.

[4] Once the hypothesis occurs to them, the Martians can engage in a version of Quine’s radical translation, but their task will be made doubly difficult by not being able to query their “informants,” A and B, by holding up objects and asking “ball?” “pencil?” and the like.

[5] The same moral can be drawn about interpreting the historical facts of evolutionary history. Even if you can describe, in matchless micro-detail, every causal fact in the history of every giraffe who has ever lived, unless you go up a level or two, and ask, “Why?”—hunting for the reasons endorsed by Mother Nature—you will never be able to explain the manifest regularities—the fact that giraffes have come to have long necks, for instance. (More on this in part VI.)