Why do I nod through dense writing and close the tab convinced I got it?
Because you are grading the text while it is still carrying you. Smooth transitions, familiar jargon and machine-polished sentences lower the effort of reading without adding one thing you can produce afterward. Your feeling of clarity measures how little friction you met, not what survived the tab closing.
That nod is earned by something real. The paragraph announced its topic, sequenced its clauses and repeated the load-bearing phrase twice. You recognized the arrangement, a different act from building it, and when the tab closes the scaffolding goes with it.
Why does one paragraph take three passes when every word in it is familiar?
Because the cost sits in the sentence's shape, not its vocabulary. Long dependencies, stacked nouns and clauses that hide their subject make you hold more in mind at once. These are processing-cost features, measurable in the text rather than in your head, and they predict how much work a sentence will demand.
A 2026 study built a corpus from 14 language models and measured six of these features: low-frequency word combinations, dependency length, object relative clauses, stacked noun compounds, denominal verbs, and inanimate subjects with animate verbs (the processing-profile analysis). All nine frontier and near-frontier models used more inanimate subjects with animate verbs and more low-frequency word combinations than human writers did, and most also produced longer dependencies and more stacked nouns. The authors trace the pattern to training on maths, code and reasoning, where density beats clarity, and support it by fine-tuning a model under a length penalty and watching its usage drift from human norms.
The boundary travels with the result. They measured features linked to processing difficulty, not reader outcomes, so nothing here tests whether people understood it less well.
Does a high readability score mean you understood the page?
No. Readability formulas were built to predict comprehension test scores from sentence length and syllable counts, and they turn out to be poor predictors of how hard a text actually feels to read. A clean score tells you the sentences are short, not that you could rebuild the argument with the page closed.
The clearest test used eye tracking, which measures reading as it happens instead of asking people to rate a text afterward (the 2025 readability evaluation). Traditional formulas, NLP scoring methods, commercial systems and frontier models were all poor predictors of real-time reading ease in adults, and simple word-level properties from psycholinguistics did a better job.
So shortening is the wrong repair. If the difficulty lives in syntax, halving sentences leaves the tangled ones tangled. Rewrite the sentence that names an actor, an action and a consequence, then say it back.
Why do people at work use fancy words when nothing specific has happened?
Sometimes they are compressing shared meaning, and a room that already knows the context moves faster in that dialect. Often nobody owns the meaning, so the sentence sounds like a plan without naming an actor, an action or a cost. The tell is whether anyone can translate it into a decision.
The second habit has a measurable cost. In experiments, replacing words in graduate admission essays with their longest thesaurus synonyms lowered the author's perceived intelligence, and simplifying a long-winded dissertation abstract raised it. Processing fluency mediated the effect (Oppenheimer, Applied Cognitive Psychology, 2006), so the words did not display intelligence, they blocked the reader's access to it.
That reframes the office problem. Long words are not a signal of expertise, they are a tax paid by whoever joins late, and it lands hardest on the person with the least standing to object.
The cheapest move is asking once, out loud: I am not following, can you tell me what this means in a sentence. Do it in private, then write the translation down, because a decision recorded in plain words survives the meeting.
Why is my job turning into arguments with someone else's chatbot?
Because the reply arrives faster than the understanding behind it. Colleagues forward model output they have not restated, so you end up negotiating with text nobody in the chain can defend. The failure is not the tool but a document that left the building before anyone could say what it decided.
An analysis of 5.1 million pages from 71 Common Crawl snapshots estimates that about 24 percent of crawlable pages contained AI-generated or AI-edited text by August 2026, up from near zero before 2022 (the share estimate). The growth is uneven: 32.8 percent of English-only pages against 17.0 percent of non-English pages, and only about 6 percent on .edu and .gov domains. It counts pages a model generated or revised, sees only crawlable pages, and is a preprint, so other denominators are not comparable.
The practical version is smaller. Any document nobody in the chain restated is one nobody can defend, and when the decision goes wrong, the argument is about whose model wrote the sentence instead of what the page decided.
How do you catch yourself not understanding, mid-page?
Close the page early, on purpose, and say the last paragraph out loud in your own words. Fluency cannot survive that test, because production has no shortcut. Do it at every section break, not at the end, when the whole structure feels obvious in hindsight.
Keep three lines per section: the claim in your own words, the strongest reason, and the question still open. The same check applies to a document you send, because if you cannot restate your own summary, it is not finished.
Absorb is built around this. Short daily sessions of reading, listening and visual material ask you to answer from memory instead of rating how clear the page felt, and the assessment gives you three separate scores. If your gauge has been telling you that smooth text is understood text, see what you actually absorbed with Absorb on the App Store.