Predictive language processing is the use of linguistic context, prior experience, and knowledge of the world to establish expectations about language that has yet to arrive. As someone listens to a sentence or reads a paragraph, the meaning of earlier words changes which interpretations and continuations are likely. This can help the language system recognize incoming material efficiently, allocate attention, and update an unfolding interpretation. A reader encountering “She spread the warm bread with…” can anticipate a food-related continuation even before seeing another word. The strength, specificity, and consequences of that anticipation are measurable questions in psycholinguistics.
Research establishes robust relations between contextual predictability and word-processing behavior. More predictable words are often read faster, skipped more often, and associated with smaller N400 responses in electrophysiological experiments. These observations, however, leave open which computations cause the advantage: anticipatory activation of a particular word, preparation of semantic features, efficient retrieval after the word arrives, probabilistic structural interpretation, or interaction among these processes. Contemporary reviews emphasize that prediction is a family of possible mechanisms rather than a single observation or a universal instruction executed identically on every word. Kuperberg and Jaeger (2016); Pickering and Gambi (2018); Staub (2025).
This article examines the scientific mechanisms and evidence for prediction during human language comprehension: what can be anticipated, how expectations are quantified, what eye movements and brain potentials reveal, why some celebrated findings have not replicated, and how different models explain the same data. It also examines conversational context, reading and listening, individual variation, and statistical language models, while preserving the distinction between a model's predictive accuracy and the cognitive process that generated human behavior.
What Is Predictive Language Processing?
Prediction as a graded expectation
Every word in a sentence narrows or reorganizes a distribution of possible continuations. “The gardener watered the…” makes some nouns more likely than others; “The gardener planted the…” changes the distribution. Prediction in its broad statistical sense refers to those context-sensitive probabilities. A strong prediction may favor one particular word, while a weak prediction may favor a category of objects, an event type, a grammatical construction, or several alternatives without selecting any single winner. Human comprehension usually works with uncertainty, so a useful description represents expectations as graded rather than all-or-nothing guesses.
The term anticipatory processing is more specific when it identifies an operation occurring before a relevant signal arrives. For example, the word “eat” may direct a listener's gaze toward the only edible object in a displayed scene before the noun naming it is spoken. By contrast, an unexpectedly long reading time observed after a word appears is evidence that context affected processing; it cannot, by itself, locate the effect before the word. Distinguishing an expectation from a demonstrated pre-onset operation is essential to causal interpretation. Altmann and Kamide (1999); Huettig (2015).
Predictability, anticipation, facilitation, and integration
Predictability is a property assigned to a continuation relative to a preceding context and to a specified probability-estimation method. Anticipation describes how a cognitive system may use that distribution before new information appears. Facilitation describes the processing benefit observed when an incoming word is compatible with contextual expectations. Integration refers to constructing or updating a coherent representation after encountering the material. These operations are related but not interchangeable. A highly plausible word can be easy to integrate without having been individually pre-activated, and a strong expectation can coexist with successful comprehension when a different word arrives.
Researchers therefore separate the question “Does context affect processing?” from “Did the comprehender represent this specific feature in advance?” and from “Was a particular prediction necessary for the observed speed advantage?” The first has broad empirical support. The second depends on the level of representation and the method. The third requires a theory and evidence able to distinguish anticipatory preparation from alternatives. Van Petten and Luka (2012); Wong, Reichle, and Veldre (2025).
An evolving interpretation rather than a stream of isolated words
Human language comprehension is incremental. Listeners combine an emerging sound sequence with knowledge of vocabulary, syntax, meaning, discourse, and the immediate situation. Readers continuously update the structures associated with prior text. Each new item can confirm a favored interpretation, revise a low-level assumption, introduce a new referent, or change which events are plausible. A prediction at one moment is therefore conditional on a provisional interpretation that may itself be revised. Anticipation is embedded in ordinary comprehension rather than functioning as a separate guessing game.
A conversation offers much richer evidence than a single sentence: who is speaking, what has already happened, which objects are visible, and what an utterance is intended to accomplish. In Dutch spoken-story experiments, expectations created by discourse affected processing at an adjective whose grammatical marking was inconsistent with a strongly anticipated later noun. The effect appeared before the anticipated noun, supplying a more direct form of evidence for preparation than a difficulty measured only when the noun was encountered. Van Berkum and colleagues (2005).
What Can the Mind Anticipate?
Semantic features and event knowledge
An upcoming word has multiple kinds of information associated with it. A context about baking may activate ingredients, utensils, actions, and relevant events. The upcoming expression may be unexpected as an exact word yet share semantic features with the favored continuation. Experiments comparing an expected noun with an unexpected noun from the same semantic category and an unexpected noun from another category found graded neural responses sensitive to this relationship. This supports the possibility that contextual expectations recruit knowledge structured by meaning, while leaving room for accounts involving semantic retrieval after the new item appears. Federmeier and Kutas (1999).
Semantic expectancy can concern a theme or role as much as a lexical item. After hearing that a child is eating, a listener may anticipate an edible patient of the action rather than any exact noun. In a visual setting containing one edible object, that anticipation can translate into selective attention toward the appropriate referent. The representation guiding the gaze might include semantic suitability, a prospective referent, or a lexical label; one gaze result rarely distinguishes all three. Altmann and Kamide (1999).
Lexical identity and word form
A stronger proposal holds that a listener can prepare a specific lexical candidate, potentially including features of its sound or spelling. English offers a clever experimental test: the article “a” typically precedes a consonant sound, and “an” precedes a vowel sound. If an unfinished sentence strongly suggests “kite,” a response to the preceding article might reveal whether the sound form of “kite” was anticipated before it appeared. DeLong, Urbach, and Kutas (2005) reported an article-level ERP effect that became a landmark result in the debate about word-form prediction.
The evidential picture changed after a registered direct replication across nine laboratories with 334 participants. The larger project reproduced the expected relation between noun cloze probability and noun N400 amplitude but did not reproduce the originally reported article effect with the planned analyses. Exploratory analyses suggested that any article effect was smaller than the original estimate. The appropriate conclusion is specific: noun predictability affects processing reliably, while routine pre-activation of the exact phonological form of a future noun is not established by this paradigm. The replication did not show that language users never anticipate word forms. Nieuwland and colleagues (2018).
Syntax, grammatical category, and thematic structure
A sentence can support expectations concerning syntactic category, agreement, case marking, argument structure, or how a verb's participants relate to each other. These are predictions about linguistic organization, sometimes without any single expected word. In English and Japanese visual-world experiments, grammatical and semantic information from earlier constituents guided anticipatory gaze to participants in described events, including cases where the relevant verb had not yet appeared in the Japanese sentence. The result demonstrates that word order and information distribution influence when expectations become available. Kamide, Altmann, and Haywood (2003).
A grammatical expectation can be misleading. Readers sometimes settle temporarily on a likely syntactic analysis and later need to revise it. Probabilistic accounts represent this as changes in the distribution over possible parses. Structural revision is related to predictive processing, but garden-path interpretation has its own experimental problems concerning attachment preferences, ambiguity, and reanalysis; the present focus is the broader way contextual probabilities constrain comprehension. Levy (2008).
Discourse reference, pragmatics, and speaker information
Language prediction extends beyond word combinations learned in isolation. Mentioning a particular object makes later references to that object more accessible. A contrastive conversation may make an otherwise unlikely adjective important. A question establishes which information is sought; shared experience changes which descriptions will be informative. A comprehender can therefore assign different probabilities to the same word depending on the speaker, communicative goal, and discourse history. A tightly controlled sentence-completion task captures only part of that contextual richness. Van Berkum and colleagues (2005); Kuperberg and Jaeger (2016).
Pragmatic expectations can be uncertain or socially negotiated. A listener may anticipate what a speaker is likely to say while maintaining several plausible interpretations of why the speaker is saying it. The scientific question is how such information affects immediate linguistic processing, with conversational repair and interpersonal skills treated as neighboring domains. Broad conversational dynamics also depend on attention, social cognition, and interaction history.
How Scientists Measure Language Predictability
Cloze probability: completing an unfinished context
A cloze task presents a sentence fragment and asks a sample of people to supply a natural next word. The proportion who produce a particular continuation is its cloze probability for that sample and instruction. If 70 of 100 respondents supply “butter,” the estimated cloze probability of “butter” is 0.70. The number describes the distribution of overt completions, not a direct reading of each respondent's neural state. The response may also reflect response strategies, grammatical conventions, population-specific knowledge, and the language used in the experiment.
Cloze studies remain useful because they reveal how strongly a specific context constrains a possible completion and enable reproducible manipulation of expectancy. Their limits are important. Participants asked to explicitly complete a sentence have time to generate a deliberate response, whereas comprehension in ordinary speech may be fast and partly implicit. Cloze data typically provide incomplete sampling of less frequent possibilities, and results may vary with the participants' linguistic background. Treating a cloze norm as an exact probability for every individual listener overstates its precision. Wong, Reichle, and Veldre (2025).
Surprisal: the information carried by an arriving word
Surprisal formalizes how unexpected a word is under a probability model. For a word w after context c, its value is S(w|c) = −log2 P(w|c), commonly measured in bits. A continuation with probability 0.5 carries one bit of surprisal; a continuation with probability 0.125 carries three bits. The scale increases as modeled probability decreases. These simple numbers explain why a change from a probability of 0.50 to 0.25 has the same one-bit increase as a change from 0.25 to 0.125.
The probability is conditional on the model and the context representation. A language model trained on a large corpus may estimate one value, while a person with specialized experience and a different interpretation of the sentence assigns another. Surprisal is consequently a formal predictor of processing outcomes, not a recording of subjective surprise or proof that neurons calculate a particular logarithm. The theoretical achievement of expectation-based accounts is that they produce quantitative predictions testable against reading times and other measurements. Levy (2008); Smith and Levy (2013).
Entropy: uncertainty about what comes next
Surprisal is assigned to an actual continuation; entropy summarizes uncertainty across the possible continuations before one occurs. A distribution with a dominant candidate has comparatively low entropy. A distribution spread evenly across many alternatives has higher entropy. Entropy and surprisal can therefore vary independently: a reader might be uncertain before a word appears, and the eventual word might nevertheless be moderately probable; alternatively, the reader might be highly confident and then encounter an extremely improbable word.
This distinction matters when interpreting neural measurements. Studies of naturalistic story listening have related computational estimates of both next-word entropy and realized-word surprisal to brain activity, and these measures showed differing regional patterns in one fMRI investigation. Such correlations support research on uncertainty and contextual probability while requiring caution about attributing a unique computation to a brain region. Willems and colleagues (2016).
Three different probability estimates
Cloze probability measures what participants overtly complete. A statistical corpus model predicts from language distributions learned during training. An experimental condition may instead manipulate the event structure or visible scene in ways that change plausible continuations. These measures often correlate, but they are not interchangeable. A reader can strongly expect an object because it is physically visible even when its written label is uncommon; a corpus model may assign a high probability to a frequent phrase that is pragmatically odd in the specific conversation. Good research states which probability is being measured and tests what cognitive operation that estimate could plausibly represent.
The Main Experimental Evidence
Anticipatory eye movements during spoken language
The visual-world paradigm records eye movements while a person listens to speech and views a scene. In a classic study, participants heard sentences about a boy who would either eat or move an object. When the scene included one edible object, listeners looked toward it earlier after the restrictive verb “eat,” sometimes before hearing the object's name. The manipulation is informative because the emerging language changes the set of plausible referents while the visual scene remains available. The result offers evidence of prospective allocation of attention as speech unfolds. Altmann and Kamide (1999).
The interpretation remains method-dependent. Visible candidate objects constrain the search space; eye movements take time to plan and execute; and attention may be guided by semantic affordances without activation of an exact word form. Related experiments in English and Japanese extend the findings to different orders of information, suggesting that anticipation adapts to language structure. The visual-world evidence strongly supports anticipatory use of context under appropriate conditions, while the representational target of the anticipation must be established separately. Kamide, Altmann, and Haywood (2003); Wong, Reichle, and Veldre (2025).
Reading-time and eye-tracking effects
Contextual predictability is a robust determinant of reading behavior. Predictable words tend to receive shorter fixations and are more likely to be skipped than less predictable words, though the size of the effect depends on word length, frequency, position, parafoveal information, the surrounding text, and the reading task. Eye-tracking measures can distinguish early from later processing effects and reveal regressions when an interpretation needs reconsideration. Their temporal richness is valuable, but skipping a word is not equivalent to already knowing its identity. Wong, Reichle, and Veldre (2025).
In analyses spanning large reading datasets, Smith and Levy identified a broadly logarithmic relation between estimated word probability and reading time. The finding matters because even low-probability words were sensitive to relative changes in probability, rather than predictability affecting only the few words that people could confidently guess. The result supports formal probabilistic models and challenges a narrow picture in which comprehension prepares one favored word and ignores the remainder. It does not uniquely identify whether processing costs arise from pre-activation, information uptake, or a combined system. Smith and Levy (2013).
The N400: contextual meaning sensitivity
The N400 is an event-related potential, typically observed as a negative-going component whose amplitude is sensitive to semantic and contextual properties of a stimulus. It was identified in pioneering experiments comparing meaningful with incongruous sentence endings. Since then, N400 responses have been observed for a broad range of linguistic and nonlinguistic materials, including spoken, written, and signed language, pictures, and meaningful environmental stimuli. The component provides a temporally precise index associated with access to and processing of meaning. Kutas and Hillyard (1980); Kutas and Federmeier (2011).
An expected word generally elicits a smaller N400 than an unexpected word in a supportive context. Unexpected words that are semantically related to an anticipated continuation can yield intermediate effects. N400 amplitude is thus informative about the interaction between context and incoming semantic information, but several cognitive theories can explain a reduced response. It is an error to interpret each N400 difference as a direct readout of a prediction error signal or as a demonstration of word-form pre-activation. Federmeier and Kutas (1999); Kutas and Federmeier (2011).
Earlier and later brain responses
Neural responses outside the N400 window also contribute. Some paradigms report differences at earlier latencies that may reflect perceptual or lexical preparation. Later positive ERP components can occur when a plausible but unanticipated continuation requires a change in an expectation, or when an expression violates a strong contextual or structural constraint. The scalp distributions and functional interpretations of these positivities vary. It is therefore useful to distinguish facilitation of expected information from additional processing prompted by a disconfirmed expectation. The same incoming word may evoke both a semantic facilitation pattern and later reinterpretation depending on context. Van Petten and Luka (2012); Brothers, Swaab, and Traxler (2015).
Neither timing nor scalp location directly determines the psychological function of an ERP component. A finding becomes more persuasive when targeted manipulations distinguish alternatives, a predicted effect replicates, and independent behavioral measures tell a compatible story. Brain recordings add important temporal detail to the problem of expectation, but the interpretation depends on experimental design and comparison among models.
Natural speech, stories, and ecological validity
Experiments that present isolated sentences one word at a time allow precise control over stimulus timing, but comprehension in daily life unfolds amid prosody, gesture, visual context, and an evolving topic. Researchers increasingly study uninterrupted stories, natural reading, and continuous speech while fitting computational measures of probability to each encountered word. Such studies reveal statistical relations between expectation and neural or behavioral variation outside the most tightly scripted laboratory conditions. They also introduce additional complications, including correlations among lexical frequency, sentence position, information structure, acoustic prominence, and semantic coherence. Frank, Otten, Galli, and Vigliocco (2015); Willems and colleagues (2016).
A naturalistic correlation is valuable when it predicts new material and survives controls for plausible confounds. It becomes stronger when it is combined with experiments that selectively manipulate relevant cues. Neither high ecological realism nor perfect control alone guarantees a faithful account of language understanding. The broader research program links them by testing the same computational claim across distinct tasks, groups, languages, and measurements.
Competing Explanations of Prediction Effects
Predictive pre-activation accounts
Pre-activation accounts propose that earlier context activates features or candidate representations before the corresponding input is received. The activation may involve likely meanings, syntactic categories, morphological properties, phonological features, or specific lexical items. If the future word matches the prepared information, recognition or comprehension can be facilitated. These accounts are particularly persuasive when an experimental effect appears before the disambiguating word, as in certain anticipatory gaze and grammatical-cue studies. The amount of advance information prepared is an empirical parameter, rather than an assumption about total advance knowledge. Kuperberg and Jaeger (2016).
Different pre-activation theories make different predictions about uncertainty. One model may prepare many candidates with strengths weighted by probability; another may devote resources primarily to a single high-probability continuation; a third may prepare abstract event features while leaving exact lexical forms unselected. These models can agree that expected words are easy yet disagree sharply about costs when the expected word fails to occur. Experiments testing unexpected but plausible alternatives can discriminate among them. Frisson, Harvey, and Staub (2017).
Facilitation through semantic memory and integration
Context can make an encountered word easier to access or incorporate without requiring advance representation of its identity. On one account, semantic information activated by earlier words changes the accessibility of related concepts in long-term memory. On another, a coherent continuation is easier to integrate into an existing event or discourse representation. The central empirical challenge is that both mechanisms may produce the same direction of reading-time or N400 effects as active prediction. A theoretically decisive experiment must identify observations that one account expects and the other does not. Kutas and Federmeier (2011).
A difficult sentence ending may cause retrieval, evaluation of plausibility, revision of an interpretation, or several processes at different times. These possibilities cannot be collapsed into the generic phrase “the brain was surprised.” Work on semantic illusions and later positive components illustrates why different ERP signatures are assigned different functional roles and why the debate over lexical retrieval versus semantic integration remains active. Brouwer, Fitz, and Hoeks (2012).
Expectation-based and surprisal theories
Expectation-based theories model comprehension as probabilistic inference over linguistic continuations or structures. Encountering a low-probability word forces more extensive updating or information processing than encountering a high-probability word. A distinctive strength is quantitative precision: processing cost can be predicted from the logarithm of conditional probability and tested across naturally occurring distributions of words, not merely against a contrast between “predictable” and “unpredictable.” Levy (2008); Smith and Levy (2013).
Surprisal models differ in how they obtain probabilities, whether their representations are syntactic, lexical, semantic, or distributed, and what kind of cognitive resource they associate with information update. A strong fit between surprisal and reading times does not establish that the model's internal representations or learning procedure are psychologically identical to those of readers. A 2025 synthesis evaluates the promise of surprisal alongside questions about nonlinearity, prediction measures, individual differences, and residual variance unexplained by current models. Staub (2025).
Prediction by production and forward modeling
Another proposal is that comprehension recruits processes related to speech production. A listener may partly simulate what a speaker would say next, using linguistic production knowledge to prepare an upcoming meaning or form. This family of theories has an intuitive connection to the fact that fluent speakers possess detailed experience of selecting words and constructing utterances. Pickering and Gambi develop a framework in which prediction by production may be especially effective under conditions supporting a simulation of the speaker's communicative intention. Pickering and Gambi (2018).
A 2024 review examines production-related prediction accounts and studies involving auditory perception, speech production, and atypical language populations. Its synthesis helps specify what kind of relation between production and comprehension should be experimentally tested. The evidence does not establish that every comprehender must internally articulate each predicted word, or that motor simulation is obligatory for comprehension. The role of production mechanisms, the levels on which they contribute, and their dependence on task and individual resources remain theoretically contested. Gastaldon, Bonfiglio, Vespignani, and Peressotti (2024).
Hierarchical predictive coding and error correction
Predictive coding theories in perception often describe hierarchical systems that send predictions downward and use differences between predicted and incoming input to update internal representations. They are sometimes extended to language, where higher-level representations of discourse, syntax, or meaning might constrain lower-level processing. This is a theoretical family with multiple formal versions, not one uniquely established neural architecture for sentence comprehension. To evaluate a predictive-coding account of language, investigators must specify the units predicted, the temporal dynamics, the form of an error signal, and how the proposal differs from a general probabilistic theory. Kuperberg and Jaeger (2016).
The wider assumptions and evidence concerning sensory inference are examined in the published Predictive Processing in Perception: Theory, Evidence, and Debate. Language processing contributes an unusually fine-grained setting for testing such principles, but a finding about word probability cannot automatically verify an entire theory of perceptual hierarchy. The relation between linguistic prediction, attention, and sensory inference must be determined at the explanatory level appropriate to each dataset.
When Predictions Fail: The Scientific Importance of Unexpected Words
A surprise is not necessarily a comprehension failure
An unexpected ending can be perfectly grammatical and sensible. “For breakfast, she drank warm water” may violate a reader's preferred prediction of “coffee” without making the sentence incoherent. Another ending may be both unexpected and semantically anomalous. The distinction matters: a violation of likelihood, a violation of plausibility, and a violation of grammar can cause different processing demands. Well-designed experiments compare these properties rather than treating every low-cloze word as the same kind of event. Van Petten and Luka (2012).
A theory in which the comprehender commits to a single word often predicts an additional cost when that word fails to arrive. A theory maintaining several graded candidates may instead predict a smooth dependence on probability or semantic similarity. In eye-tracking studies of reading, Frisson and colleagues found robust predictability effects but did not observe a general additional penalty on an unexpected word merely because a strongly expected alternative was available. Their findings are consistent with broadly distributed activation and caution against equating a predictability benefit with an obligatory discrete prediction-error penalty. Frisson, Harvey, and Staub (2017).
What the DeLong replication changed
The 2005 article–noun ERP result became influential because it seemed to offer unusually direct evidence that the specific phonological form of an unseen noun had already been prepared. The 2018 multicenter replication separates the robust part of the empirical phenomenon from the stronger interpretation. In 334 participants across nine laboratories, noun N400 amplitude varied with cloze probability, while the preregistered article analysis did not show the predicted original pattern. An effect at the actual noun thus cannot be taken as sufficient evidence that its form was predicted before noun onset. DeLong, Urbach, and Kutas (2005); Nieuwland and colleagues (2018).
This example also illustrates why a failed replication needs precise language. The study challenged the magnitude and reproducibility of the article-based phonological prediction effect. It did not erase visual-world anticipatory evidence, semantic expectation effects, or every possible context in which phonological preparation might occur. The scientific update is a narrower claim about routine word-form pre-activation and the experimental conditions under which it can be detected. A serious model must accommodate both reliable contextual facilitation and the absence of a particular pre-onset effect in a sufficiently powered replication.
What a late positivity may indicate
A word can be comprehensible yet force a reader to abandon a favored continuation. Later positive ERP components have been interpreted as reflecting consequences of violated expectation, updating, attention, or reanalysis in different circumstances. These signals may provide information about processes distinct from the semantic access associated with N400, but there is no one-to-one mapping from a positivity to a single mental operation. The size and distribution of the response depend on sentence meaning, plausibility, and experimental design. Van Petten and Luka (2012); Brothers, Swaab, and Traxler (2015).
Prediction in Reading, Listening, and Conversation
Reading: the interaction of lexical and visual information
During reading, predictions can influence which words are recognized quickly and where the eyes move next. Yet eye movements are shaped by both language and the visual properties of the page. Parafoveal preview gives readers partial information about words that have not yet been fixated, and word length influences skipping even when sentence context is strongly constraining. A prediction account must therefore model the joint influence of linguistic expectancy, visual accessibility, and oculomotor control. The most recent extensive review of reading prediction analyzes cloze, fixation durations, skipping, information-theoretic models, and methodological disagreements in this combined setting. Wong, Reichle, and Veldre (2025).
The reading case is especially informative because probability can be manipulated independently of visual word form in carefully constructed comparisons. It is also unusually vulnerable to hidden confounds when naturally occurring words differ in frequency, length, orthographic neighborhood, and position in the sentence. Readers may adopt different strategies under comprehension questions, proofreading instructions, or speeded tasks. Accordingly, a single result from a particular reading paradigm should not be generalized to all natural conversation without supporting evidence.
Listening: rapid adaptation to a continuous signal
Spoken words arrive over time, and listeners begin processing them before their sound forms are complete. Speech rate, coarticulation, prosody, background noise, and talker familiarity influence what incoming information is available. Context can restrict the set of plausible referents and linguistic continuations, sometimes enabling gaze toward a relevant object before the anticipated noun arrives. Because the acoustic signal itself unfolds continuously, researchers must separate anticipation driven by prior context from emerging recognition of a partially spoken word. The placement of the experimental measurement relative to word onset is crucial. Altmann and Kamide (1999); Huettig (2015).
Prediction may become useful when part of the signal is uncertain, but it can also mislead if contextual expectations dominate weak sensory evidence. Understanding speech in noise depends on perception, attention, linguistic knowledge, and listening conditions. The related published work Selective Attention: How the Mind Filters Information addresses attention as its own construct; a difficulty following conversation in noise should not be interpreted as a clinical language impairment on the basis of prediction theory alone.
Conversation: anticipating an utterance in a shared situation
Everyday dialogue involves changing goals and shared knowledge. Listeners use a speaker's previous statements, topic, syntax, and likely communicative intention to construct expectations. A person asked “Where did you put the keys?” may anticipate a location phrase. A speaker who sees only one red cup in the room may need a different description from someone facing several cups. Such expectations can influence comprehension before a sentence is complete, although the precision of a predicted word varies greatly. The laboratory evidence for immediate use of discourse context shows how a broader story can affect processing of a grammatical cue before the expected noun is produced. Van Berkum and colleagues (2005).
Successful conversation also requires flexibility. Speakers change topics, use irony, repair misunderstandings, and reveal information the listener did not know. An effective comprehension system must accommodate surprising continuations while maintaining coherence. This flexibility helps explain why graded expectations are often a more useful theoretical starting point than imagining that the listener possesses a single fixed script for the speaker's next sentence.
Learning, Experience, and Individual Differences
Linguistic experience shapes probability distributions
Predictive expectations depend on the regularities a person has encountered. Vocabulary, educational background, topic knowledge, discourse experience, and familiarity with a speaker may affect which continuations are probable to that person. A specialist reading a familiar scientific passage can find its terminology predictable while a newcomer experiences each technical noun as unexpected. Statistical models trained on different domains can provide corresponding differences in estimated surprisal. These observations make population-average cloze norms useful approximations rather than universal individual expectations. Kuperberg and Jaeger (2016).
Research published in 2026 tested domain-adapted language models against eye-movement data from German students reading biology and physics texts. Models better aligned with the readers' experience showed improved prediction of some later reading-time measures compared with less reader-aligned estimates. The findings support the methodological value of representing reader-specific experience; they do not establish that a machine model learns or understands language in the same manner as the readers. Škrjanec and Demberg (2026).
Age, literacy, multilingual experience, and task demands
A claim that “everyone predicts every next word” cannot be inferred from average effects in a sample of university students. The speed and specificity of anticipation may depend on linguistic skill, literacy, attention, processing resources, age, and task demands. Some experiments require an unusually constrained sentence or a visible set of objects; others measure natural reading with no explicit anticipation instruction. The relevant question is how the task and experience change the mechanism and the usefulness of preparing information in advance. Reviews place these individual and contextual constraints at the center of a balanced theory. Huettig (2015); Pickering and Gambi (2018).
Multilingual comprehension adds language-specific frequencies, cross-language activation, code-switching context, and differences in proficiency. The existence of contextual prediction in one language or task does not establish a uniform pattern across all languages and writing systems. Cross-linguistic experiments that vary word order and grammatical marking are especially useful because they reveal whether a proposed mechanism uses abstract event knowledge or language-specific cues. Japanese and English anticipatory eye-movement results demonstrate how differing orders of input can still support forward-looking interpretation. Kamide, Altmann, and Haywood (2003).
Typical variation and clinical boundaries
Differences in anticipatory behavior occur within the ordinary range of language experience and task performance. A person may show weaker expectation effects because a context is unfamiliar, the task taxes attention, the signal is difficult to perceive, or the individual uses a different comprehension strategy. None of these observations alone establishes dyslexia, developmental language disorder, aphasia, stuttering, apraxia, or any other diagnosis. Research involving atypical language or sensory populations can test mechanistic hypotheses, but an experimental group contrast does not translate directly into a clinical test for an individual. Gastaldon and colleagues (2024).
Clinical assessment requires an appropriate professional evaluation and independent evidence of functional difficulty, developmental or acquired history, relevant language and dialect, and other possible contributing conditions. The current article concerns the cognitive science of expectation during comprehension. Diagnostic criteria, differential assessment, and treatment effectiveness are separate specialized questions, as are the learning-disorder mechanisms addressed within established dyslexia research.
Prediction and Large Language Models
What model probabilities contribute to psycholinguistics
A computational language model estimates conditional probabilities for upcoming text. Researchers can assign each word a surprisal value and ask whether the resulting series predicts fixation duration, reading time, neural responses, or comprehension difficulty. This procedure gives abstract claims about expectancy a quantitative form. Corpus-based models have been used to account for naturalistic reading and ERP findings, and their limitations can be tested by comparing predictions with human responses. Frank, Otten, Galli, and Vigliocco (2015); Staub (2025).
Modern language models are valuable for evaluating alternative probability estimates across large materials. Their predictions may fit certain human measurements well without matching the human learning environment. Model size, training distribution, tokenization, contextual window, and exposure to the test material can change the resulting probabilities. A machine's success at assigning lower surprisal to the same words that humans process faster is evidence of a useful statistical correspondence; it is not sufficient to identify a human neural implementation.
Why machine prediction and human comprehension must be distinguished
Human language comprehension involves perception, bodily and social context, acquired world knowledge, discourse goals, memories, and interactions with other people. A model that generates plausible text demonstrates a capacity to compute and use regularities in language. Comparison with human comprehension requires explicit behavioral and neural benchmarks and careful treatment of the different information sources available to each system. The 2026 study on reader-experience alignment illustrates one route to better model evaluation: match the relevant exposure and then test the predictions against human behavior. Škrjanec and Demberg (2026).
Questions concerning subjective experience, artificial selfhood, and the philosophical status of artificial reason belong to a distinct level of inquiry. Neither low word surprisal nor fluent generative output establishes subjective experience or human-equivalent understanding. Psycholinguistic model comparisons can illuminate specific computations while leaving ontological questions open to their appropriate philosophical and scientific methods.
How to Judge the Strength of a Prediction Study
Identify the phenomenon before selecting a mechanism
A finding that contextual expectancy improves recognition is evidence for contextual facilitation. A gaze shift before a noun provides evidence for anticipation toward a referent. A pre-onset grammatical-cue response can provide more direct evidence for preparation of certain features. A reduced N400 after a word appears shows sensitivity to preceding context and incoming meaning. A surprisal–reading-time correlation supports quantitative prediction of behavior. These findings have different inferential strengths and should be described in their own terms. Kuperberg and Jaeger (2016); Wong, Reichle, and Veldre (2025).
Control the right alternative explanations
Important alternatives include semantic relatedness, lexical frequency, word length, orthographic and phonological familiarity, sentence plausibility, syntactic complexity, explicit task strategies, and properties of the experimental display. In EEG studies, preprocessing, baseline choices, time windows, and component definitions can alter estimates. In corpus-based modeling, differences among probability estimators and the training material complicate causal interpretations. A useful design measures or manipulates the variables that distinguish the theories, rather than simply obtaining another significant effect of high versus low predictability.
Interpret replication and evidence quality precisely
A preregistered multi-laboratory result can narrow the plausible size of an effect and reveal where an influential mechanism claim was overstated. The 2018 replication of the DeLong article effect is a case in point. The correct scientific response is to update the specific claim about word-form pre-activation while retaining supported evidence for other kinds of contextual sensitivity. Replication also depends on linguistic materials, task timing, participant samples, measurement reliability, and the analysis chosen. Good reviews distinguish robust behavioral effects from disputed mechanistic interpretations. Nieuwland and colleagues (2018); Staub (2025).
Combine converging measures and meaningful comparisons
The most informative research integrates eye movements, reaction time, comprehension accuracy, neural responses, and explicit computational predictions, with each method contributing different constraints. A theory that explains N400 modulation but incorrectly predicts a large pre-onset article effect needs revision. Another theory that captures reading times but fails on discourse-dependent expectations may require richer contextual representations. Explanatory progress comes from predictions that are independently falsifiable and from models calibrated against genuine out-of-sample human observations.
Practical Implications of Predictive Language Processing
Clear communication and contextual support
Communication is easier when speakers and writers establish what a discussion is about, introduce unfamiliar terms, and connect new information to relevant prior context. Predictable language can reduce certain processing demands, especially in tasks with limited time or noisy input. At the same time, important communication often depends on novelty: a clear explanation must make room for information the reader could not already infer. The useful principle is to support coherent interpretation rather than maximize the frequency of obvious words. Smith and Levy (2013); Wong, Reichle, and Veldre (2025).
Reading and learning
Previews, topic cues, coherent headings, and appropriately introduced vocabulary can supply a framework within which new terms become interpretable. These techniques are consistent with the general importance of context in language processing, but a study of word-level expectation does not, by itself, prove the effectiveness of a particular educational intervention. Teaching decisions require independent learning-outcome evidence. The distinction keeps an established cognitive mechanism from being prematurely converted into a guaranteed instructional method.
Designing experiments and language technologies
Researchers can use predictability norms and computational surprisal values to select materials, compare theoretical models, and anticipate which words might be difficult for readers. Technology developers can use these quantities when assessing readability and language interfaces. The results should be calibrated for intended users, tasks, and domains, particularly when groups have different language histories. Statistical predictions from a model are research tools whose validity must be established against appropriate human outcomes.
Frequently Asked Questions
Does the brain predict the next word while someone reads or listens?
Human comprehension is sensitive to the contextual probability of upcoming language, and some experiments demonstrate anticipatory attention or grammatical preparation before the relevant word appears. Evidence that a particular person routinely pre-activates every exact upcoming word is much weaker. Prediction varies by representational level, context, and task. Huettig (2015); Nieuwland and colleagues (2018).
What is the difference between cloze probability and surprisal?
Cloze probability is estimated from people's overt completions of a language fragment. Surprisal is the negative logarithm of the probability assigned to an incoming item by a specified distribution. Cloze measures can help estimate probabilities, while computational models can assign surprisal across large text corpora. Neither directly observes the neural mechanism behind a reading-time effect.
Does a smaller N400 prove that a word was predicted?
A smaller N400 indicates context-sensitive processing of meaning and often accompanies higher predictability. The effect can be explained by multiple mechanisms involving semantic access, preparation, and integration. Specific pre-activation claims require experimental evidence able to discriminate those alternatives. Kutas and Federmeier (2011).
Why is the 2018 nine-laboratory replication important?
It reproduced the noun N400 predictability effect but not the celebrated preceding-article pattern interpreted as evidence for anticipatory phonological word-form activation. The result supports a more precise account of which levels of prediction have strong evidence. It does not show an absence of all anticipatory language processing. Nieuwland and colleagues (2018).
Does an unexpected word always slow comprehension?
Less predictable words often take longer to process on average, but the outcome depends on plausibility, similarity to other expected continuations, task requirements, and subsequent sentence structure. Studies have found robust predictability effects without a general extra penalty for disconfirming a single highly expected word. Frisson, Harvey, and Staub (2017).
Is prediction the same as predictive coding in the brain?
Predictive language processing is an empirical subject involving contextual expectations and comprehension. Predictive coding refers to a broader family of formal models of information exchange, error signals, and hierarchical inference. A linguistic expectancy effect can be compatible with several architectures and is not a unique validation of one neural coding theory.
Can a language model predict human reading difficulty?
Language-model surprisal can explain measurable variation in human reading and neural responses. How well it performs depends on the model, training data, stimulus domain, and measure of human behavior. Matching a statistical curve is useful evidence for model evaluation and does not establish identity of human and machine cognition. Škrjanec and Demberg (2026); Staub (2025).
Scientific Synthesis
Predictive language processing describes how contextual knowledge shapes expectations during ongoing comprehension. Robust findings show that predictability is associated with reading behavior, semantic brain responses, and some anticipatory attentional shifts. Formal surprisal models connect those patterns to quantitative probabilities and illuminate how a graded expectation distribution influences information processing. Direct pre-onset evidence supports anticipation in particular visual and grammatical contexts, while large-scale replication has constrained strong claims about the routine prediction of exact phonological word forms.
The central scientific challenge is now mechanistic precision: identify which representation is anticipated, which information supports it, when the operation occurs, what processing benefit or cost follows, and whether an alternative account predicts the same result. Integrating experiment, replication, naturalistic measurement, and carefully matched computational models turns prediction from an appealing metaphor into a research program with testable claims. Language comprehension emerges as a flexible interaction among learned regularities, situated meaning, attention, and new input, supporting understanding even when an utterance takes a surprising turn.
Related Articles
Semantic Priming: How Related Words Influence Processing — How related concepts affect lexical recognition and why priming must be distinguished from anticipatory word prediction.
Lexical Ambiguity: How Context Resolves Multiple Word Meanings — How contextual expectations guide the interpretation of words with competing meanings.
Discourse Comprehension: How We Understand Connected Language — How a developing discourse model supports reference, coherence, contextual inference, and expectations.
Cluster Language, Thought and Psycholinguistics — The parent scientific chapter situating prediction within language comprehension, production, learning, and thought.
Predictive Processing in Perception: Theory, Evidence, and Debate — A separate treatment of the broader predictive-processing framework and its evidential limits.
Semantic Memory: Facts, Meaning, and Knowledge — The structure and retrieval of conceptual knowledge that can contribute to contextual expectancy.
Selective Attention: How the Mind Filters Information — The allocation of attention that affects listening, visual-world behavior, and recognition.
Tip-of-the-Tongue Phenomenon: Why Words Feel Just Out of Reach — A distinct question about lexical retrieval rather than anticipation in comprehension.
References
Altmann, G. T. M., and Kamide, Y. (1999). Incremental interpretation at verbs: Restricting the domain of subsequent reference. Cognition, 73(3), 247–264. https://doi.org/10.1016/S0010-0277(99)00059-1
Brothers, T., Swaab, T. Y., and Traxler, M. J. (2015). Effects of prediction and contextual support on lexical processing: Prediction takes precedence. Cognition, 136, 135–149. https://doi.org/10.1016/j.cognition.2014.10.017
Brouwer, H., Fitz, H., and Hoeks, J. (2012). Getting real about semantic illusions: Rethinking the functional role of the P600 in language comprehension. Brain Research, 1446, 127–143. https://doi.org/10.1016/j.brainres.2012.01.055
DeLong, K. A., Urbach, T. P., and Kutas, M. (2005). Probabilistic word pre-activation during language comprehension inferred from electrical brain activity. Nature Neuroscience, 8, 1117–1121. https://doi.org/10.1038/nn1504
Federmeier, K. D., and Kutas, M. (1999). A rose by any other name: Long-term memory structure and sentence processing. Journal of Memory and Language, 41(4), 469–495. https://doi.org/10.1006/jmla.1999.2660
Frank, S. L., Otten, L. J., Galli, G., and Vigliocco, G. (2015). The ERP response to the amount of information conveyed by words in sentences. Brain and Language, 140, 1–11. https://doi.org/10.1016/j.bandl.2014.10.006
Frisson, S., Harvey, D. R., and Staub, A. (2017). No prediction error cost in reading: Evidence from eye movements. Journal of Memory and Language. https://doi.org/10.1016/j.jml.2017.04.007
Gastaldon, S., Bonfiglio, N., Vespignani, F., and Peressotti, F. (2024). Predictive language processing: Integrating comprehension and production, and what atypical populations can tell us. Frontiers in Psychology, 15, 1369177. https://doi.org/10.3389/fpsyg.2024.1369177
Huettig, F. (2015). Four central questions about prediction in language processing. Brain Research, 1626, 118–135. https://doi.org/10.1016/j.brainres.2015.02.014
Kamide, Y., Altmann, G. T. M., and Haywood, S. L. (2003). The time-course of prediction in incremental sentence processing: Evidence from anticipatory eye movements. Journal of Memory and Language, 49(1), 133–156. https://doi.org/10.1016/S0749-596X(03)00023-8
Kuperberg, G. R., and Jaeger, T. F. (2016). What do we mean by prediction in language comprehension? Language, Cognition and Neuroscience, 31(1), 32–59. https://doi.org/10.1080/23273798.2015.1102299
Kutas, M., and Federmeier, K. D. (2011). Thirty years and counting: Finding meaning in the N400 component of the event-related brain potential (ERP). Annual Review of Psychology, 62, 621–647. https://doi.org/10.1146/annurev.psych.093008.131123
Kutas, M., and Hillyard, S. A. (1980). Reading senseless sentences: Brain potentials reflect semantic incongruity. Science, 207(4427), 203–205. https://doi.org/10.1126/science.7350657
Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106(3), 1126–1177. https://doi.org/10.1016/j.cognition.2007.05.006
Nieuwland, M. S., et al. (2018). Large-scale replication study reveals a limit on probabilistic prediction in language comprehension. eLife, 7, e33468. https://doi.org/10.7554/eLife.33468
Pickering, M. J., and Gambi, C. (2018). Predicting while comprehending language: A theory and review. Psychological Bulletin, 144(10), 1002–1044. https://doi.org/10.1037/bul0000158
Škrjanec, I., and Demberg, V. (2026). Language models that match reader experience are better predictors of reading times. Journal of Memory and Language, 146, 104677. https://doi.org/10.1016/j.jml.2025.104677
Smith, N. J., and Levy, R. (2013). The effect of word predictability on reading time is logarithmic. Cognition, 128(3), 302–319. https://doi.org/10.1016/j.cognition.2013.02.013
Staub, A. (2025). Predictability in language comprehension: Prospects and problems for surprisal. Annual Review of Linguistics, 11, 17–34. https://doi.org/10.1146/annurev-linguistics-011724-121517
Van Berkum, J. J. A., Brown, C. M., Zwitserlood, P., Kooijman, V., and Hagoort, P. (2005). Anticipating upcoming words in discourse: Evidence from ERPs and reading times. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31(3), 443–467. https://doi.org/10.1037/0278-7393.31.3.443
Van Petten, C., and Luka, B. J. (2012). Prediction during language comprehension: Benefits, costs, and ERP components. International Journal of Psychophysiology, 83(2), 176–190. https://doi.org/10.1016/j.ijpsycho.2011.09.015
Willems, R. M., Frank, S. L., Nijhof, A. D., Hagoort, P., and van den Bosch, A. (2016). Prediction during natural language comprehension. Cerebral Cortex, 26(6), 2506–2516. https://doi.org/10.1093/cercor/bhv075
Wong, R., Reichle, E. D., and Veldre, A. (2025). Prediction in reading: A review of predictability effects, their theoretical implications, and beyond. Psychonomic Bulletin and Review, 32, 973–1006. https://doi.org/10.3758/s13423-024-02588-z