I’ve finally gotten around to reading this famous book, and I have to say that its fame is quite merited. This is one of the best examples I’ve come across of philosophy interacting with its neighboring scientific disciplines. Unlike say, Dennett, however, Dreyfus writes not in the perspective of a fellow researcher who tries to perform some philosophical examination of a thorny problem within the discipline, but rather in the perspective of an outsider who wants to rectify the outsized claims of researchers in a fledgling field. So it’s easy to understand why he was so excoriated by AI people in the past.
However, the book is not actually as negative as one would think, especially in light of the fact that for Dreyfus the failure of AI is nothing less than an empirical refutation of a philosophical doctrine that is deeply entrenched in the Western tradition: namely, the epistemological assumption that everything can be described by formal, context-free rules, and the ontological assumption that the world consists of independent, context-free facts1. In fact, he attempts to categorize intelligent behavior to pinpoint the types of tasks that AI would be successful in (from Type I tasks that involve only associationistic reasoning, up to Type III tasks that involve formal rules that in principle can be performed by a computer but whose complexity requires the use of heuristics), and of course to pinpoint also the types of tasks with which it would not fare so well (Type IV tasks that require nonformal reasoning and context sensitivity, such as understanding natural language).
While his criticisms of the symbolist paradigm is spot on, I am less enthusiastic about the alternative paradigm that he outlines in Part III. Probably this stems from the fact that I don’t have much of a background in phenomenology, so the reading in this section was harder going. As a first impression, I can see why Edward Feigenbaum would call this “cotton candy”---but for different reasons. “To the things themselves,” as Husserl says, means the attempt to theorize about everyday reasoning experience; but at the same time the force of Dreyfus’s arguments against symbolist AI adumbrate the very impossibility of this task. So there seems to be a self-defeating air about this part of the book. But at the very least, I think it’s a better foundation for understanding human cognition than symbolist AI.
It is interesting to read this book over forty years after it was first published. Of course AI as a field has changed quite a bit since that time, so Dreyfus’s text has some glaring omissions, the most important of which is the discussion of machine learning. I am not sure how well Dreyfus’s arguments would extend to this subfield of AI. On one hand, it does not share the epistemic and ontological assumptions of symbolist AI. Machine learning algorithms need not assume that the regularities of the world can be captured in or, more strongly, consists of independent facts whose relations can be captured in context-free rules; they only assume that the world has regularities at all, which can be studied as a black box using probability and statistics. However, my hunch is that Dreyfus’s assessment of machine learning would be similarly negative as of his assessment of symbolist AI unless it integrated an “embodied” approach, which would require insights from robotics2.
Footnotes
-
It is easiest to see the force of Dreyfus’s criticisms in light of, say, the logical positivists. But his claim that the entire tradition, from Socrates up to (but not including) Wittgenstein and Heidegger is party to these assumptions seems to me a stretch. However, I think this is not really crucial to Dreyfus’s central points, as he only connects the philosophical underpinnings of symbolist AI with these assumptions to draw a contrast with the alternative “embodied” paradigm. ↩
-
His negative remarks concerning Watson, the AI that won Jeopardy! against Ken Jennings, seem to vindicate this hunch. See his piece on the Times here, where he constrasts Watson with the more embodied approach by the “subsumption architecture” that Rodney Brooks pioneered in robotics. ↩