How to Not Think Like a Bot
In my high-school English classes, I often tell my students that they write their papers backwards: They devise a thesis and then look for evidence to support it. They’ll find what they want to see, I tell them, but they won’t be surprised. They might say that Ishmael, the narrator in Herman Melville’s Moby-Dick, is appalled by Captain Ahab, but they won’t notice that the two characters also sound a lot alike.
The most exciting moments for a teacher come when students stumble onto something unexpected—when they run to my office to tell me about a new twist in their thinking about birds in Sula or the discovery of yet another biblical reflection in Housekeeping. Those revelations come only when they survey the text as it is, not as they assume it to be.
Many students assume that they know what they’ll see—in books, people, or anything else—because that’s what they’ve been taught, relentlessly and often subliminally. By the time they were learning to read, many schools had moved away from teaching traditional phonics. Instead of asking kids to sound out letters, the new approach, called “whole language,” had them divine words from only the first letter, the context, and maybe a picture. In some cases, teachers covered up the word and had students guess what it was. The implication was that language is reliable, consistent, and predictable, which is awfully boring—and dangerous.
Perhaps unsurprisingly, that approach to teaching literacy didn’t work. In recent years, particularly after a 2020 research report and a 2022 investigative podcast brought this flawed pedagogy into the public eye, whole language has fallen out of fashion. Parents, teachers, and just about anyone who knows how to read have piled on with a common-sense rebuke: How can you read a word without seeing it?
[John McWhorter: How I taught my kid to read]
But even if we’ve moved past the heyday of whole-language reading, we clearly still believe that we can draw conclusions from context rather than from all the facts before us. Rarely has the adoption of a new technology been so quick and insidious as what we’re witnessing in the popularization of large language models such as ChatGPT.
LLMs imitate human language by way of probabilistic guessing. Trained on books, the contents of the internet, and more, the models use patterns to predict outputs. If you plug in “I woke up and ate,” ChatGPT might guess that the next word is breakfast, pancakes, or cereal. Rarely, rat. Rarer still, mitochondria. Chatbots don’t understand; they don’t read; they don’t think. They create responses that reflect the probability of each word—or, to use the technical term, token—appearing in a sentence.
Generative-AI chatbots basically adopt the same flawed understanding of the world that informs the whole-language approach to reading: that one can best discern meaning by looking at external clues rather than straight at the facts. The resulting text may astonish with its plausibility, but a guess, even one of extraordinary mathematical complexity, is a guess.
[Read: The people who marry chatbots]
The problem with guesswork is not just that the results can be wrong. It’s that even the most educated of guesses reduce the world to a sequence of likely tokens. Sure, some predictions are warranted and even useful. The sun does rise every morning. My pen falls every time I release it from my grasp. But a chief delight of being human is witnessing the world’s capacity to surprise.
Reductive presumptions are even worse when it comes to people. Stereotypes steamroll the beautiful mess of individual desires and idiosyncrasies into something simplistic and guessable, when in fact we contradict ourselves; we contain multitudes. We are not actually so predictable. We can’t be guessed.
My students, like most people, sometimes approach conversations guessing what the other person will say, as if human thought consisted of probabilistic tokens. They see a woman with a nose ring or a man in a pickup truck and preempt the argument that they assume is coming. Cover up the word; guess from the context. Find the evidence that supports your thesis.
When my students read without expectations, they discover the unexpected. That’s how one senior, who didn’t know the word fust and had to look it up, noticed other elements of mold and decay in Hamlet and ended up writing about the rot of Denmark’s “unweeded garden.” It’s what led a ninth grader to see the Everglades, in Their Eyes Were Watching God, as a place that combines the stability of land and the mutability of water, a place where the characters, like the plants they cultivate, can grow.
I encourage my students to approach texts with no notion of what’s likely, not only because it yields better, more interesting papers, but also because I believe that it will make them better citizens. I hope that they learn to see the world, as they do texts, for what it is, in all its guess-defying beauty.