Can AI actually read? Sort of. The deed is part of a two-stage process. Say I want ChatGPT (or some other chatbot) to write a plot summary of Mary Shelley’s Frankenstein. The bot first needs to “read” the text, which resides in the LLM’s dataset. Based on what it “reads,” the bot then “writes” the summary I requested. Cranking out summaries is just the beginning. Today’s emerging AI reading apps can deftly compare multiple documents, answer questions about texts, and even assume the persona of a fictional character with whom we’re invited to converse about the book in which it dwells.
When humans read, the act is sometimes an end in itself. Think of curling up with a fast-paced thriller. However, readers might be stirred to reflect on what they’ve read (always a good idea), speak (perhaps discussing the book with a friend), or write about what their eyes and brain have absorbed. Especially in school settings, this last move, from reading to writing, is common. To determine whether students have done the assigned reading, professors commonly ask them to write about it. As the literary scholar Robert Scholes aptly observed, “we can see writing… but we do not see reading.”
Sounds like generative AI.
Yet AI’s kind of reading has little in common with what humans do. At the operational level, AI essentially “reads” individual words (technically, tokens). When humans read, our eyes often scan ahead (forward saccades), while our minds cast what we’re reading into the larger tapestry of sentences, paragraphs, pages, and chapters. What about cognitive processing? AI-as-reader is not in the sense-making business. It just executes statistical predictions. When humans read, the whole point is to derive linguistic meaning. We do so based not only on the text at hand, but likely drawing upon our life experiences—other works we’ve read, places we’ve been, conversations we’ve had, or maybe challenges we’ve faced.
Another way of thinking about the difference between human reading and AI’s version is to look at AI’s mediating role in what had been a direct connection between writer and reader. The traditional relationship between authors and readers looks schematically like this:
human author → author’s text → human reader of human author’s text
That is, humans directly read what human authors have written. In the case of AI, the schema changes to this one:
human author → author’s text → tokenized text → AI-as-reader → AI-as-writer → human reader of AI’s text
Humans still do some reading themselves, but now they read strings of words spun out by AI, not text created by a human author. It’s this mediation that radically changes what people end up reading.
When AI reads and then writes, it conjures up results that are eerily akin to what human reading- plus-writing might come up with, at least in terms of compositional quality. Why does this matter? Because in answering AI’s siren call to do the reading for us, we risk weakening our own reading skills.
As literate humans, we learn when it’s OK to skim and when the text calls out for weighing each word and phrase. AI doesn’t distinguish between the two approaches. Without sustained practice in when to read which way, we humans can be tempted to speed up everything. The same challenge plays out for analytical skills. Developing them is hard work, as is keeping them honed. If AI can analyze texts for us, why expend the effort? You can bet that, as a university professor with more decades of teaching experience than I need to admit, I’m worried how much AI-as-reader is poised to lure students into taking a pass on doing their own work.
“Thinking for ourselves” is a handy catch-all phrase that brings in analysis, evaluation, and judgment calls. But there’s another part of the meaning we shouldn’t ignore. That’s the ability—and drive—to develop our own perspective, our own point of view. The challenge with AI, for both reading and subsequent writing, is that it can silence our inner voice, which we rely on to drive our own interpretations. Using AI to read for us can stifle opportunities to work out our own take on a text, a situation, or even a person, based on material floating about in AI’s data soup. A nasty consequence of too much enchantment with a glittering technology.
We need to find balance between the reading we cede to bots and the reading we do ourselves. Why? Because there’s a lot at stake if we let slide the benefits that literacy brings, including the less tangible ones.
From Growth to Empathy
Among its many virtues, reading has long been heralded for supporting individual growth. The kind of growth I’m talking about includes understanding ourselves and striving to become better people. To build our emotional and moral character, to forge a sense of aesthetics. When AI reads for us, I can’t imagine this kind of personal development resulting.
One facet of growth is seeing the world through other people’s eyes. We feel their pain. In a word, we empathize. C. S. Lewis’ essay “An Experiment in Criticism” speaks of how literature enables us to “see through the eyes of others.” With a nod to Lewis, Pope Francis urged people to read more literature. His Vatican letter of July 17, 2024, “On the Role of Literature in Formation,” addresses “the value of reading novels and poems as part of one’s path to personal maturity.” While the epistle was directed to Catholic seminarians, the import of his message extends to a far broader audience. Francis explains that by reading literary texts, “we develop an imaginative empathy that enables us to identify with how others see, experience and respond to reality.”
Is Reading Worth the Effort
A lot of reading is enjoyable. But sometimes it entails real effort. How much is worth expending?
The concept of mental effort has intrigued social scientists for decades. One of the earliest was George Zipf, who coined the term “principle of least effort” to describe what he took to be an underlying principle of both individual and group behavior.
In their landmark 1974 paper “Judgment Under Uncertainty,” psychologists Amos Tversky and Daniel Kahneman argued that people often rely on a handful of “heuristic principles which reduce the complex tasks of assessing probabilities and predicting values to simpler operations.” In other words, to save time and mental energy, we often make quick judgments. Sometimes the strategy works; other times, we get trivial problems wrong because we don’t think them through.
In his later book Thinking, Fast and Slow, Kahneman developed the notion of using heuristics in decision-making into what he dubbed System 1 thinking—thinking fast. That’s when you speed through the decision-making intuitively, relying on heuristics. By contrast, System 2— thinking slowly—is more deliberate and effortful, naturally demanding more time. Our task is knowing when System 1 will do and when we need the more brain-intensive System 2. If we always relied on System 2, we wouldn’t get much done. (More drastically, our ancestors might have been too slow to escape an approaching tiger.) Cognitive psychologists remind us that human time and cognitive resources are limited. Accordingly, we sometimes take shortcuts in our reasoning process and decision-making. We’re being pragmatic, not lazy.
Such thinking options also apply to how we approach reading. Will consulting a single Wikipedia article tell us what we need to know about the norovirus, or should we widen our research net? The same goes for saying one-and-done after retrieving a response from Claude or Llama or ChatGPT. In each case, can we safely assume the information is accurate? Are we satisfied with the comparative analysis an AI app produces of that clutch of articles we were tasked to review, or should we read the articles ourselves?
Bias or Aversion
When AI reads for us, we economize on effort. One justification is an assumption that AI can do at least as good a job as we might, maybe better. Given the incredible feats AI has shown itself able to pull off, it’s understandable for us to be biased towards AI as a worthy option. That said, we’re still human beings. We have our pride and aren’t eager to be discarded on tech’s refuse pile. It’s understandable to have natural aversion to technology replacing us, both in efficiency and quality. Plus, we care a lot about authenticity, meaning that the label says “Made by a Human.”
Negotiating this see-saw from bias to aversion isn’t a new challenge and hardly one unique to reading. Technologies—from tractors to supercomputers— can reduce human workload, often accomplishing tasks more successfully. But a crucial question is when to trust them. To get a sense of how the bias-aversion dichotomy evolved in the world of technology, let’s start with the earliest days of aviation.
The Wright brothers made their first historic flight in 1903, with Wilbur flat on his stomach, manipulating the levers to keep the plane balanced upright. Hardly an efficient way to steer. Nine years later, Lawrence Sperry used an invention of his father’s called the gyrocompass to automatically balance a plane as it flew. The invention marked the first form of autopilot.
By the 1970s, this mechanical approach to assisting pilots was yielding to computer-based controls running on digital algorithms. The goal was to reduce both pilot workload and human error. Since most plane accidents result from human mistakes, it was reasonable for pilots to increasingly trust autopilot. Except when it wasn’t. One downside of autopilot was that human pilots were losing the skills necessary to handle potentially dangerous situations.
The problem was automation bias.
Even in the early days of harnessing computers in the service of automation, psychologists had begun to fear that handing over control to a machine could undermine human skills. More than forty years ago, Lisanne Bainbridge wrote about “ironies of automation,” where “the result is the direct opposite of what might be expected.” While the new technology was supposed to improve industrial production, workers might no longer know what to do when automated systems misbehaved. As she put it, while current workers (who used to perform the work manually) could manage by “riding on their skills,” future generations of operators wouldn’t have this training to fall back on.
Bainbridge was concerned about the unintended consequences of computers running physical machinery. These days, we need to ask about the potential impact on human skills when AI-driven computers began to do cognitive work such as reading and writing. The threats of de-skilling make for a close parallel.
Meanwhile, as technology has evolved, the discourse has shifted from automation to algorithms. You’ll hear talk of both “algorithm aversion” and “algorithm appreciation,” reflecting whether people prefer advice from humans or machines.
Both algorithms and humans are fallible. The question is, in the moment, which do you trust. If we’re using AI to read mammograms, it’s easy to envision a partnership between the AI and human radiologists. This kind of teaming up is akin to the long-standing medical practice of getting a second opinion. And unlike pilots needing to make split-second decisions, time is less critical. The issue of when to use AI to read for us is less straightforward. The only way to “verify” what AI read and wrote is then to do the reading yourself. But if you do, why bother using AI in the first place?
Attitudes towards relying on algorithms in decision-making run the gamut from total trust to negotiated partnership to outright aversion. That aversion may reflect the belief that some tasks are best left to humans—such as creating art or writing poetry or critiquing what those seeking your professional judgment have written. Readers sometimes give higher marks to works they’re told were written by humans, even when the texts were written by AI. Our human predilection for “Made by a Human” keeps surfacing.
As a member of the species, I find this propensity redeeming.
Excerpted from Reader Bot: What Happens When AI Reads and Why It Matters by Naomi S. Baron, published by Stanford University Press, ©2026 by Naomi S. Baron. All Rights Reserved.
