Everyone knows something. Some people know a lot. But no human being knows as much, apparently, as Watson, the IBM computer that defeated the greatest champions of the television quiz show Jeopardy at their own game.
Faced with the prompt Aeolic, spoken in ancient times, was a dialect of this, Watson effortlessly answered Ancient Greek (or rather, using Jeopardy’s answers-as-questions format, What is Ancient Greek?). Confronted with Classic candy bar that’s a female Supreme Court justice, Watson shot back Baby Ruth Ginsburg. Very impressive. But does Watson understand what it’s talking about? Answering that question will point the way to the distinction between mere knowledge and true understanding.
When given This “insect” of a gangster was a real-life hit man for Murder Incorporated in the 1930s & ’40s, Watson answered James Cagney. Surely it knows that James Cagney was an actor, not a “real-life” gangster? And asked a question in the category U.S. Cities, Watson notoriously replied Toronto. So it knows that Aeolic is a kind of ancient Greek, but not that Toronto is in Canada?
As the commenters later explained, Watson does not figure out the answers to these questions in the way that humans do. Whereas we summon up a list of gangsters or U.S. cities and then ask ourselves whether they meet the other criteria explicitly imposed or implicitly suggested by the clue, Watson consults a sophisticated table of statistical associations between words, extracted from a massive stock of written material, from newspaper reports through encyclopedia articles. Aeolic, in the few places in which it appears, is linked closely to Ancient Greek. The same goes for Baby Ruth and candy bar and of course Ruth Bader Ginsburg and Supreme Court justice; Watson is then clever enough to see the overlap. Unfortunately, the same strong connection is found between James Cagney and various terms associated with organized crime. Watson has a great deal of information at its cybernetic fingertips about gangsterism, gangster movies, and the key figures in both, but it seems not to understand what it means to be a real-life gangster.
And how could it? How could a table of statistical associations comprehend the difference between fact and fiction, between murder and make-believe?
Watson knows a lot of things—assuming that knowledge is a matter of fast, reliable, retrieval of the facts. But it has little or no understanding of the things that it knows. It knows that Aeolic is a dialect of Ancient Greek, but it does not know what it is to be a dialect, or even—in spite of the fact that it is connected to the outside world only by words—what it is to be a language. It can talk (at least in the context of a quiz show) as informatively and as accurately as some of the most knowledgeable people on the planet, but its grasp on the facts it conveys is even less certain than that of a cocktail party habitué who has read all the latest reviews but has never glanced at a page of the books themselves.
What is Watson missing? I have given it a name—understanding. But what is that? There are two kinds: understanding language and understanding the world. Consider this sentence: Δέδυκε μεν ἀ σελάννα. You won’t understand it unless you read Aeolic Greek. But if I tell you it means “The moon has set”, you grasp immediately what the sentence is about: you know what it is for the moon to disappear below the horizon. You do not understand the sentence (unless you speak Aeolic), but only because you do not understand the language, not because you fail to understand its subject matter.
Watson’s problem is that it does not understand the world. (Some philosophers would argue that it does not genuinely understand language either, but I put that question aside.) It gives answers, but it has no grasp of what makes its answers correct.
You do not have to be a computer to find yourself knowing without understanding. There are some facts that real flesh and blood people know only in a Watson-like way. Perhaps you know that Bach wrote fugues, but you don’t understand what a fugue is—the more so, perhaps, if you are tone deaf. Another case: when I was a boy, intoxicated by science, I knew—in a trivia-contest-winning sort of way—that the hydrogen and oxygen molecules in water were held together by covalent bonds, but covalent was not much more to me than a glamorously technical word.
A little later I learned that a single electron could be in a superposition of two different places at once. All physicists know this, yet arguably, no one yet really understands what it means. We have the sentences, or mathematical formulae, to represent superposition, but we don’t know what, deep down, these sentences are talking about. And wouldn’t we love to? Knowledge is good, but isn’t understanding much better still?
I am, however, supposed to be analyzing, not acclaiming, understanding. What is it that Watson does not grasp about the movies, that the younger me did not grasp about covalent bonds, that no one perhaps grasps about quantum superposition?
One way to answer this question is to ask how we might distinguish facts that are truly understood from facts that are merely known Watson-style. It seems easy to make such distinctions from the inside, about our own knowledge. My bad conscience as a young boy told me that I was only feigning chemical expertise; the non-musical know-it-all is well aware that they don’t really understand what a fugue is.
We all know, by contrast, that we have a grip on the setting of the moon. Close your eyes and imagine: the familiar orb drifts steadily downward, is sucked into the horizon, is gone. Or think: as the Earth turns on its axis we stationary observers on its surface speed toward, then away from the moon; eventually our planet comes to occlude it entirely. Or feel: the setting of the moon measures the passing of time, the distance traveled by a departing lover, the passing of life.
Watson misses all of that, you might suppose. But so what? Watson has many peculiarities: it is blind, has no lovers, and is theoretically immortal. Surely none of this, however, stands in the way of understanding. We may relate to the moon through our senses and emotions, but might not other beings take a different but no less profound approach?
A different test for understanding investigates abilities rather than internal imagery, feelings, and thoughts. It is easy to discover that something important is missing in the youthful Michael’s grasp of covalent bonds. Ask me to define “covalent”, and I would have faltered. Or better, ask the younger me to explain how to solve problems in quantum chemistry, or ask someone who has just read Bach’s Wikipedia entry but who has no interest in music to tell a fugue from a passacaglia. The gap in understanding emerges soon enough.
Watson will not crack so easily. Imagine a more versatile version of Watson, proficient in answering questions generally, not just on Jeopardy—exactly the kind of expert system that IBM is using its Watson technology to build. Such a system would have no trouble defining covalent, fugue, or any other term that you throw at it. Presumably, it might learn to solve problem sets in a science class or to classify works of music, using the same statistical techniques that work so well in Jeopardy to distinguish the right moves from the wrong moves.
Why does the machine seem all the same not to achieve understanding? One answer is that its expertise is parasitic: it learns the right moves by examining the moves already made in the vast body of text that its programmers supply. Arguably, though, most of us require a similar degree of assistance—most of what we know we learn from others, rather than by figuring it all out for ourselves. A deeper answer is that there is something about Watson’s statistical ways of knowing that is incompatible with understanding.
Watson and you both answer questions by seeing connections between things. But they are different kinds of connections. Watson picks up from things it reads that there is a correlation between a sphere’s rotating and a fixed point on its surface having a constantly changing view of the rest of the world. You grasp why this correlation exists, seeing the connection between the opacity of the Earth, light’s traveling in straight lines, and geometry of the sphere itself. For you the statistics are a byproduct of what really matters, the physical and causal relations between things and people and what they do and say. Grasping those relations is what understanding consists in. Watson lives in a world where there are no such relations: all it sees are statistics. It can predict a lot and so it can know a lot, but what it never grasps is why its predictions come true.
1. Could a machine ever understand things in the way that we do?
2. Is understanding a matter of having knowledge of certain special facts, such as causal facts? Or is it a matter of having a special kind of knowledge of facts: transparent, deep, luminous, or something like that?
3. Many scientists believe that, at bottom, our thought is implemented in neural networks that make statistical associations. Does that mean that we are no better than Watson? That our sense of understanding is an illusion?
The point of differentiating understanding and knowledge is to get a better grip on understanding. In my essay, I used the example of Watson, IBM’s Jeopardy-playing machine, as a case study of a system that displays plenty of knowledge but that, because it computes its answers by finding statistical associations between words, apparently has little or no understanding of the facts that it emits to win the game.
So what is it to understand a fact? Several possible components of understanding were mentioned in my essay and pushed by various commenters:
1. Some sort of grasp or knowledge of the underlying structures that give rise to the fact to be understood, such as causal structures.
2. Some sort of direct experience of the subject matter.
3. Some sort of direct experience of the underlying structures.
A few commenters wondered whether we have direct experience of any of the world’s underlying structure. On these skeptical views, it is either impossible to grasp causal structure, or causal structure is an imposition of the mind that does not reflect the structure that’s really out there, the objective structure that you would need to be acquainted with to have true understanding. Does that mean that we are no better positioned than Watson to understand what’s going on in the world? Not necessarily: perhaps it is not essential, after all, to have deep knowledge of structure in order to have understanding.
Other commenters focused on direct experience itself. How closely is it connected to consciousness? Must you be conscious of the world to have understanding? A theme that came up several times was whether other people were the natural locus of understanding. We never got to the bottom of this, but one reason you might suppose that we are better placed to understand people than things is that we have direct experience of minds—namely, our own minds. Causality might be foreign to us, but thought is surely not. We can understand other people because we can, to some extent, know what it is like to be them.
This naturally leads to the question whether understanding other people has anything in common with understanding physical processes. Is psychology, like physics, just a matter of knowing the causes of things? An important related question is currently rather topical: is the kind of understanding you get from a university education in the humanities (literature, English, philosophy) qualitatively different from the understanding you get from the sciences?
Another line of thought in the comments pursued the topic of machine understanding, picking up on my claim that Watson understands nothing. Can a machine have knowledge of causal structure? Can it have direct experience of anything? (If experience requires consciousness, we need to know whether a machine can be conscious. That is a different big question; we put it aside.) I myself am somewhat optimistic that a machine might one day be built that is not only as intelligent as us, but that has genuine understanding. A sticking point is the notion of direct knowledge or grasping; as one commenter noted, though we may know it when we see it, we don’t have much in the way of a philosophical or psychological theory of the notion.
Near the end of the comments I raised the topic of moral understanding. How useful is a theory of causal understanding, or more generally, a theory of our understanding of the kinds of facts that turn up in Jeopardy, to thinking about moral understanding? I suggested an analogy between the two, with general moral principles standing in for causal laws. Could a computer one day have moral understanding? Could we automate the law courts?
A theme we didn’t get to—we had only a week, after all—was aesthetic understanding. That can mean, on the one hand, understanding literature and art, but also, on the other hand, using literature and art to understand both the world and our part in it. The possibility of a literary understanding of life suggests that grasping causal principles is not the only route to understanding; this takes us back to question of understanding people as opposed to things, and so the humanities versus the sciences.
Understanding is one of the biggest and oldest topics in philosophy, but it has not been much discussed in the last hundred years or so, at least in the English-speaking world. Thanks to the Templeton Foundation for helping to bring it back!
New Big Questions:
- Can a machine make moral decisions? Can it have moral understanding?
- Do we understand the behavior of people in the same way that we understand the behavior of things? Does understanding in the humanities work the same way as understanding in the sciences?
- How does literature help us to understand life?