Why write a dialogue? Our aim was to take seriously Arcade’s new colloquy format, which opens the door to a spontaneous exchange of ideas that are not yet exhaustively filtered, moderated, or rhetorically depersonalized. The ideas that follow grew out of conversations between the two of us, and although they might someday find their way into a journal article or a book chapter, it seemed worthwhile to acknowledge them for what they are now: speculative and provisional, germinal, we hope, in both senses of that word. We hope it will read as a wager on the truly collaborative future of our discipline—collaboration measured not simply by coauthorship, but by a willingness to expose one’s hypotheses to engagement and critique at all stages of their life cycle.
J.D. Porter: There’s a moment in Toni Morrison’s Jazz where the narrator seems to break the fourth wall to explain that the plot didn’t turn out as she expected, because the characters “contradicted me at every turn”; they were “busy being original, complicated, changeable—human, I guess you’d say ” (220). In other words, the characters took control of the narrative, took some of its possibilities out of her hands; she was going to end the story some other way, but they wouldn’t let her. I remember reading that for the first time and thinking, what does that mean exactly? Is she suggesting that her characters have agency? And the more I thought about it, and recalled other authors saying similar things (the characters “took on a life of their own” or “got away from me”, etc.), the more I began to wonder if the easy answer is the correct one: The characters seem to be real, and have agency, because they are, and they do.
Hannah Walser: Cognitive literary studies has an explanation for this phenomenon ready to hand. Blakey Vermeule, for instance, claims that “the illusion of [characters’] independent agency” experienced by novelists has been speculatively linked both to “imaginary play in childhood … and to mind-reading capacities in general” (46-7). The idea is that our brains are evolutionarily primed to attribute mental states and intentions not only to humans themselves, but to humanlike entities—anthropomorphic cartoons, expressive robots, and yes, literary characters. When authors feel their characters directing them or resisting them or talking back, they’re experiencing a side effect of this overattribution of minds, just as readers are when they feel as though characters are real humans with lives outside the boundaries of the story.
JDP: The cognitive piece is crucial here, but I think this actually extends to an ontological question as well. We think of characters and fictional worlds as wholly manipulable just because they were invented; they are mere arrangements of technical features that the author could adjust in any way at all. But the suggestion that Morrison and others have made is that, once some of those technical features are in place, the author loses a little control. The characters attain some sort of actual ontological status related to the epistemological issues you raise. I think the explanation here has to do with probability, specifically Bayesian reasoning and its deployment of priors.
HW: Agreed—our judgment of others’ agency is as much a matter of probabilistic reasoning as it is of imaginative projection. When it comes to contemporary research on how we learn to understand other minds, Hume, not Kant, is the patron saint—sometimes even explicitly credited as such, by the developmental psychologist Alison Gopnik for instance (Gopnik 76). “Causal learning,” Gopnik notes, “is a notorious example of the gap between experience and truth” (75), most of all in the case of mental causes—desires, intentions, beliefs, and so on—which are by definition invisible and perhaps little more than hypothetical. But Gopnik’s idea, borne out by research into the learning processes of infants and small children, is that human reasoning about causality can be modeled according to Bayes’s concept of probability, which takes into account both the conditional probability of an event—for instance, the probability that it’s raining outside, given that I’m taking my umbrella to work—and the prior probability of the two events: that it’s raining, and that I would take my umbrella to work on any given day regardless of weather. (Wikipedia has a good breakdown of Bayesian basics.)
I’m going to explain this in detail, because Gopnik’s adaptation of Bayesian probability to social cognition strikes me as a persuasive theory of how we come to see other humans as intending agents with personalities: we notice that variations in behavior are often predictable on the basis of hypothesized individual preferences and beliefs rather than inherent features of the world. Take the setup of one of Gopnik’s experiments: a baby is confronted with an adult confederate who seems to find broccoli delicious and Goldfish crackers disgusting; then, with two bowls in front of the baby—one filled with broccoli, the other with crackers—the adult ambiguously asks the child to “give me some” (Gopnik 55-6). In essence, this experiment asks the infant to make a prediction: which food will make the confederate happy? Fourteen-month-old babies, regardless of the adult’s preferences, give her a handful of crackers: they love crackers; the prior probability of crackers leading to happiness, in a baby’s world, is extremely high. But by eighteen months, babies have begun to construct what Gopnik calls a “causal map” (39) of human minds; they’ve begun to understand visible reactions—smiling and saying “yum,” or putting out one’s tongue and saying “yuck”—as effects of invisible causal entities called mental states. So when the confederate asks for a snack, the eighteen-month-old will hand her, however incredulously, some broccoli; given this individual’s previous happy response to broccoli, the baby calculates that more broccoli will produce the same result. (Why are babies so obliging, so eager to make others happy? That’s a question for a different dialogue.)
So far, so good; bravo baby. But complications can arise when these epistemological assessments ossify into ontological assumptions. Another study by one of Gopnik’s colleagues, Elizabeth Seiver, found that American 4-year-olds were sensitive to statistical differences that implied person-based or situation-based explanations; in other words, they varied their causal explanations depending on whether they were looking at a person’s consistent behavior in multiple situations (implying that some trait specific to the person makes him act this way) or a situation that elicited consistent behavior from multiple different persons (implying that a feature of the situation itself is responsible). Surprisingly, though, American 6-year-olds were actually less accurate at this task: they tended to explain all behavior in terms of personal traits (Seiver et al. 450). Why would two extra years of learning and being in the world decrease the children’s empirical accuracy? Seiver et al., drawing on cross-cultural research in developmental psychology, suggest that children’s causal explanations draw not only upon “covariation information” (i.e., statistical patterns) but also on “adult trait language,” the terms that other members of their culture use to explain behavior (451-2). Although Seiver and colleagues don’t mention it, literature is surely one of the main vehicles for communicating these terms: when Western children read stories about children who behave in particular ways because they are brave or kind or clever, they’re internalizing a trait-based explanatory system—whereas they would infer a quite different causal system from stories where children do things because those things are easy or safe or because their parents want them to.
JDP: This has fascinating implications when it comes to novelistic ontologies, because it seems to me that a well-developed character actually attains agency. I don’t mean this in a rhetorical or metaphoric sense: I mean real agency. And to clarify an important term, by well-developed, I mean realistic, which, as you say, means variations of behavior that fall within some band of predictability. In some of your other work you mention a sort of sweet-spot between total randomness and total predictability: a random-letter generator and a machine that prints “dog” over and over don’t seem agential; a machine that answers questions with unique sentences, even if they are not particularly coherent, seems kind of agential. To all that I’d add that a realistic character should be further constrained by the kinds of behaviors that make one person seem unique relative to all others. As a simple example, if you’ve been writing a sweet old lady villager character for 300 pages, she can’t suddenly become a harsh space ship captain, no matter what your plot machinations are. The old lady controls the possibilities by virtue of being a plausible old lady; you write her past, but that makes her into some sort of entity that determines her future. This is why we’re calling this Bayesian: The author sets priors, but these priors then affect the probabilities of future events. Once you make her a sweet old lady villager, the sweet-old-lady-villagerness becomes a real, existing factor that has to be contended with. The epistemological unrecognition that would be caused by deviation from those priors stems to some extent from the ontological status of her field of probable/plausible actions.
HW: If you, as a reader, encountered that story where a sweet old lady turned into a brusque spaceship captain, the deviation from your prior probabilities—that is, the “causal maps” and assessments of likelihood you’ve developed through interacting with humans in the world—would be so great that you’d probably reject the story outright, or at least shift the grounds of your interpretation. (You might, for instance, start thinking about the text less as a mimetic representation of characters’ personalities and motivations and more as a complex game of authorial intention.) It’s perhaps for this reason that we don’t see too many novelists who completely dispense with their culture’s ideas of characterological consistency.
But now imagine a different narrative in which the sweet old lady became an accessory to murder. Still quite surprising, but now perhaps assimilable enough that you would actually use this information to revise your priors. In this case, for instance, if the old lady was placed in circumstances where inaction led to murder, you might shift your baseline causal map just a little more toward situation-based (rather than trait-based) explanations. A different story, if it showed you a man whose disposition completely changed after he sustained a traumatic brain injury, might lead you to increase the role of physiological factors in your behavioral predictions. Such narratives won’t necessarily lead to a more accurate understanding of human behavior, but the key is that they’re doing more than simply reinforcing cultural and cognitive defaults; rather, they function as evidence, less potent than real-world experience but still influential, that our brains can use to rewrite their probability assessments. This is another form of textual agency: the probability-worlds we encounter in fiction have the potential to change our priors in the real world.
JDP: This concept of priors is such a great way to think about the ontological layering at play in literature. A text is an event (not just an object), comprising a blend of things the author makes up, things from the real world, things the reader interprets, and, we’re saying here, things that the text winds up imposing of its own…well, something like volition. So we can think about the text becoming a plausible story (maybe even recognizable as a story) only by virtue of correspondence with real-world priors. If you imagine a narrative that totally departed from those, who knows what you’d have—some Beckettian slog. It’s fun to imagine what that would even look like—wildly inconsistent characters in rapidly changing settings, plots that ignore their own history, unexplained departures from the laws of physics…very interesting, but not much of a narrative. So the real world matters in making the story. But over the course of the narrative, the text acquires its own set of priors, as an inevitable side effect of this first process; the ones it borrows turn into its own. And so it’s worth emphasizing that this blended ontology of the text is also a blended ontology of the real world—after all, both things are being blended here. And, as you say, the text can even, in practical terms, push back, either on Toni Morrison or on a reader’s priors.
HW: These distinct (but interwoven) ontologies are, I’d argue, inextricable from—maybe even identical to?—distinct epistemologies. Let’s say that the term “agent” labels a particular region on the probability band, a particular regime of prediction that’s based on inferred intentions and mental states. Then concepts like “object” and “machine” and so on can be imagined as occupying neighboring regions and entailing different probability assessments. In other words, the ontological categories of “agent,” “object,” and so on both derive from and imply differing probability patterns in perceptual data. If this is the case, then agency should be understood as a kind of continually revised assessment of the likelihood that any given behavioral event is causally linked to the unique individual in question.
Why is this important? Well, most extant cognitive theories of literature place all of the power and agency in the reading encounter on the side of the brain; the brain sets the terms of the reader’s encounter with the text, which in turn can only represent phenomena that fit with the brain’s own categories and heuristics. Our purpose here is, if not exactly to flip that agency, at least to redistribute it: the text is a relatively autonomous system of probability, not identical to the probability-system of the world although related to it, and the brain makes inferences and predictions on the basis of data from both these systems. In other words, the text creates the conditions within which the brain constructs its ad-hoc ontologies.
JDP: There’s an obvious objection here that I want to address: Couldn’t we change the characters? Maybe they or their texts have certain probability restrictions that make some changes less likely than others, but at any time the author could make the old lady a space captain; for that matter, the reader can imagine her as a space captain at any time. It feels as though just this possibility is an important departure from extra-novelistic ontologies; we can’t just make actual old ladies into grizzled space captains whenever we want. The option does not meaningfully exist, even for our most powerful regulatory agencies. I have two responses to this:
1. Well, maybe we could do this, but we don’t. Empirically, this just isn’t how stories work. When something like this does happen, it’s strange enough to stand out as a signature feature of the work, evidence of postmodernism perhaps—but not just another tool in the standard narratological arsenal. If it’s important that we could defy probabilistic narration, it’s also important that we almost never do.
2. We can adjust actual ontology. It happens all the time. See, for example, the temperature of the earth, social fictions like race and gender, the question “Is there a city called Carthage”, etc. This is a signature issue in, for example, ecocriticism: Passenger pigeons once had undeniable ontological presence; they could literally blot out the sun. Now they do not exist; we, people, did that. In the 1940’s we even started getting rid of atoms by turning matter into energy. I don’t want to say that novelistic ontology is the same as real-world ontology; but its amenability to adjustment is not evidence against its ontological weight.
HW: Maybe we can explain the difference between real-world and fictional ontologies in terms of their resistance to exactly that kind of adjustment: literary beings put up less resistance than nonliterary ones, but they still put up some. I’d go even further, in fact: maybe the difference between the ontologies of agency that I was discussing above—conscious entities vs. nonconscious but animate entities vs. machines, etc.—can also be understood in terms of the degree and quality of resistance to imaginative and material intervention that they mount. I can’t talk an apple off a tree, although I can talk a person into helping me; I can’t (unless I’m empowered by the state or backed by violence—again, seeds I’d like to plant for future conversations) make a person do what I want by changing the condition of his body, but I can make a machine do what I want by changing its physical construction. Humans move through the world with interests and desires, after all, and our predictions and probabilistic assessments of future events are in practical terms assessments of how likely we are to be able to manipulate or change those events. Perhaps the sense of fictional characters’ agency shared by both writers and readers results partly from the complexity of these textual “machines” and the difficulty of rewiring them.
JDP: The world is a little less different from a text than we often think, especially when you look at it through this lens of resistance, where ontological status is not a binary but an amount. This is another useful concept to borrow from Bayes: That the answer to a question is very often not “yes” or “no”, but “some”. Will Jeb Bush win the primary is best answered with a percentage: He has, say, a 45% chance. Similarly, is this poem a sonnet? might best be answered “This poem is 80% a sonnet”; Is this person an agent? by “She is 60% an agent”. In this view, texts and characters and fictions are a little more real; apples and machines and people a little more fictional.
Gopnik, Alison. The Philosophical Baby. New York: Farrar, Straus and Giroux, 2009.
Morrison, Toni. Home. New York: Alfred A. Knopf, 2012.
—————. Jazz. New York: Vintage International, 1992.
Seiver, Elizabeth, Alison Gopnik, and Noah D. Goodman. “Did She Jump Because She Was the Big Sister or Because the Trampoline Was Safe? Causal Inference and the Development of Social Attribution.” Child Development 84.2 (2013). 443-454.
Vermeule, Blakey. Why Do We Care About Literary Characters? Baltimore: Johns Hopkins University Press, 2010.
 This moment is actually a little ambiguous; the speaker might be a character of some sort. Still, it certainly suggests my reading as well, especially since this kind of thing is a common move in Morrison’s work—elsewhere in Jazz, but also in Home, for instance, when the protagonist tells the narrator she got the story wrong: “Earlier you wrote about how sure I was that the beat-up man on the train to Chicago would turn around when they got home and whip the wife who tried to help him. Not true. I didn’t think any such thing” (69).