Read this fragment: “The day was breezy, so the boy went outside to fly…” Most readers land on the word kite before their eyes even reach it. Editors see a version of this constantly: a sentence that gives away its own ending, so that a reader’s attention has already moved past the words that haven’t arrived yet.
What is less obvious, and what a run of neuroscience over the past two decades has been quietly establishing, is that this is not a quirk of predictable prose. It is what a reading or listening brain does on every sentence, whether the writing is predictable or not.
An old experiment that caught the brain guessing
The clearest early evidence came from a study built around a small, stubborn fact of English grammar: “a” precedes a consonant sound, “an” precedes a vowel sound. Researchers at the University of California, San Diego gave people sentences like “the day was breezy, so the boy went outside to fly a kite,” then swapped in an equally sensible but less expected ending, such as “an airplane.” Recording electrical activity from the scalp, researchers found in 2005 that brain signals shifted the moment readers hit the article itself, before the actual noun had even appeared on the page.
The size of that shift tracked, in a graded way, how strongly readers had expected a different word. The brain was not simply reacting to “airplane” once it showed up. It had already committed, in some partial and revisable way, to “kite,” and the mismatch registered at the article that gave the game away.
A signal that fires ahead of the word, not after it
What that experiment captured is a piece of a larger picture now called predictive processing: language comprehension as a running bet, updated word by word, rather than a passive intake of finished sentences. The brain does not wait for a word to land before deciding what it means. It pre-activates likely candidates based on everything that came before, then measures the gap between what it expected and what it got. That gap, in EEG recordings, shows up as the amplitude of the N400 response, a signal long known to track meaning but shown here to be sensitive to prediction specifically, not just plausibility after the fact.
The story is not as clean as it first looked
Good editing means being honest about what a piece of evidence actually supports, and this literature has its own complicating chapter. In 2018, a consortium of nine laboratories ran a large, pre-registered replication of the DeLong design with 334 participants. Mante Nieuwland reported in eLife that the core effect held for nouns: readers still showed graded surprise tied to how expected a word was. But the more precise claim, that readers pre-activate the specific phonological form of an upcoming word closely enough to register a mismatch at the article “a” versus “an,” did not replicate as cleanly across sites. The broader claim, that comprehension involves ongoing prediction of upcoming meaning, stands on firm ground. The narrower claim, about exactly how far in advance and how specifically the brain commits, is still being argued over.
The same objective driving the software
What makes this more than a footnote for linguists is a separate line of work asking whether the brain’s prediction habit resembles what large language models do mechanically. Researchers led by Ariel Goldstein, working with Uri Hasson’s lab at Princeton and collaborators including Google researchers, recorded electrical activity directly from the brains of patients listening to a 30-minute spoken story, then compared those recordings to what GPT-2 was doing internally on the same story. The results, published in Nature Neuroscience in 2022, identified three principles shared by the human brain and the model: both continuously predict the next word before it arrives, both compare that prediction against the actual incoming word to register surprise, and both represent words using context rather than fixed dictionary meanings. In a 2024 talk covered by Princeton’s Center for Statistics and Machine Learning, Hasson put it plainly: “the human brain is engaged in optimization of an objective which is very similar to the objective we optimize when we train language models.”
Where the comparison stops being useful
Hasson was careful, in that same talk, to note where the resemblance breaks down. A language model learns its predictive habits from a training run across a slice of the internet, billions of words no single person will ever read.
A human learns to predict language from a much smaller, much richer diet: the sentences spoken around one particular child, in one particular life. Hasson’s own lab is now running a project recording infants at home for twelve hours a day across their first thousand days, precisely to see what predictive language processing looks like when it is trained the way people are actually trained, rather than the way a model is.
The shared computational trick, guessing what is coming and adjusting on the gap, does not mean the two systems arrived at that trick by anything like the same route.
What this means for the sentence on the page
For anyone editing sentences for a living, this is not an abstract curiosity. If a reader’s brain is already running a few words ahead, guessing at where a sentence is headed, then a badly built sentence is not just ungrammatical or clumsy. It is one that fights its own reader’s prediction unnecessarily, forcing a costly re-read: a misplaced modifier that points to the wrong noun, a subject buried three clauses after the verb that was supposed to belong to it, a pronoun with no clear antecedent left dangling in the gap between guess and confirmation.
Manuscripts we work on are full of sentences where the writer knows exactly what they mean and the reader’s predictive machinery is quietly building the wrong sentence in the meantime. Good editing, at the level of the individual clause, is largely the work of noticing that mismatch before a reader does: either resolving it, or, when the writer wants the jolt of a genuine surprise, making sure the sentence has earned it rather than stumbled into it by accident.
The boy still goes outside to fly something. Whether it turns out to be a kite or an airplane, the reader’s mind has already placed its bet by the time the word shows up. The job of the person editing that sentence is to decide, deliberately, whether to pay that bet off or spend it on something better.
