How to read a novel as a theory of behavior
Stories can model the one thing computation never can: experience
“What I cannot create, I do not understand.” — Richard Feynman,
the de facto mission statement of the cognitive scientist
When the field of cognitive science was founded in the 1950s, its architects had a clear goal. It was to develop a theory of the human mind so powerfully insightful, so comprehensively accurate that it could be implemented in a computer.
Unlike previous approaches to psychology, the point was not to understand the mind by drawing sophisticated diagrams of boxes with vague labels connected by arrows, nor to convey one’s theory in an incisive essay. No, the point was to show (not tell) that your theory was correct because you could construct something that, when presented with the same set of stimuli, would respond in the same way as a human.
The foundational belief of the early cognitive scientists was that whatever the mind happened to be, not only could we know about it: we could build one.
This is the mission of cognitive science I fell in love with as an undergraduate. I was captivated by the idea that if you really figured out how the mind works—if your theory was sufficiently true—then you should be able to demonstrate that knowledge by recreating a mind in the laboratory.
And it is in this vein that pretty much every cognitive scientist reads Feynman’s claim that engineering is the ultimate proof of understanding. What counts as “creating” a mind is to replicate its fundamental processes via computational models. If you get that right, then everything else can simply be added on top. Having a sophisticated computational model, in this sense, is like having a good soup stock on hand: one can always add in vegetables, seasoning, and other accoutrements later on, all of which are undoubtedly present in any given soup but are not the soup itself.
And yet a stock is not a soup: a model of the mind’s fundamental processes is not a mind itself. The pursuit of the computational cognitive scientist is to build a model of the mind in general. It is to create a paired down version of the mind’s most basic structures of thinking, remembering, reasoning, and understanding that are common to all of us. And don’t get me wrong. These are things worth knowing about.
But there is another way of reading Feynman’s call to action.
It is not to synthesize a version of the mind in its most general and abstract form, a kind of lowest common denominator of what we all share. It is to capture the mind in its specificity, to account for what makes the act of thinking—given one’s particular life experience and circumstances and personality and social milieu—so irrepressibly unique to a given individual.
An individual’s psychology is not made up of sterile algorithmic processing. It is inextricable from the experiences, the significant life events, the relationships with other human beings, the sources of meaning which give form to its function. Cognitive science, as it is current practiced, is happy to forfeit its claim to modeling this kind of cognition. Everyone gets everything they want. And for their sins, cognitive scientists will get a facsimile of the mind divorced from what makes our minds not just information processors—but us.
Luckily, there is a group of people whose goal is also to construct a facsimile of human behavior, but not the paired down one of underlying, fundamental computational principles. It is closer to what an individual human life looks like in all of its complicated glory. This kind of model cannot process novel stimuli, or tell you which images feature cats and which don’t, or compose insipid college application essays (all instances of a soup insufficiently seasoned). But it is nonetheless, as Feynman’s dictum requires, a way of showing one’s understanding of something by attempting to create it for yourself. The specific goal of this approach is to reconstruct the human experience from the point of view of the person experiencing it. And the people who create this kind of model of the mind and its consequences are not computational cognitive scientists. They are novelists.
At first glance, novels and computational models couldn’t be more different. But they have more in common than you might initially think.
For instance, both require their creators to confine behavior to a highly-controlled realm, where all the levers and dials sit laid out for fine tuning by the person behind the curtain. For the novelist, this is a world that is made up, composed at the author’s god-like discretion. It is one which may resemble ours in some ways, though it will inevitably deviate from it in others. But the cognitive scientist shares the same requirement: the laboratory is itself a kind of fictional world.
To take in-lab behavior seriously requires suspension of disbelief. We must believe, even temporarily, that the only thing which matters is one specific influence on behavior. This is whatever is being experimentally manipulated. If that weren’t the case, if there were any number of things which could be influencing a participant’s behavior, then it wouldn’t be a very well-controlled experiment.
And though we know that’s not how things actually work, that behavior is always determined by far more than just one variable, we agree to set aside that knowledge for the sake of learning what the experiment (and the model designed to predict people’s behavior in it) can show us.
The fictional world of the novelist is also a laboratory, one of a different sort. Perhaps some of it is overtly unrealistic. There might be spaceships, or US presidents who never existed, or streets in London which aren’t actually there. No one actually did the things that are being described, at least not exactly as the events have been depicted by the author. But keep in mind: in a controlled lab experiment, no one would be taking those actions either if the experimenter didn’t arrange things just so. The job of the author, like the scientist, is to fix the parameters of the world at hand, then within those constraints create a realistic simulation of what, all else being as it is, might actually occur.
This is the implicit contract between the author and the reader:
If the reader accepts the fictionalized circumstances—the controlled variables which can be set up in a fictional environment but are not guaranteed, or even possible, in real life—then the rest of it will be real.
This is also the contract between the author and reader of a scientific paper, as well.
For both a well-designed experiment or novel, the promise of its architect is to create a scenario in which the behavior under scrutiny is uniquely revealed by the circumstances. If the reader accepts the limitations of the fictional world, the author promises to give them a privileged vantage into the topic at hand.
And, in a way, this is already how people naturally read novels.
They don’t view them as purely “fictional” in the sense of being wholly unrelated from reality. They view them as hypothetical versions of what really exists, the consequence of the author’s own simulation of how the world works. A cognitive scientist simulates behavior based on an understanding of the mind in terms of its most fundamental, decontextualized processes. And when a novelist simulates behavior, they do based on an understanding of a specific mind in the full-breadth of its social context.
When people read novels, they are implicitly trying to reverse engineer the generative model the author used to create these events. This is why people get worked up and emotionally-moved (or, alternatively, worked up and offended) by stories. The events are made up. But the stakes of the underlying generative model are real.
In a fictional world where anything can happen, what the author chooses to have happen says a lot about what they think does, in fact, happen.
In my final episode of Cognitive Revolution, I presented a rough sketch of this idea to Stephen Kosslyn. Kosslyn is an eminent cognitive scientist, the kind whose name permeates any introductory cognitive psychology course. He’s been in the game a long time; he was Steven Pinker’s PhD advisor. We were talking about the nature of computational models, and what they can (or can’t) tell us about the mind, when I made this claim about novel being a kind of models of mind. “So let me ask you a question,” he said. “Let’s say that a novel is a construction of an understanding of psychology, writ large, in all of its ‘complicated glory.’”
“How do you know whether the author got it right?”
With a computational model, there’s a straightforward answer. The model is a good one if it predicts the behavior of human participants in an experiment. And if the model bares no resemblance to the behavior of your participants, then you’ve failed to provide evidence that it’s a good model. What the equivalent mechanism for evaluating a novel as a model of behavior?
When he first posed this question, I didn’t know how to answer it.
It seemed like an obviously worthwhile follow-up question. After all, it’s the most direct violation of the scientific method committed on a regular basis by novelists—lack of experimental rigor!
But I’ve come to think that it’s actually not the right question to ask about novels, because it’s not the right question to ask about models.
What changed my mind was another Cognitive Revolution interview, one with Sam Gershman, whose lab I worked in for two years. At one point in the conversation, I asked him about his general approach to modeling.
What he told me was that models aren’t about creating an unimpeachably accurate representation of the world. Rather, they are tools for thinking.
“Particularly when it comes to models, I don’t see models as literal descriptive statements of how the world works, or how cognition works, or whatever. The models, really—when they’re good—are tools for thinking. And they’re part of an intellectual dialectic that will eventually produce better models, which are better tools for thinking about some problem domain. And, of course, along the way you’ll produce models that are better descriptively. But I don’t really see that descriptive goal as really the primary goal of modeling, at least for me...
“In general, we have to approach our theoretical arguments with a sense of humility and not believe them too literally—treat them more as a conversation piece rather than a literal statement about the world.”
The measure of a model is not its ability to simulate the world as it really is. Instead, it is how well it articulates a way of how things might possibly be. Computational models are a means of outlining the space of possible worlds, any one of which we may happen to find ourselves in—or not. A model is good to the extent that it can help you think more clearly about the topic at hand [1].
This changes the frame for evaluating a novel as a model of human behavior. The relevant question, then, becomes not “Did the author construct the most accurate picture of reality?” but “Does the author give the reader a potentially-useful tool for thinking about reality?”
If the author has done her job—and we as the reader have the skills to figure out what we’re looking at—then this rendering should be a useful tool for helping us think through a particular aspect of reality in a way we would not have otherwise been able to see as clearly.
Think about it this way: If every painting had to be realism, there would be no impressionism, surrealism, cubism, expressionism. Our “pictures” of the world would be more accurate, strictly speaking, but our set of tools for representing and thinking through reality would be smaller.
“All models are wrong, but some are useful”.
This is a quote originally from George Box, and a refrain that’s nearly as common in cognitive science circles as Feynman’s.
My goal in making this argument is not to suggest that computational models from cognitive science are any more wrong or not-useful than anyone within the field might personally believe. Rather, in acknowledging their wrongness, as well as the limits of their utility, I would like to indicate beyond the usual scope of what is considered within the space of legitimate models of the mind. In particular, novels are good for modeling the one thing that computation never can: experience.
There is a reason that novels, as models of experience, are not verifiable. It is that experience itself is not verifiable.
All any of us has is our own fiction. What is “true” for any of us, ultimately, is not the events themselves. It is that each of us lives our own version of them. Novels give us a template for how construct our own narrative, our own way of telling the story of who we are. Whether or not that way of telling it is generalizable in some pan-human sense of fundamental cognitive processes is beside the point. The only universal human fact is that we are all living our own complicated human experience, and it’s up to us to make sense of it.
Still, I understand why scientists don’t like novels, why they’re not an obvious object to incorporate into scientific protocols. They lack generalizability. They are not verifiable in the positivist’s sense, and for the most part they are immune to falsification. They do not lend themselves to replication; unlike the sleek generative models of the cognitive scientist, the model of the novelist is a single simulation affair. The novel offers little in the way of an ability to sample from a large number of probabilistic iterations in order to determine what happens on average [2].
But nonetheless, I can’t help but feel the endeavor of cognitive science will remain incomplete until it has a means of dealing with experience.
My proposal is not for cognitive scientists to forfeit their computational models and devote themselves to literary criticism full-time. But rather my claim is that the scientific method alone isn’t enough to reckon with the entire scope of mind, meaning, and experience. Those who are interested in studying the mind need not abandon the scientific enterprise. But I do think we need to take seriously the methods of interpretation, traditionally the purview of the humanists, if we want to deal with questions of human meaning. Without this, cognitive science will one day find itself in possession of a bowl of something warm and hearty, filled to the brim with a base of undoubted quality, as well as a lot of the ingredients you’d hope to find. But it’ll be missing something. The soup will forever remain slightly off.
[1-2] See comments for footnotes
2] When you judge your model in terms of its simplicity, you will tend to study things which lend themselves to simple explanations. It’s not that mathematical explanations are overly simple—perhaps theoretical elegance is the preferred term here—but that this approach can only investigate phenomena that hold together well when you reduce them to their essence. This is true of the kind of things studied by computational cognitive scientists. They can point to the equation and say, “Here, at core, is what is happening.”
Novels, like experience itself, don’t work that way. They don’t hold together well when you reduce them to a bare bones plot. Their essence dissipates. Likewise there is no fundamental computational model that can distill experience into its primordial form. Unlike computational models, novels are not their essence. That’s why the first thing you learn in high school English is not to write “plot summary.” If the summary was all that was needed to get the author’s point across, they would’ve just written that to begin with.
Some aspects of cognition may lend themselves to the reductionism of computational models. But humans, in their wholistic experience, are not simple.
Occam’s razor only takes you so far.
1] And here’s another thing to consider… Cognitive scientists don’t actually evaluate models solely on whether they are empirically validated. Most models languish in scientific obscurity. The p value maybe close to zero; but so is the citation count.
Rather, a small subset of models which go through the rigors of experimental verification are actually determined to be useful within the larger scientific community. These are the ones that get cited.
So besides simple popularity metrics such as citations, how do cognitive scientists determine whether a model is any good? There’s a simple answer: interpretation. Someone with the relevant expertise—one who has thought critically about other models which attempt to do the same thing—must look closely at the model and determine what works and what’s lacking. Then other people with a similar expertise must debate whether that evaluation is correct or not.
I think it’s somewhat provocative that these are essentially no different than the mechanisms we have for evaluating novels. There are simple popular metrics such as book sales, which probably tell us at least something about whether the story resonates with readers, whether they find the model useful in one capacity or another. Then beyond that we have to rely on the in-depth considerations of a community of thoughtful, well-trained readers.