AI and Reading Comprehension

Tim Shanahan

Teacher question: I am writing with a question about AI and reading comprehension instruction. Our school recently circulated an AI research framework indicating levels of permissible AI use in students’ research tasks. Level 0 prohibits generative AI entirely. Level 1 allows students to use AI to translate or simplify texts. Level 2 allows AI to locate sources. Level 3 allows AI-generated explanations as sources of information. Level 4 allows students to use AI to summarize or synthesize multiple sources. Level 5 involves analyzing AI output itself as the subject of research. The policy states that AI cannot replace students’ own thinking.

I wondered whether any aspects of it might raise concerns from the perspective of how students develop comprehension through reading. Allowing AI to provide explanations of sources or to summarize and synthesize multiple texts seems as though it could bypass some of the processes that develop comprehension, such as grappling with complex texts, comparing sources, resolving inconsistencies, and constructing one's own synthesis of information.

From a reading comprehension standpoint, do you see any aspects of this kind of AI research framework as potentially concerning?

Shanahan responds:

I’ve been receiving different versions of this question for about three years. This one is particularly articulate; I’ve only included an excerpt of it. The author herself provided a good answer to her own question.

Why haven’t I responded earlier to these queries?

To tell the truth, I didn’t know how to answer. I knew of some AI tutoring schemes that seemed potentially hopeful, but that was about it.

As this query points out, AI may undermine learning as much as it could help it along. This danger seems especially pertinent to the development of reading comprehension.

Accordingly, this two-part blog/podcast entry will address some of those concerns.

It makes sense to start with my definition of reading comprehension. It is the ability to make sense of information conveyed in written language – the ability to negotiate the affordances and barriers of text.

This is an “active reader” kind of definition. It includes understanding, inferring, judging, interpreting, relating, and remembering – all actions that readers may have to implement to make sense of a text.

But the definition also emphasizes the social aspects of this sense making. Texts are not natural phenomena; they are written by or for somebody who has a communicative purpose. As such, texts may include definitions, descriptions, graphic elements, explanations, examples, analogies, repetitions, and so on, all aimed at helping the imaginary readers (those readers the author imagined would be reading the text) to get the point. Those are the kinds of linguistic and conceptual features that act as affordances. They are efforts an author makes to helping the readers to grasp the message.

Of course, authors differ in how well they do this imagining or in how effectively they address their readers’ communicative needs. However, even if they were to do these things to perfection, there’s always the chance that some unexpected reader may come along who will be unable to make use of some of these affordances. They might even be perplexed by them. Perhaps the author’s diction excludes somebody, a metaphor miscarries, or a reader simply doesn’t know what to do with a complicated graphic element. When that happens, these conceptual and linguistic elements themselves may become barriers to understanding.

Readers, to comprehend, must take advantage of the affordances to a sufficient degree and surmount enough of the barriers to make sense of a text.      

Learning to read means learning to do this with a wide variety of texts; texts that vary in content, style, purpose, structure, genre, language features, and, yes, difficulty.

Since the 1940s, teachers have been told that kids learned reading best when taught with texts at “their reading levels.” Nevertheless, research has overwhelmingly rejected that idea (Shanahan, 2025). Kids make greater gains when they get a chance to try to make sense of texts they can’t yet read reasonably well. Such texts give them the opportunity to grapple with and figure out some of those affordances and barriers.

Given that starting point, the idea of having AI “summarizing or synthesizing” texts for students seems like a really bad idea.

I find exercise to be tiring, sweaty, and often boring. Whether I’m running, swimming, bicycling, or lifting weights, I’ve often fantasized about hiring someone to do those things for me. That way I could easily exercise 3-4 times a week and still do everything else I want to do.

Now, Cyndie, my wife, is a bit of killjoy. She’s been downright discouraging about my hiring idea. She is steadfast in her belief that I will have to do my own exercise if I’m to benefit. The person who does the exercise is the one with stronger bones and muscles and clearer lungs.

It’s the same when it comes to learning to read. No one else can do that for you – not even machines that have consumed hundreds of billions of words.

READ MORE...