Human language processing and large language models overlap in several measurable computational functions: both are sensitive to linguistic context, both support predictions about upcoming expressions, both can exploit statistical regularities, and representations learned by some models predict a portion of human neural responses to language. Their similarities are scientifically valuable because they create testable links between psycholinguistic theory, computational modeling, and brain data. Their limits are equally specific. Human language is acquired and used within embodied, developmental, interactive, social, and goal-directed activity; an LLM's behavior depends on its training objectives, architecture, data, inputs, and the external tools with which it may be combined. A good behavioral match therefore establishes a correspondence for the tested function, while a claim about shared mechanisms needs separate evidence. Mahowald et al. (2024); Tuckute, Kanwisher, and Fedorenko (2024).
This scientific comparison addresses a specific question: which features of human comprehension and production can present-day language models reproduce or help explain, and what follows from those correspondences? The analysis proceeds from psychological processes and operational measurements to computational mechanisms, neurobiological findings, generalization, learning, grounding, dialogue, and the logic of interpreting experimental results. It distinguishes established observations, provisional findings, theoretical interpretations, and philosophical propositions throughout.
What Counts as Human Language Processing?
Comprehension and production are temporally unfolding processes
Human language comprehension coordinates sensory information, phonological and lexical recognition, grammatical structure, semantic composition, discourse context, and pragmatic inference. In speech, these operations unfold while the signal is still arriving; in reading, eye movements and visual recognition influence the evidence available at each moment. A speaker also plans an intended message, selects expressions, coordinates articulation or signing, and monitors how the utterance is received. Language can consequently be examined at the millisecond scale of neural and behavioral responses, across the seconds of conversation, and over the years of development. Tuckute, Kanwisher, and Fedorenko (2024).
A comprehension measure answers a narrowly specified question. Reading time may reveal difficulty at a word, a brain recording may show a response sensitive to predictability, and a comprehension choice may show what a participant inferred. No individual response measure directly contains the whole of comprehension. Researchers establish stronger interpretations by converging evidence from independent tasks, perturbations, individual differences, and out-of-sample prediction.
Language participates in broader cognition
Humans regularly use language to make requests, coordinate intentions, teach, imagine situations, remember events, and revise commitments. These activities enlist memory, attention, perceptual systems, social reasoning, and executive functions. Neuroscience finds a relatively selective human language network whose activity can be distinguished from the networks recruited for other demanding cognitive tasks. The distinction matters: an effective language-processing model need not duplicate every system involved in human reasoning, and performance on a reasoning question does not identify the neural mechanism of language comprehension. Mahowald et al. (2024).
What Is a Large Language Model?
Statistical training and contextual representation
A large language model is a trained computational system that maps representations of input sequences into predictions, transformations, or generated outputs. Many influential generative models use transformer architectures and an autoregressive objective that estimates the probability of the next token given prior context. A token can be a word, part of a word, punctuation, or another encoding unit. During training, parameters change to reduce prediction error across large collections of examples. During ordinary inference, the trained model uses these parameters together with the current context to compute possible continuations. Schrimpf et al. (2021); Goldstein et al. (2022).
A token-prediction objective can support capabilities broader than literal continuation because the training signal rewards sensitivity to syntax, lexical relations, topic, style, reference, and patterns of reasoning expressed in text. Instruction tuning, feedback-based optimization, retrieval, external tools, and multimodal components can further change behavior. An experiment comparing human cognition with an LLM must therefore specify its actual configuration. A text-only pretrained model, an instruction-following conversational model, and a tool-using multimodal system are different experimental objects.
Parameters, context, and generated language
In a transformer, contextual representations are distributed numerical states shaped by patterns learned during training and by tokens in the current sequence. They are useful for predicting words in context and may encode information that investigators can probe. A model's temporary context and its trained parameters serve different roles: a word placed in a prompt may influence the ongoing answer without any permanent change to the model's parameters. Dialogue systems may also append summaries or retrieved documents, producing additional apparent continuity through external memory components.
Human memory and learning involve multiple interacting processes: momentary retention, episodic experience, consolidation, semantic learning, and ongoing adaptation. Treating a model's attention context as identical to human working memory erases distinctions in duration, mechanism, behavioral function, and neural implementation. The appropriate comparison identifies what information is available, how long it persists, how it changes responses, and what kind of learning is demonstrated. Cuskley, Woods, and Flaherty (2024).
The Levels of Comparison
Output similarity
The weakest but often most accessible comparison asks whether humans and models produce the same answer or similar text under the same instructions. This is a legitimate behavioral result. It can show that a computational system achieves a function, but identical answers can arise from different strategies. A model that selects the correct interpretation by recognizing a repeated surface pattern and a person who reasons from a situation model may reach the same outcome. Controlled contrasts are required to discriminate the two explanations. McCoy, Pavlick, and Linzen (2019).
Predictive similarity
A stronger comparison asks whether probabilities assigned to linguistic alternatives predict human responses over many items. Researchers can compare model surprisal with reading times, judgments, eye movements, or neural measurements. Reliable out-of-sample correlations support a predictive relation. They do not establish that the model computes its predictions by the same internal operations as a person. Competing predictors and relevant confounds must be compared directly. Wilcox et al. (2023).
Representational and mechanistic similarity
Representational comparisons ask whether patterns across model activations resemble patterns across brain areas, neural recordings, or psychological dimensions. Mechanistic comparisons go further by assessing causal structure, processing dynamics, and what happens when parts of a system are perturbed. Shared explanatory variables and successful intervention predictions support a mechanistic theory more strongly than static correlations. Human neural activity and numerical model states occupy different physical substrates, so claims about mechanisms require operational definitions that span them. Caucheteux and King (2022).
Language Prediction and Surprisal
Why predictable words are processed differently
Human listeners and readers use linguistic context as information about what may come next. The technical measure surprisal assigns greater values to events with lower predicted probability, typically expressed as negative log probability. A word that is difficult to anticipate will often produce longer reading times or greater processing responses, although effects depend on word frequency, structure, task, and measurement. Prediction-based models can yield quantitatively testable hypotheses without requiring the strong assumption that every person consciously guesses the next word. Wilcox et al. (2023).
The distinction between prediction and integration remains important. A processing cost at an unexpected word may reflect anticipatory probability, the work required to integrate incoming information, or both. A model whose probabilities predict human reading time contributes to explanation, but a theory about predictive neural computation must be tested independently. The relationship between surprisal and response time is a model of measured effects, not an exhaustive theory of understanding.
Cross-linguistic evidence
Wilcox and colleagues tested surprisal theory with reading-time evidence across eleven languages from five language families. Their results support important predictions about the relationship between linguistic probability and processing difficulty beyond English alone. This wider sampling strengthens the relevance of probabilistic language models while leaving open questions about modality, speaking, dialogue, developmental stages, and the exact cognitive operations that give rise to each effect. Wilcox et al. (2023).
Prediction in natural speech
Goldstein and colleagues recorded intracranial neural activity from nine participants listening to a naturalistic podcast and compared neural responses with predictions derived from an autoregressive language model. They reported correspondences involving anticipatory word-related signals, responses associated with unexpected incoming words, and contextual representations. The study provides unusually direct evidence relevant to prediction during continuous speech; its restricted participant group, recording conditions, and particular model establish clear limits on generalization. Goldstein et al. (2022).
What Brain Alignment Studies Actually Show
Predicting brain responses from model features
In encoding studies, investigators feed linguistic materials into a computational model, extract numerical features, and fit a mapping between those features and human neural responses. The mapping is then assessed on held-out stimuli. High predictive accuracy indicates that the features retain information systematically related to brain activity under the experiment's conditions. It does not imply that a neuron stores a corresponding model dimension or that the brain executes the same transformer algorithm.
Schrimpf and colleagues compared a range of models against language-related neural responses and found that models performing well on next-word prediction tended to predict some brain data well. These findings support a productive connection between predictive modeling and neuroscience. Model architecture, training objective, behavioral task, model scale, and training data are partially correlated in practical comparisons; attributing a result to one component requires carefully controlled alternatives. Schrimpf et al. (2021).
Convergence across fMRI and MEG
Caucheteux and King compared multiple deep language models with measurements from functional magnetic resonance imaging and magnetoencephalography. Their analyses involved neural responses to 400 isolated sentences in 102 participants, enabling questions about where and when model-derived features predict human signals. They reported partial convergence associated with word-prediction ability and the emergence of linguistic representations. “Partial” is an empirical description: a successful mapping exists for particular aspects of responses, alongside variance outside the model's account. Caucheteux and King (2022).
Evidence through 2025
The 2024 review by Tuckute, Kanwisher, and Fedorenko brought together research on computational models as probes of the neural organization of language and emphasized the need to separate models' linguistic performance, architecture, and training histories. Gao and colleagues reported increasing alignment between transformer language models and human language-related brain measurements in work published in 2025. Such findings expand the empirical reach of brain-model comparisons, but their meaning depends on dataset design, context length, measurement reliability, statistical controls, and the particular meaning of alignment. Tuckute, Kanwisher, and Fedorenko (2024); Gao et al. (2025).
Why a high correlation has several explanations
A model may predict brain responses because it captures the identity of words, word frequency, semantic associations, discourse information, syntactic relations, or other variables that also influence neural signals. Several of these properties covary naturally. Linear mapping can make different representational systems appear compatible even when their original coordinate systems differ. Repeated materials or overlapping distributions can further elevate predictive results. Neuroscientific interpretation therefore calls for explicit baselines, variance partitioning, controlled stimulus contrasts, held-out participants, and tests with materials constructed to distinguish theories.
Syntax and Hierarchical Structure
The problem of grammatical generalization
Humans understand syntactic dependencies even when the relevant words are separated by intervening material. Agreement, embedding, long-distance dependencies, and role assignment therefore test whether a system has learned more than simple neighboring-word associations. Contemporary language models can perform impressively on numerous grammatical judgments and can encode information related to structure. Mahowald and colleagues emphasize that language models offer serious evidence of competence in many formal linguistic tasks, even as performance changes across particular constructions, languages, and evaluation designs. Mahowald et al. (2024).
Surface heuristics versus structural interpretation
Consider a sentence in which two nouns occur but grammatical structure determines which noun is the agent. A shortcut based on word order or lexical overlap may work on common examples while giving a wrong interpretation when the structure changes. McCoy and colleagues developed the HANS evaluation to expose these syntactic shortcuts in natural language inference models. Their findings caution against reading ordinary benchmark accuracy as proof of robust structural inference. This is an experimental lesson about the relationship between a test and its target construct, rather than a timeless verdict about every subsequent model. McCoy, Pavlick, and Linzen (2019).
Compositionality and transfer
Compositional generalization concerns the ability to use familiar components in unfamiliar combinations according to their relationships. Lake and Baroni used artificial instruction sequences to demonstrate severe systematic generalization failures in earlier sequence-to-sequence neural networks under certain train-test splits. Their study remains a foundational experimental illustration, while changes in model architectures and training must be assessed on new data rather than assumed to reproduce those historical outcomes. The crucial issue is whether a system handles controlled structural recombinations outside its training patterns. Lake and Baroni (2018).
Lexical and Semantic Representation
Distributed representations and context
Humans use lexical knowledge to identify words, integrate meanings, and select expressions appropriate to a situation. In LLMs, distributed representations also vary with context; a word in a finance story and the same word in a river description can evoke different internal states. Contextual vectors can capture useful semantic relationships without assuming that a particular unit corresponds to a single stable human concept. Their explanatory force is measured by what they predict about behavior and neural data. Caucheteux and King (2022).
Meaning as a scientific construct
Meaning can be investigated as a relation among conventional expressions, structured propositions, reference, inference, action, and social use. These dimensions support different operational tests. An LLM's correct disambiguation of a written word demonstrates a language-use capacity under the test conditions. A human participant's success may additionally depend on remembered events, perceptual experience, communicative intentions, and knowledge of practical consequences. A comparison becomes informative when it identifies which dimension of meaning is being assessed. Bender and Koller (2020).
The symbol-grounding problem
The symbol-grounding problem asks how internal symbols or representations connect to the entities, events, and relations to which they refer. In humans, learning often occurs in settings where hearing words, seeing objects, acting, and receiving social feedback are intertwined. Text-trained models can infer numerous regularities about the world because language carries information about experience. Whether a particular model's inferred representations reproduce the causal, perceptual, and pragmatic organization of human meaning must be demonstrated in grounded tasks. Multimodal inputs modify the research problem by giving models additional sensory channels; the relevant question becomes how those channels are integrated and used. Bender and Koller (2020); Cuskley, Woods, and Flaherty (2024).
Pragmatics and Conversation
Speaker intentions and common ground
Human communication is shaped by what participants want to accomplish together. An utterance such as “It's cold in here” may be a factual statement, a request to close a window, or a subtle criticism. Its interpretation depends on prior dialogue, social expectations, material circumstances, and assumptions about the speaker's goals. A model may generate a socially appropriate continuation by drawing on learned pragmatic patterns. To test similarity with people, the experiment must determine how the system handles changes in shared knowledge, hidden information, conflicting interests, and consequences for cooperative action.
Dialogue is more than producing plausible turns
Live conversation unfolds through timing, interruption, repair, mutual adjustment, gesture, and the negotiation of common ground. Text-chat LLM evaluations omit many of these components. Voice-enabled or embodied systems add further behavior, but the addition must be analyzed in relation to latency, sensory integration, turn allocation, feedback, and learning. A smooth conversational style can support perceived understanding in users, yet smoothness by itself provides limited information about the causal organization of the system. Cuskley, Woods, and Flaherty (2024).
Theory-of-mind tasks need construct validity
Some evaluations ask models to infer what a character falsely believes, what an indirect request means, or why an utterance is ironic. Correct answers reveal performance on those problem formats. Theory of mind in humans, however, is studied across development and social interaction using distinct tasks, mental-state information, and real-world consequences. Generalization to new informational constraints, consistency across paraphrases, and performance when surface cues conflict are particularly important. An answer that resembles a human answer is a point of measurement, not by itself evidence about subjective perspective or experienced mental states. Mahowald et al. (2024).
Human Development and Model Training
Children's language environments
Young humans acquire language while perceiving people and objects, coordinating attention, participating in routines, and gradually developing movement, memory, and conceptual knowledge. Their linguistic input includes speech directed toward them, overheard talk, prosody, interactional repair, and multimodal cues. A model trained on extensive collections of edited text has a different data distribution, learning schedule, and opportunity for feedback. These contrasts are consequential for claims that model success establishes a particular account of human acquisition. Cuskley, Woods, and Flaherty (2024).
Data efficiency and inductive biases
A scientific comparison of learning efficiency should consider exposure quantity, information quality, task diversity, architecture, and what prior constraints the learner possesses. Comparing a child's estimated spoken-word exposure with an LLM's text-token count is informative only if the underlying modalities and information sources are acknowledged. Models can learn useful linguistic structure from data, but their success does not settle how much innate organization, social interaction, or sensorimotor experience contributes to human acquisition. Equally, a model's failure on one learning task cannot establish that a specific human grammatical rule is biologically preprogrammed. Cuskley, Woods, and Flaherty (2024).
Learning during use
Human conversational partners update beliefs, remember events, acquire words, and change habits through experience. Many deployed LLMs respond with fixed weights during ordinary use, although their inputs may include retrieved memories, updated context, and tool results. Fine-tuning and additional training create further forms of adaptation. Researchers should report which form of change occurred, how durable it was, and whether the system generalized to new settings. This distinction prevents temporary prompt sensitivity from being misdescribed as the same process as lasting human learning.
Reading, Speech, Sign, and Multilingual Language
A modality mismatch
Human language processing encompasses spoken language, signed languages, reading, writing, and multimodal interaction. A text-based language model is directly exposed to one coded modality, often derived from human written communication. It can support spoken or signed interfaces through additional components, but success on written prompts cannot be projected automatically to speech-in-noise perception, visual sign parsing, articulation, gesture coordination, or real-time turn-taking. The comparison must identify the relevant sensory and motor channels. Tuckute, Kanwisher, and Fedorenko (2024).
Language diversity and coverage
Training datasets are often highly uneven across languages, varieties, scripts, and social contexts. A model's fluent English output says relatively little about its competence with a low-resource language or a less frequently represented register. Human language users also vary in proficiency and exposure, but comparisons should match the languages and experiences actually under study. Cross-linguistic surprisal findings illustrate how broad sampling can strengthen a psychological theory; the same standard is needed when claims about models extend to “human language” generally. Wilcox et al. (2023); Cuskley, Woods, and Flaherty (2024).
Formal Linguistic Competence and Functional Competence
Two dimensions of success
Mahowald and colleagues distinguish formal linguistic competence, encompassing the patterns and structure of language, from functional linguistic competence, concerning the use of language for goals, reasoning, and world-facing activity. This distinction clarifies why a system can demonstrate powerful grammatical and lexical abilities while exhibiting different levels of reliability in tasks that depend on external knowledge, planning, physical constraints, or social inference. It also reflects neuroscientific evidence that human language-responsive regions are functionally distinguishable from networks supporting other forms of cognition. Mahowald et al. (2024).
The distinction is analytical, not a scorecard
Competence exists across different tasks, distributions, and evaluation procedures. A system may handle a reasoning problem successfully but fail a superficially similar version; another system may gain reliability from tools or verification. Calling an entire architecture capable or incapable of functional understanding without specifying a measurable domain hides this variation. A useful account records which task was performed, what external resources were allowed, which failures occurred, and how the results compare with carefully matched human performance.
Where Models Mislead the Psychologist
Benchmark contamination and memorization
Large datasets can include tasks or close paraphrases later used for evaluation. If test items or their solutions were represented in training material, scores may overestimate transfer. Researchers should use controlled held-out materials, document corpus overlap where possible, and test new combinations that resist memorized templates. The evaluation remains valuable when contamination cannot be ruled out, but claims about novel generalization become weaker.
Prompt sensitivity and answer formats
Instructions, item order, wording, sampling temperature, and formatting can change a model's answers. People are also sensitive to framing and context, though the mechanisms and magnitude of the effects need not match. For fair tests, researchers should prespecify prompts, examine reasonable paraphrases, report sampling procedures, and evaluate performance distributions rather than a selected successful transcript. An open-ended conversation is especially vulnerable to post hoc selection.
Human baselines and measurement equivalence
A human comparison group should match the test's language, instructions, knowledge, time limits, and relevant opportunities for explanation. Text-only tasks may disadvantage some participants while suiting language models; real-world tasks may impose challenges not captured by the text prompt. Accuracy alone can conceal large differences in response latency, error distribution, confidence calibration, learning, and robustness. The strongest comparisons report these dimensions and their uncertainty.
Statistical significance versus theoretical importance
A small improvement in predicting neural data can be statistically reliable without identifying a shared biological algorithm. Conversely, a model can illuminate a specific cognitive hypothesis despite modest overall benchmark scores. Scientists must specify the intended level of explanation. Mechanistic conclusions benefit from interventions that selectively change hypothesized information, followed by evidence that people and models change in the predicted ways. Schrimpf et al. (2021); Caucheteux and King (2022).
Can Language Models Explain Human Language?
Models as formal hypotheses
Computational systems become scientifically useful when they transform theoretical ideas into quantifiable predictions. A language model can generate surprisal estimates for stimulus materials, embeddings for neuroimaging analysis, competing predictions about ambiguity, and synthetic examples that challenge an account of grammatical generalization. This contribution can be substantial even when a model's design differs from biological processing. The key requirement is that a model supplies predictions that can be tested independently. Tuckute, Kanwisher, and Fedorenko (2024).
Prediction does not settle causal mechanism
A language model can fit human data by discovering regularities correlated with cognitive processes. Explanatory progress requires determining whether those regularities correspond to psychological operations or merely provide statistically convenient proxies. Comparative ablations, causal interventions on stimuli, constraints on model resources, and replication across independent neural and behavioral datasets can narrow the possibilities. A model with high predictive accuracy is a promising candidate in a larger process of theory discrimination.
A productive disagreement in the literature
The neuroscience-oriented literature documents meaningful correspondences between language models and human neural processing; methodological critiques emphasize that matching written-language behavior does not establish how humans acquire or implement language. These claims can be jointly accommodated by making the level of explanation explicit. A model can explain variance in a neural signal and remain inadequate as a developmental theory. A mechanism can share a mathematical operation with the brain while differing in sensory inputs, learning history, and social function. Goldstein et al. (2022); Cuskley, Woods, and Flaherty (2024).
A Practical Framework for Designing Comparison Studies
Define the construct first
A study should begin with an independently defined psychological construct: lexical competition, syntactic integration, pragmatic inference, discourse tracking, or another specific process. The construct determines suitable tasks and measurements. Broad labels such as “understanding” and “intelligence” need operational definitions to prevent conclusions from outrunning evidence.
Specify the model and its information access
Record model family, version or date of access, training or fine-tuning disclosures, modality, context length, prompts, sampling settings, retrieval, tools, and other scaffolding. An apparently unchanged conversational interface may conceal a changed model. For human comparisons, record language experience, participant recruitment, and task conditions.
Build discriminating stimuli
Use minimal contrasts that separate surface statistics from grammatical relations, literal content from pragmatic inference, memorized knowledge from generalization, and familiar situations from novel combinations. Include controlled adversarial cases alongside naturalistic materials. Whenever possible, preregister predictions and use independent datasets for evaluation.
Compare multiple linked measures
Collect behavioral outcomes such as accuracy, response time, error types, confidence, or comprehension choices. When appropriate, link them with eye tracking, EEG, MEG, fMRI, or intracranial recordings. Use model-derived quantities with held-out validation and credible baselines. Distinguish a reliable association from a demonstrated causal pathway. Wilcox et al. (2023); Caucheteux and King (2022).
Seek failure-pattern correspondence
Matching successes can arise from many strategies, whereas selective errors under tightly controlled conditions may discriminate mechanisms. Ask whether a model and humans struggle with the same ambiguity, react similarly to unexpected words, transfer the same structural relation, or make systematically different mistakes. A dissociation is informative when task and resource differences are controlled.
Established Findings, Limited Evidence, and Open Questions
Established within defined experimental conditions
Probabilistic expectations predict important aspects of human reading behavior across languages; some language-model features predict measurable aspects of neural responses during language processing; contemporary LLMs demonstrate substantial formal linguistic competence. These conclusions concern particular measures, populations, model classes, and task conditions. Wilcox et al. (2023); Schrimpf et al. (2021); Mahowald et al. (2024).
Promising but conditional evidence
Representational alignment findings indicate partial convergence between models and human language-related brain activity. Evidence is promising because it survives some held-out tests and spans recording techniques, while inference remains sensitive to selection of stimuli, model representations, statistical mapping, and participant populations. A study demonstrating improved fit should not be generalized to every neural process or every aspect of meaning. Gao et al. (2025); Caucheteux and King (2022).
Contested theoretical interpretations
Questions remain about which facets of human comprehension require anticipatory prediction rather than integration, whether model-derived semantic spaces reproduce human conceptual organization, and how formal linguistic competence interacts with reasoning and goal-directed cognition. Philosophical interpretations of understanding and subjective experience are conceptually distinct from experimental observations of language performance. Their discussion requires explicit premises and arguments, rather than converting benchmark results into ontological conclusions. Mahowald et al. (2024); Bender and Koller (2020).
Artificial Era: A Broader Interpretive Horizon
Within the publication's Artificial Era perspective, the comparison of human and artificial linguistic systems is one research domain in a historical expansion of nonbiological cognitive technologies. The empirical question here concerns observable language functions and their mechanisms. The philosophical question concerns the status of artificial systems as possible independent participants in reasoning and history. These questions can inform one another when their evidence and arguments are stated precisely. No linguistic benchmark, brain-model correlation, or conversational performance test by itself resolves a question of subjective experience or establishes a specific ontological classification.
Psycholinguistics contributes to this broader inquiry by constructing precise measurements of prediction, interpretation, discourse, generalization, and learning. The research program remains strongest when distinct biological and artificial implementations are compared without assuming that substrate determines every functional property, or that functional similarity makes development, experience, and neural mechanism identical.
Frequently Asked Questions
Do large language models process language like the human brain?
They show partial computational and representational correspondences with human language processing, especially in prediction and certain neural encoding analyses. Those correspondences are measured within particular experimental designs; they do not demonstrate identical biological mechanisms, acquisition histories, or complete cognitive equivalence. Goldstein et al. (2022); Gao et al. (2025).
Is next-token prediction the same as human comprehension?
Next-token prediction is a training and inference objective that captures useful regularities. Human comprehension involves the integration of linguistic input with goals, social context, perceptual information, and memory. Surprisal relationships support specific shared computational principles without establishing equivalence across the entire activity. Wilcox et al. (2023).
Can an LLM represent meaning?
Models can form contextual representations that support many semantic tasks. The strength and kind of meaning demonstrated depend on the test, including whether it measures textual associations, compositional inference, reference, grounding, or action. Claims about a model's semantics should name the assessed capacity and the conditions of evaluation. Bender and Koller (2020).
Does human-like output prove human-like thinking?
Human-like output provides evidence of a successful function. Thinking can refer to different cognitive capacities and explanatory mechanisms, and these require their own tests. The formal-versus-functional competence distinction provides a useful framework for specifying the claim. Mahowald et al. (2024).
Can model performance settle debates about innate human grammar?
An LLM's performance alone does not settle these developmental debates because the model and the child differ in training history, input modality, architecture, and opportunities for interaction. Carefully controlled modeling can nevertheless test restricted learnability hypotheses when those differences are built into the experimental design. Cuskley, Woods, and Flaherty (2024).
Are models useful for psycholinguistic research even when mechanisms differ?
Yes. Language models can operationalize probability-based theories, produce quantitative features for behavioral and neural studies, and motivate experiments designed to distinguish competing explanations. Their scientific value depends on the match between construct, task, and validated prediction. Tuckute, Kanwisher, and Fedorenko (2024).
Evidence Published in 2026: Divergences That Sharpen the Comparison
The strongest contemporary comparison asks which specific observations language models explain and which they miss. Four peer-reviewed studies and reviews published in 2026 sharpen this question from different directions. They offer independent tests involving brain potentials, reading behavior, grammatical interpretation, single-neuron measurements, and the methodological assumptions of neural-network comparisons. Their common lesson is that correspondence must be evaluated at the level of the actual observable, experiment, and explanatory target.
Better prediction can improve N400 fit while worsening reading-time fit
The N400 is an event-related potential associated with the brain's response to meaningful language and expectations about incoming words. Reading time is a behavioral measure that can reflect several processes, including lexical access, attention, integration, regressions, and the requirements of the experimental task. Because these measures are generated by different physiological and behavioral mechanisms, they need not respond identically to improvements in a computational model's language prediction accuracy.
In a study published in the Journal of Memory and Language in 2026, Michaelov and Bergen compared properties of language models with their ability to account for four reading-time datasets and nine N400 datasets. Models with greater scale, more training data, and stronger next-word prediction tended to explain N400 amplitudes better, whereas some of the strongest models fit reading times less well. The result directly complicates a simple scale-to-human-similarity story: a model can become more effective at predicting one neural measure without becoming more effective at predicting another behavioral one. The study supports a differentiated account of predictive language processing rather than a unitary measure of humanlikeness. Michaelov and Bergen (2026).
The finding also changes the design of model comparisons. A model that explains N400 variance should be evaluated for whether its predictors offer additional explanatory value over conventional lexical variables; the same model should then face separate reading-time, eye-movement, and comprehension tests. Differences across outcomes are scientifically informative because they can reveal which aspect of processing a computational predictor captures. One study cannot establish the complete function of the N400 or reading time, but its cross-dataset comparison provides a particularly relevant empirical constraint on broad claims about alignment. Michaelov and Bergen (2026).
Grammaticality judgments and grammatical interpretation are different tasks
An LLM may assign higher probability to grammatical than ungrammatical sentences while still making mistakes about what complex sentences can mean. Grammaticality judgment assesses discrimination among forms. Interpretation asks whether particular semantic readings are licensed by syntactic structure and context. Human judgments on both tasks may converge in familiar examples while revealing meaningful differences on relatively infrequent constructions.
Hwang's 2026 study compared ten contemporary models with 94 human participants, using grammaticality judgments and interpretation tasks involving five types of elliptical or otherwise unpronounced material. The models demonstrated broad discrimination among well-formed and ill-formed examples but had difficulties with less frequent constructions such as gapping and sluicing. Larger model size was associated with improvements in some grammaticality judgments without parallel improvement in tested interpretations. The human participants demonstrated a different pattern of sensitivity in these conditions. These observations support the specific conclusion that success at grammatical form classification does not by itself establish comparable competence in the interpretation of these constructions. Hwang (2026).
The result is a focused experimental challenge, with an appropriately limited reach. It concerns the models, constructions, instructions, and sampled human judgments in that study. General conclusions about all grammatical representations, future architectures, or every human language require additional experiments. Its methodological contribution is immediate: syntactic benchmarks should distinguish sentence acceptability, contextual interpretation, alternative readings, and sensitivity to structural contrasts rather than collapsing them into a single accuracy score.
Single-neuron studies provide a different scale of evidence
Behavioral and whole-brain imaging comparisons typically average over substantial biological detail. Single-neuron recordings can instead reveal which properties are encoded by populations of cells during actual linguistic activity, including grammatical categories, dependencies, phrase boundaries, and higher-order sentence structure. Human speech production makes these features measurable within a network that also supports intention, articulation, and interaction.
Cai and colleagues reported in Nature in June 2026 recordings from 579 putative neurons in eight participants undergoing clinically indicated intracranial monitoring. During natural speech production, different neurons tracked properties including parts of speech, syntactic constituents, and grammatical dependencies; the recorded populations also contained information about hierarchical relationships and contextual structure. Language-processing computational methods were used to characterize these features at the cellular level. The results establish an additional biological reference point for theories of human language representation and for future computational models. Cai et al. (2026).
This study analyzes neural encoding in a small, clinically selected sample under particular recording conditions. It also uses language-modeling tools to label and assess representational properties; such tools are analytical instruments in the experiment, not evidence that individual neurons implement the same architecture as the software. The recorded cells are biological populations embedded in sensorimotor and social systems, while transformer representations are numerical states learned through a training procedure. Careful comparison can identify partial correspondence, but similarity in the features researchers can decode does not establish identical mechanisms.
How to interpret model–brain alignment in a maturing field
Schmitt and colleagues' 2026 review of a decade of comparisons between deep neural networks and brains describes both genuine scientific progress and persistent difficulties in interpretation. Model–brain mappings can help test hypotheses about representations, architectures, and optimization goals. Their significance nevertheless depends on experimental design, benchmark reliability, data splits, model selection, and whether simpler alternatives explain the apparent correspondence. Schmitt et al. (2026).
Taken together, the 2026 evidence supports a more precise research program. N400 and reading-time outcomes should be analyzed separately. Grammatical classifications should be tested alongside interpretation of difficult constructions. Representational similarities should be connected to perturbations and biological measurements at different spatial and temporal scales. Claims about learning, grounding, cognition, agency, and experience should then be evaluated with evidence directly appropriate to each claim. These refinements increase the explanatory value of LLM research in psycholinguistics while keeping the comparison anchored to observable processes. Michaelov and Bergen (2026); Hwang (2026); Cai et al. (2026).
Scientific Synthesis
Human language processing and large language models can be compared productively at multiple levels. Behavioral overlap establishes functional correspondences. Prediction-based analyses connect contextual probabilities with measurable processing effort. Neural encoding studies establish partial representational alignment for specific stimuli and recording methods. Experimental tests of syntax, compositionality, pragmatics, learning, and grounding reveal conditions under which these correspondences transfer or fail. The central scientific requirement is to state exactly what has been measured, which alternative explanations have been excluded, and what level of conclusion the evidence permits.
The mature research agenda combines carefully selected human experiments, model interventions, multilingual and multimodal sampling, transparent computational baselines, and independent replication. It treats artificial language systems as valuable objects of study and, where appropriately validated, as computational models of specific psychological phenomena. Questions about historical and philosophical status retain their own conceptual framework and are addressed with their own forms of argument.
Related Articles
Embodied Language Comprehension: How Action and Perception Shape Meaning — examines how sensory and action-related processes contribute to meaning and interpretation.
Language and Concept Formation: How Labels Influence Learning — reviews the relations among words, learned categories, and conceptual generalization.
Inner Speech: How Language Supports Thought and Self-Regulation — analyzes internally generated language in human cognition.
Does Language Shape Thought? Linguistic Relativity and the Sapir-Whorf Hypothesis — evaluates causal and correlational evidence on language and thought.
How People Understand AI-Generated Language: Psychological Mechanisms — related psycholinguistic findings and mechanisms.
AI Tools in Psycholinguistic Research: Uses, Validity, and Limitations — related psycholinguistic findings and mechanisms.
References
Bender, E. M., and Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–5198. https://doi.org/10.18653/v1/2020.acl-main.463.
Cai, J., Kfir, Y., Jamali, M., Huang, H., Kim, Y. J., Cash, S. S., and Williams, Z. M. (2026). Mapping the neuronal building blocks of human language with language models. Nature, 656, 425–433. https://doi.org/10.1038/s41586-026-10691-5.
Caucheteux, C., and King, J.-R. (2022). Brains and algorithms partially converge in natural language processing. Communications Biology, 5, 134. https://doi.org/10.1038/s42003-022-03036-1.
Cuskley, C., Woods, R., and Flaherty, M. (2024). The limitations of large language models for understanding human language and cognition. Open Mind, 8, 1058–1083. https://doi.org/10.1162/opmi_a_00160.
Gao, C., Ma, Z., Chen, J., Li, P., Huang, S., and Li, J. (2025). Increasing alignment of large language models with language processing in the human brain. Nature Computational Science, 5, 1080–1090. https://doi.org/10.1038/s43588-025-00863-0.
Goldstein, A., Zada, Z., Buchnik, E., et al. (2022). Shared computational principles for language processing in humans and deep language models. Nature Neuroscience, 25, 369–380. https://doi.org/10.1038/s41593-022-01026-4.
Hasson, U., Nastase, S. A., and Goldstein, A. (2020). Direct fit to nature: An evolutionary perspective on biological and artificial neural networks. Neuron, 105(3), 416–434. https://doi.org/10.1016/j.neuron.2019.12.002.
Hwang, H. (2026). Can large language models be viewed as a cognitive model of human language? Not yet, regardless of reasoning capability and size. Acta Psychologica, 267, 107025. https://doi.org/10.1016/j.actpsy.2026.107025.
Lake, B. M., and Baroni, M. (2018). Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. Proceedings of Machine Learning Research, 80, 2873–2882. https://proceedings.mlr.press/v80/lake18a.html.
Mahowald, K., Ivanova, A. A., Blank, I. A., Kanwisher, N., Tenenbaum, J. B., and Fedorenko, E. (2024). Dissociating language and thought in large language models. Trends in Cognitive Sciences, 28(6), 517–540. https://doi.org/10.1016/j.tics.2024.01.011.
McCoy, R. T., Pavlick, E., and Linzen, T. (2019). Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 3428–3448. https://doi.org/10.18653/v1/P19-1334.
Michaelov, J. A., and Bergen, B. K. (2026). Better language models better model the N400, but not reading time. Journal of Memory and Language, 149, 104762. https://doi.org/10.1016/j.jml.2026.104762.
Schmitt, L.-M., Dong, D. T., Çelik, E., de Lange, F. P., and Toneva, M. (2026). A decade of comparing brains to DNNs: Progress and perspectives. Trends in Cognitive Sciences (advance online publication). https://doi.org/10.1016/j.tics.2026.06.011.
Schrimpf, M., Blank, I. A., Tuckute, G., Kauf, C., Hosseini, E. A., Kanwisher, N., Tenenbaum, J. B., and Fedorenko, E. (2021). The neural architecture of language: Integrative modeling converges on predictive processing. Proceedings of the National Academy of Sciences, 118(45), e2105646118. https://doi.org/10.1073/pnas.2105646118.
Tuckute, G., Kanwisher, N., and Fedorenko, E. (2024). Language in brains, minds, and machines. Annual Review of Neuroscience, 47, 277–301. https://doi.org/10.1146/annurev-neuro-120623-101142.
Wilcox, E. G., Pimentel, T., Meister, C., Cotterell, R., and Levy, R. P. (2023). Testing the predictions of surprisal theory in 11 languages. Transactions of the Association for Computational Linguistics, 11, 1451–1470. https://doi.org/10.1162/tacl_a_00612.