What you lose when you let a machine read for you

When AI summarizes for us, we stop thinking for ourselves

Life is too short to read everything within our reach.

There's a famous episode of Seinfeld (1994) where George joins a book club without thinking it through too much. He just has to read Breakfast at Tiffany's, a "little book" of less than two hundred pages. Days go by and when Jerry asks him what it's about, George can barely stammer. "You haven't read a single page, have you?" Jerry asks, and it's easy to empathize with the response: "No… I wish I were [brilliant], but I'm not." To which Jerry replies: "Why don't you just rent the movie?"

When faced with a shortcut, the temptation is to cheat. And, like George, the risk is failing miserably.

When there's no movie to "save" us the effort of reading the book, there's no alternative but to face it or look for some kind soul who left a good summary on Wikipedia. Or at least that's how it was until a few years ago. For all the George Costanzas out there, perhaps the greatest promise of our marvelous water-thirsty technological wonders is to spare us the displeasure of having to read and understand.

The scene, ripe for a meme, characterizes our age of deceit: from a half-hearted set of instructions, one machine drafts an email so another can break it down into "the important parts." The filler, the formalities, the feigned kindness, from this perspective, barely increase electricity consumption in some datacenter by a tiny percentage. The electronic slaves chat among themselves while we convince ourselves of our undeserved efficiency.

Shamelessly, the tech boys are selling us the ability to summarize as a "superpower." The world was overwhelming before ChatGPT stuck its snout out, and the solution is nothing other than renouncing our higher cognitive capacities in favor of programs that swear to reduce any text to a handful of bullet points capable of making us sound smart.

Automated summaries wreak havoc when they make up facts, but they also subject us to a timid lobotomy when they work perfectly. Maybe it's because many emails really could be summed up in one sentence, or many books are only relevant until the third chapter and then repeat the same ideas until the publisher agrees to publish them. But it's surprising how we convinced ourselves that compression is not only unrestrictedly desirable but must apply to everything we read.

In Claude Shannon's original sense, a message can be compressed if it contains redundancy. But a summary is not compression in that sense: it doesn't eliminate redundancies but decides what matters and what doesn't. Compressing is an operation on structure; summarizing is an operation on meaning. There are texts whose meaning cannot be separated from their form: their power depends on rhythm, accumulation, digressions, and repetitions. In such cases, reducing them doesn't clarify them but replaces them.

The effort of reading, interpreting, discarding the trivial, and synthesizing the fundamental isn't the barrier — it's the very process of learning and understanding. By delegating this task, we explicitly renounce any attempt to comprehend what we're facing. The result before our eyes (the summary) might be the same regardless of how it was obtained, but the result stored behind our eyes will not be.

A recent study sought to empirically measure this phenomenon. Researchers compared people trying to learn about a topic using good old Google — opening tabs, comparing sources, closing ads, etc. — with those who followed a summary regurgitated by an LLM. The former developed superficial knowledge, wrote generic advice about the topic, and felt less engaged with the task. It turns out that the friction inherent in facing a challenge is what "leads to the development of a deeper and more original mental representation of the subject in question," its authors write.

Without active and laborious engagement with reading, the brain discards up to ninety percent of new information within just a week, as Hermann Ebbinghaus noted in 1880. The act of underlining, doubting what we read, and connecting it with what we already know is what encodes knowledge into long-term memory.

Mechanical summaries flatten our mental lives: a loss of individual depth that reflects in a widespread increase of superficial information. As Collin Jennings argues, hyperlinks were never mere navigation tools but a living representation of the architecture of human thought. In his words, following a link is tracing someone else's mental process, seeing how one idea led to another. By replacing that rich and chaotic web with a single block of machine-generated text, voices lose their texture and individuality, reduced to a flat, monotonous tone and, crucially, without sources.

It's thought itself that gets flattened: a vibrant ecosystem of intertwined ideas transforms into a monologue that often sounds like nothing.

If the price for the tool working well is the atrophy of our cognition, the picture becomes even darker when doubt is fundamental. A study from Carnegie Mellon University along with Microsoft became famous for confirming suspicions: the greater the confidence we place in generative AI tools, the lower the user's cognitive effort. Those who surrender their brains in favor of these tools gradually stop forming hypotheses until they begin assuming the tools have validated their sources and simply accept the final summary.

Like the climate crisis or the frog in the pot, the decline in critical thinking won't happen abruptly. Cognitive atrophy dangerously disguises itself as efficiency: we get used to summaries because we move faster, everything looks neater, the day flows more smoothly. It feels comfortable, modern, optimized. But eradicating the friction of thought turns us into operators of systems whose functioning escapes us. The muscle of questioning atrophies.

Another whole collection of problems arises when we remember that hallucinations are most likely impossible to eliminate due to the very nature of LLMs and the transformer architecture, a conclusion reached by researchers from OpenAI itself.

Perhaps it's worth revisiting the philosophical concept of bullshit proposed by Harry G. Frankfurt in 1986 to describe exactly what these models produce. These machines don't lie in the strict sense; they produce bullshit: the liar knows the truth and makes a deliberate effort to hide it. But the bullshit producer, on the other hand, is indifferent to truth. They are "stochastic parrots" or autocomplete systems on steroids, designed to prioritize fluency and user complacency over factual accuracy.

None of this invalidates their everyday usefulness: of course they're useful, that we can no longer imagine ourselves without them, and that we hope they'll improve in the future. In understanding these two opposing issues lies the main challenge of an industry obsessed with LLMs that seems to have forgotten that "artificial intelligence" is a much broader field that emerged way before 2017 with that famous paper.

For anyone paying attention, the consequences of renouncing our reading comprehension are plain to see. Last year a journalist tested five of the most popular bots to analyze complex texts; the results bordered on absolute negligence: the summary of a novel set during the American Civil War completely omitted slavery, and when analyzing legal documents it overlooked critical clauses in rental contracts that gave the landlord the right to enter the dwelling at any hour. Not to mention what happens when we trust Google's AI-generated summaries in its search results on health matters.

Not every use of these tools demands the same level of distrust, but it's also unwise to pretend context neutralizes the problem. Using a machine to catch up on a very long chat thread isn't the same as asking for a summary for a medical exam or signing a business agreement. The difference isn't trivial, even though the underlying logic is the same: delegating cognitive effort to a black box.

Reading doesn't guarantee comprehension, just as producing a summary doesn't guarantee having thought. But part of what we learn when we tackle reading is to make it an active exercise in meaning-making. And the more we read, the easier it becomes.

The consequences of our intellectual laziness threaten the future of the web and deepen its fragility. If we rarely go beyond the message returned by an AI, and its regurgitated inaccuracies seep uncorrected into academic research, corporate decisions, and public documents, it feeds a "disinformation loop" that poisons our collective memory and even endangers trust in any information we might find on the Internet.

We're not going to stop relying on machine-made summaries. They're everywhere, lurking around every corner. Our chances of banning them are much smaller than our chances of changing the perspective from which we read them.

But we can confront them and poke holes in them. Read them with relentless distrust. Look for where they oversimplify, where they omit the uncomfortable, where they sound too sure of themselves. Not to show off, but to reconstruct the path of the original text, the one the summary couldn't follow.

When something strikes us as flawless and convincing, we can be suspicious. By subjecting that synthesis to pressure, we deliberately reintroduce the cognitive friction they wanted to save us from. It's in that clash that the brain finds its best chance to assert itself: not in the passive absorption of informational mush, but in fierce dialogue with what's presented to us until we recover its complexity.

We can treat summaries as a destination or as a starting point.

Or we can just wait for the movie to come out.

This text was originally published in 2026 in Spanish.