Why AI-generated text is so irritating to read

You know the feeling. You open an email, a blog post, a report from a colleague, and within two sentences something in you sags. Nothing in it is wrong. It is grammatical, organised, polite, even a little elegant. And yet you find yourself skimming, then scrolling, then quietly resenting whoever sent it. People describe it as "hollow", "sludgy", "like eating cardboard". Most of us cannot say exactly what triggered the reaction, only that it is real and getting more common. The research on how the brain reads gives a surprisingly clear answer, and it comes down to one thing.

The short version: Reading is a social act. Your brain does not treat text as information to be decoded; it treats it as a person to be understood. The same neural circuitry you use to work out what a friend across the table is thinking lights up when you read a paragraph, because a paragraph has always meant that someone, somewhere, wanted to tell you something. AI text is irritating because it activates that circuitry perfectly and then leaves it with no one to find. You turned up for a conversation and the other chair is empty. Every other reason below is a variation on how that emptiness gets delivered.

Reason one: your brain is looking for a mind, and there is none there

Neuroscientist Raymond Mar spent years mapping what happens in the brain when people read stories, and the finding that keeps reappearing is that comprehension runs on the "mentalizing" network, the regions of the medial prefrontal cortex and temporoparietal junction that we use to infer other people's beliefs and intentions. Reading does not borrow this network occasionally; it depends on it. Developmental work by Tauzin and Gergely shows that even preverbal infants interpret an exchange of signals as one person handing information to another. We arrive wired to assume a communicator.

This matters because most of what makes a piece of writing land is not on the page at all. It is what you infer about why the writer chose this word rather than that one, why they stopped where they stopped, what they were worried you would misunderstand. When a human wrote it, those inferences pay off. When a model wrote it, they lead nowhere, because no choice was made in the sense your brain is looking for. You keep reaching for the intention and closing your hand on air. The technical name for the sensation is hard to pin down, but everyone recognises it: the text is talking at you rather than to you.

Reason two: it breaks the bargain that makes reading worthwhile

Dan Sperber and Deirdre Wilson's relevance theory, the standard account of how communication works, says that every deliberate utterance carries an implicit promise: what you get out of this will be worth the effort of processing it. That promise is what lets you trust a stranger's sentence enough to read it to the end. Relevance is a ratio, cognitive reward divided by cognitive cost, and good writers keep the ratio high by leaving out everything the reader does not need.

Generic AI prose runs the ratio into the ground. In 2025, Shaib, Chakrabarty, Garcia-Olano and Wallace tried to measure what people mean when they call text "slop". They had nineteen experts build a taxonomy of the problem, then had professional copy-editors apply it to hundreds of passages. The strongest predictors of a slop verdict were low relevance and low information density. And the one feature every annotator agreed on, regardless of domain, was verbosity: the text simply said less per sentence than a reader has a right to expect. So when you feel cheated by a page of smooth, on-topic, well-formed prose, you are not being fussy. You paid full attention and got a fraction of the usual return, and the brain has a word for that transaction, which is irritation.

Reason three: your brain runs on surprise, and the text has none

Underneath everything, the brain is a prediction machine. It guesses the next word before it arrives, and it learns from the difference between the guess and the reality. Neuroscientists can watch this happen: the N400 brain response grows larger the less predictable a word is, and recent studies show that the surprise value computed by a language model is the best single predictor of that response. Prediction error is not a bug in reading. It is the mechanism by which reading feels like anything.

Human writing feeds that mechanism naturally. It is bursty: a long, winding sentence followed by a blunt one; three paragraphs on a single detail, then a jump. Altmann and Piantadosi documented this clustering in ordinary prose years before anyone worried about chatbots. Language models, by construction, do the opposite. They are trained to produce the most likely continuation, which places their output at the statistical centre of everything ever written. Sentence lengths flatten. Surprise flattens. Researchers describe the resulting probability curves as unnaturally smooth, and readers describe the resulting prose as something they can read effortlessly but cannot stay awake through. Both are describing the same thing: a prediction engine with nothing to correct. That is the biology behind the strange complaint that AI text is "easy to read but impossible to get through".

Reason four: the effort is upside down, and that feels like disrespect

In 2011 Buell and Norton published a series of experiments on what they called the labour illusion. People value a result more when they can see the effort that went into it, and the reason is reciprocity: someone worked on my behalf, so I owe them attention in return. The same logic explains why handmade goods sell for more and why a two-line note from a friend can mean more than a printed card.

AI text turns the exchange rate inside out. The sender spent a minute prompting; the reader spends twenty minutes trying to extract what was meant. When Stanford's Jeffrey Hancock and BetterUp Labs surveyed 1,150 American desk workers in September 2025, four in ten had received what the study called "workslop" in the previous month. Fifty-three percent said it annoyed them and twenty-two percent said it offended them. On average, each piece took nearly two hours to sort out, about twenty minutes longer than if the sender had simply done the task. Forty-two percent trusted the sender less afterwards. Read those numbers together and the emotion is unmistakable. This is not the frustration of reading something badly written. It is the feeling of being handed someone else's unfinished homework and being expected to be grateful.

Reason five: the polish has become the tell

Ease of reading is usually a point in a text's favour. Psychologists call it processing fluency, and decades of work show that fluent statements are judged truer and more likeable. But fluency only works as a signal because, historically, it was hard to fake. Once smoothness comes free, it stops meaning "skilled" and starts meaning "automated".

The evidence that this flip has happened is striking. Raj, Berg and colleagues ran sixteen preregistered experiments with more than 27,000 participants and found that simply believing a piece of writing was AI-generated lowered people's rating of it every single time, that the effect ran through a drop in perceived authenticity, and that no intervention they tried could remove it. Mandel and Imas found the same pattern in art auctions with real money on the table: aversion was nonlinear, with an outsized penalty for even a trace of AI involvement. The psychology is closer to contamination than to quality judgment. Once you suspect a machine, the very symmetry and completeness that used to impress you become evidence against the text, and you start reading for tells instead of meaning.

Reason six: you have learned the pattern, and the brain hates a pattern

The last reason is the most familiar and the most mechanical. The brain economises on anything it has seen before; repeated stimuli get down-weighted so that attention can go where it is needed. Say a word enough times and it briefly loses its meaning, an effect called semantic satiation that shows up as a shrinking N400 in the lab. AI prose is unusually rich in repeated stimuli. Kobak and colleagues analysed fifteen million scientific abstracts and found a sudden, post-ChatGPT surge in a small vocabulary of style words, enough to conclude that at least one abstract in eight in 2024 had passed through a model, and closer to two in five in some fields.

Readers learn these fingerprints within weeks. The lists of three. The cautious hedge on every claim. The answer that arrives as bullet points. The closing paragraph that restates the opening. Each one, once recognised, gets processed as a template rather than a message, and recognising a template while trying to read for meaning is exactly the kind of interruption the brain finds aversive. There is a quieter loss underneath it. Sourati and colleagues, analysing 880,000 texts, found that heavy model use compresses the variety of human writing by between a fifth and a half, and strips away the cues that tell you a writer's age, values and worldview. Part of the pleasure of reading has always been the contact with a specific other person. Homogenised prose removes the person, and what remains is the sensation of reading the same anonymous author for the thousandth time.

Putting it together

Set the six reasons side by side and they stop looking like a list and start looking like one event seen from different angles. Your mentalizing system engages and finds no mind. The relevance bargain is broken, so effort exceeds reward. The prediction engine goes hungry because nothing surprises it. The effort ran in the wrong direction, which reads as contempt. The polish that would once have reassured you now makes you suspicious. And the whole thing arrives in a pattern you recognise before you have finished the first paragraph. No single one of these would be enough to explain the strength of the reaction. Together they describe why a page of competent text can produce something close to anger.

It also explains why the fix is not stylistic. Removing the tells helps at the margin, and so does cutting length, but what readers are actually reacting to is the absence of a person with something at stake. Text stops being irritating the moment it becomes obvious that someone meant it: a specific claim, a detail that could only have come from being there, a judgment the writer might be wrong about. Give the reader a mind to model and most of the rest takes care of itself.

A note on the limits of the evidence: The mechanism is robust; the boundaries will keep moving. The link between low surprise and measurable reading effort is well established in the lab. The further link to boredom and irritation is strongly supported by the slop research and by the workslop survey, but it has not been isolated in a single clean experiment. Statistical detectors based on perplexity no longer reliably separate human from machine text, so much of the reaction is driven by suspicion rather than certainty, and Shaib's team are explicit that humans can write slop and models can avoid it. Agreement between people on what counts as slop is only moderate, and the penalty for disclosed AI use shrinks among readers who use these tools heavily themselves.