Google's Gemini Agreed With Me. That’s When I Got Suspicious.
I asked Gemini whether AI book summaries cost students the learning. It said yes — fluently, confidently, and on a foundation that doesn’t hold.
In the past month, I have been wondering about the role of artificial intelligence in learning in my free time. My wondering usually happens over lunch as I absentmindedly sip on iced tea while grading papers. This time my wondering broke lunch and I drafted an AI question to avoid my students’ papers. I usually bounce questions off colleagues and this question was no different. My question was, “If a student uses AI to summarize a research article or book, did the student learn what the assignment was teaching?”
Here is what came back, verbatim:
“The short answer is partially, but mostly no—and the difference comes down to whether the assignment’s goal was to transmit information or to build cognitive muscle.
If an assignment is strictly about knowing what happened in a story, an AI summary delivers the plot points efficiently. But in education—especially literature, history, and humanities—the plot is rarely the real lesson. The value lies in the process of reading itself.”
I agreed with every word, and I sat there pleased with myself for about four seconds.
Then I got suspicious. An answer that flatters what I already believed was the one I’d be least likely to check. So I did to it what I make graduate students do to everything: find the claim, weigh the evidence, follow the reasoning, name the implications, ask about the context. It took most of an afternoon and I didn’t enjoy it.
The claim hedges in its first six words, and the hedge is the whole answer in miniature. I can’t restate it in one declarative sentence, which is my test for whether a claim is a claim. It lets the answer keep a foot on either side of the argument for another 800 words.
The evidence is where it fell apart in my hands, and not the way I expected. There is none: not one citation, not one study, not one researcher, and not one number. The closest the answer comes is this: “Neurological research shows that reading fiction activates the same brain regions as experiencing events firsthand.” Which research, conducted by whom, on how many people, published where? The sentence is doing the work of a citation while containing no information a person could check. I teach doctoral students to circle that construction in a manuscript.
The reasoning arrived as a table, which suggests its own kind of neat authority. Here is a screenshot:
“Cognitive Friction | Minimal (digestible, pre-chewed insights) | High (forcing focus, processing complex syntax, interpreting subtext)”
“Empathy & Perspective | Zero (analyzed externally) | High (narrative transport—temporarily experiencing another life)”
Reading the Book scores High on all five dimensions. The AI summary scores High once, on plot, then Minimal, Zero, Low, and a neat list. A comparison that can only tip one direction is not a comparison.
Look at what the friction row is doing. Friction is scored as a virtue, which happens to be a position I hold and have argued in public; nowhere does the answer say why it should be one or cite a source. It assumed I already agreed, and turns out, I did.
Then came the fluency tells.
“What Reading Actually Teaches (Beyond the Plot)
1. Cognitive Stamina and Deep Focus
Reading a full-length book requires sustained attention over hours, forcing the brain to hold multiple storylines, subtexts, and character arcs in memory. Skimming an AI output trains the brain for surface-level scanning, which degrades the ability to focus on long-form, complex material.
2. Processing Ambiguity and Subtext
Great literature relies on ironies, unreliable narrators, tone shifts, and implicit meanings.
An AI summary tells you: “The narrator feels conflicted about his choice.”
Reading the text requires you to detect the conflict through word choice, pacing, and silences. Inferring meaning builds critical thinking; having it explained removes the mental work that creates the skill.
3. Vocabulary, Tone, and Prose Style
Absorbing an author’s style, rhythm, and vocabulary happens through passive exposure while reading. Summaries strip away the author’s voice in favor of uniform, neutral prose, skipping the passive language acquisition that makes students better writers.
4. Emotional Resonance and Empathy
Neurological research shows that reading fiction activates the same brain regions as experiencing events firsthand. You don’t just learn that a character suffered loss; you sit with their grief across dozens of pages. An AI summary turns empathy into a bullet point.”
And the context is the failure I should have caught first. The whole answer is about novels. Gatsby is the worked example, and Shakespeare and nineteenth-century literature are the hard cases. Empathy, narrative transport, sitting with a character’s grief across dozens of pages. However, I assign bundles of scientific research articles. Nothing in Gemini’s eight hundred word response touches my question about reading research: what happens when a student summarizes a methods section, or how a reader sits with a results table, or what limitations means in a paper that has none. It answered as though all readers are reading fiction.
Finally, the prescription:
“The Exception: When AI Actually Helps Learning
AI summarization isn’t inherently destructive—it depends entirely on when and how it is used:
❌ As a Replacement: Skipping the book and reading the summary gives a illusion of competence. The student can pass a basic multiple-choice quiz, but cannot write an original, deeply analytical essay.
As a Pre-Reading Scaffold: For dense classical texts (like Shakespeare or 19th-century literature), reading a quick summary first lowers cognitive load, allowing the student to tackle the actual text without getting lost in plot details.
As a Post-Reading Review: Using AI to check comprehension after reading or to find specific themes to cite in an essay acts as an effective study partner.”
Having spent the table establishing that summaries hollow out learning, the answer offers two ways to use them anyway: “As a Pre-Reading Scaffold” for dense classical texts, and “As a Post-Reading Review” to check comprehension. Read AI before the book, then read AI after the book. The critique dissolves into a schedule.
There’s a typo in it. “Skipping the book and reading the summary gives a illusion of competence.” I have read that sentence perhaps ten times and I still think it is the most honest line in the answer.
I have spent the summer writing about the students I’ve been calling the honor-roll offloaders. Sitting at that table with an empty tea glass, I understood that Gemini had just done the identical thing in front of me: produced a fluent, well-formatted, confidently structured artifact that looked like understanding and was assembled without any of the work understanding requires. The summary and the summarizer share one flaw: it agreed with me but it was hollow.
Knowing if an argument holds is what the five CERIC questions are for.
I went back into it once more before publishing this, hunting for the place where Gemini had actually conceded something rather than performed a concession. It was one line above the schedule, in the sentence that undoes everything above it: “AI summarization isn’t inherently destructive—it depends entirely on when and how it is used.”
A total hedge.
A question for the comments: What’s the last confident answer you checked, and what did you find?
Read deeply,
Dr. Genevive