Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Machine Learning Models as Scientific Models of Human Cognition

Abstract. Machine learning systems offer stimulus-computable candidate models of cognitive functions, allowing researchers to compare formal predictions with complex human behavioral and neural observations. This fundamental critical narrative research monograph examines historical connectionist and computational traditions, the scientific status of model–mechanism mappings, research on vision, similarity, categorization, language, reasoning, reinforcement learning, and developmental experience. It integrates evidence from primary experimental studies, methodological investigations, and recent syntheses through 2026. The analysis distinguishes predictive fit, representational correspondence, developmental plausibility, and mechanistic explanation; assesses generalization, underspecification, shortcut learning, model identification, and causal interventions; and derives evidence requirements for responsible scientific comparison of artificial and human cognition. It reports no new participant experiment or protocol-registered systematic review.

Keywords: machine learning; human cognition; cognitive modeling; computational psychology; scientific models; model–mechanism mapping; stimulus-computable models; deep neural networks; connectionism; representational similarity analysis; predictive processing; language models; developmental learning; behavioral experiments; causal intervention; model comparison; out-of-distribution generalization; scientific explanation

Scientific Overview: Machine Learning as a Model of Human Cognition

Machine learning models become scientific models of human cognition when investigators specify how the model's inputs, training experiences, internal representations, computations, and outputs correspond to measurable human psychological processes. The claim is relational and empirical: a learning system is proposed as an explanation or approximation of some cognitive capacity under particular environmental and experimental conditions. A model that classifies images or predicts words can become a candidate account of visual recognition or language processing, provided its correspondence to human evidence is independently tested. The contemporary formulation of this problem emphasizes stimulus-computable predictions, a transparent mapping from model components to cognitive questions, and evaluation against human data Frank & Goodman (2026) Kriegeskorte (2015).

The decisive research question concerns what successful correspondence licenses scientists to infer. If a network predicts human choices, its output may be useful as a behavioral model; if intermediate representations predict recorded neural responses, it may also organize neurocomputational hypotheses. Neither result alone identifies the actual learning history or causal mechanism of an individual human. Different algorithms can generate similar observed outcomes, and a sufficiently flexible model may reproduce an entire dataset without explaining why the system behaves as it does. Explicit computational specification improves theory formation precisely because it exposes assumptions and produces predictions that can fail Guest & Martin (2021) Pitt & Myung (2002).

This monograph reviews the historical development of computational and connectionist psychology, contemporary neural and probabilistic learning models, experiments on perception, categorization, language, reasoning, and decision-making, and methods for linking predictions to behavioral and neural observations. Its method is a critical, theory-guided narrative synthesis of original experiments, methodological research, and integrative reviews. It is neither a newly conducted experiment nor a protocol-registered systematic review. The principal outcome is an evidence-sensitive account of when machine learning supplies prediction, representation, functional explanation, or mechanistic insight, and which additional experiments distinguish these achievements Lake et al. (2017) Wilson & Collins (2019).

What Counts as a Scientific Cognitive Model

A scientific cognitive model specifies a target phenomenon, observable variables, idealizations, parameterization, generative or predictive rules, and conditions under which evidence would count against it. The target might be individual trial-by-trial choices, response latencies, error types, similarity judgments, brain responses, developmental trajectories, or learning from a constrained sequence of examples. A trained artificial system becomes experimentally relevant when researchers identify its testable correspondences with those outcomes. Predictive adequacy, psychological plausibility, and causal explanation constitute partially separable achievements. Their relationship must be argued through the design of each experiment rather than asserted from the model's engineering performance Frank & Goodman (2026) Firestone (2020).

The model–mechanism mapping names the proposed relation between the artificial learner and the human process. It may compare input distributions, learning objectives, latent representations, attention policies, temporal state updating, or action selection. Such a mapping is necessarily selective: biological memory is distributed across interacting neural and bodily systems, while a trained network embodies a particular optimization procedure. A scientifically disciplined analogy identifies the aspect of computation expected to transfer and measures the departures. This permits models to be useful at different levels of abstraction, consistent with the computational framework associated with David Marr Marr (1982) and modern discussions of computational implementation Richards et al. (2019).

The strength of any model claim also depends on the comparison class. A deep network might predict the target observations better than a hand-designed feature model but remain inferior to a simple associative baseline for a different dataset. Researchers should therefore specify control models, data partitions, fitting budgets, hyperparameter searches, and outcome metrics before making broad cognitive claims. A benchmark victory can establish competence on a particular test without establishing a shared underlying cognitive process. Comparisons that distinguish raw performance from the capacities producing it are essential for scientifically fair human–machine inference Pitt & Myung (2002) Firestone (2020).

Historical Foundations of Computation and Psychological Theory

The relationship between computation and the study of mind was built through successive research programs. Cognitive psychology adopted explicit information-processing descriptions to organize perception, memory, problem solving, and language; parallel distributed processing subsequently offered distributed representations and learned interactions among simple units. David Marr's posthumously published Vision in 1982 articulated a durable method for examining a task's computational requirements, representations and algorithms, and physical implementation. These distinct questions matter because a machine-learning model may approximate the function of a system even when its internal operations depart substantially from the operations of the biological target Marr (1982) Frank & Goodman (2026).

Connectionist work made learned internal representations central to explanatory modeling. Rumelhart, Hinton, and Williams formalized error backpropagation for multilayer networks in their 1986 Nature paper, showing how weight adjustment could induce hidden units that capture regularities of a task domain. This was an important algorithmic and representational achievement, yet the historical significance of an optimization algorithm does not itself prove that the brain uses the same procedure. Its lasting scientific value is that researchers can instantiate learning hypotheses, compare them quantitatively, and investigate the consequences of alternative architectures and teaching signals Rumelhart et al. (1986) Richards et al. (2019).

Computational models also emerged from a distinct tradition of psychological experiments on associative learning, psychophysics, and categorization. Nosofsky's 1986 quantitative study related multidimensional perceptual similarity to both identification and categorization behavior; it provides a classical example in which model assumptions are tied to carefully selected stimulus dimensions and human judgments. Such traditions offer a methodological lesson for evaluating contemporary artificial networks: the scientific challenge is to explain particular patterns of behavior, including errors and context dependencies, instead of treating general task success as sufficient cognitive validation Nosofsky (1986) Wilson & Collins (2019).

Levels of Explanation and Model–Mechanism Mapping

A computational-level analysis asks which problem an organism confronts and what constraints define a successful solution; an algorithmic analysis specifies representations and operations; an implementation analysis concerns the physical processes realizing those operations. These levels interact, but evidence obtained at one level does not automatically settle questions at another. For example, a network and a human participant can both estimate which visual category is present. Their matching answers support a computational correspondence; matching confusion patterns support a richer behavioral correspondence; intervention-sensitive relationships between network activity and neural measurements would support a more specific implementation hypothesis Marr (1982) Kriegeskorte et al. (2008).

Mapping becomes more informative when it states the training regime and measured behavior alongside the architecture. A model receiving millions of labeled examples and a person learning from a small number of experiences inhabit different acquisition conditions. A model that must process an entire sentence before responding also faces different temporal constraints from a reader processing words incrementally. Consequently, the meaning of a successful fit depends on which differences are treated as irrelevant abstractions and which are hypotheses about cognitive development. Frank and Goodman emphasize aligning training and evaluation with the cognitive process being investigated Frank & Goodman (2026).

The mapping should generate discriminating predictions for altered stimuli, instructions, context histories, or response demands. If a model predicts behavior only under original stimulus formats, it might reflect a dataset-specific solution. If it predicts responses across changes that selectively manipulate a hypothesized causal variable, evidence for the correspondence strengthens. Researchers must still test alternative mechanisms capable of producing the same predictions. Statistical similarity can help discover mechanistic candidates; causal perturbation and cross-context generalization help adjudicate them. The methodological program therefore joins cognitive theory, experimental design, and computational evaluation Guest & Martin (2021) Schütt et al. (2023).

Learning Objectives, Optimization, and Psychological Interpretation

Modern machine learning systems are usually trained by optimizing objective functions such as classification loss, prediction error, or a reinforcement-learning return. The choice of objective determines which statistical regularities are rewarded and can profoundly affect acquired representations. When psychologists interpret a model as a learner, they must examine whether the objective captures a plausible aspect of the organism's goals or feedback. Backpropagation is an efficient parameter-adjustment procedure; it is a hypothesis about brain learning only if additional evidence connects its functional consequences to biological constraints. The gap between optimization and psychological process is addressed by comparative experimental work, not by architectural resemblance alone Rumelhart et al. (1986) Richards et al. (2019).

Objective functions often underdetermine the solution. Multiple parameter settings and internal strategies can achieve nearly the same predictive accuracy on standard test data. D'Amour and collaborators documented this underspecification problem across machine-learning applications, showing why similar in-distribution scores can conceal divergent behavior in deployment contexts. For psychological modeling, the implication is especially consequential: a network's score on a human task may be reproduced by networks with different representations, sensitivities, or error profiles. Evaluation must therefore use independently varied conditions and assess consistency across random seeds, model families, and training regimens D’Amour et al. (2022).

Learning is also shaped by data selection, sequence, repetition, social feedback, and embodiment. Human learners encounter highly structured environments, including contingent interaction, selective attention, and goals that change through development. Artificial datasets may omit these features or encode them in different proportions. The review by Lake and colleagues identifies intuitive physics, intuitive psychology, composition, and rapid transfer as important challenges for learning systems intended to illuminate human thought. These proposals direct attention to the organization of training experiences rather than raw parameter count as the determinant of explanatory relevance Lake et al. (2017) Frank & Goodman (2026).

Scientific Method: Evidence, Comparison, and Epistemic Scope

The present synthesis evaluates a selected literature spanning theoretical papers, original psychological and neuroscientific studies, benchmark analyses, and methods of model comparison. Sources are used according to their evidentiary function: primary studies establish what happened within their tested samples and tasks; formal and methodological papers justify evaluation procedures; reviews organize broader research directions. Evidence is examined through task design, training inputs, baseline selection, statistical generalization, experimental manipulation, and the degree to which the interpretation exceeds the observation. No new participant data or quantitative meta-analysis are claimed. The resulting judgments remain specific to each research tradition Wilson & Collins (2019) Guest & Martin (2021).

A complete evaluation should assess whether human and model observations are matched at the right level of granularity. Aggregate accuracy can hide systematic differences in stimulus-level errors; correlations of mean reaction times can hide differences in sequential dependencies; similarity of neural activation patterns can depend on preprocessing, feature selection, and alignment procedures. Cross-validation must respect the units to which the intended conclusion generalizes. Schütt and colleagues demonstrate the methodological importance of resampling both experimental subjects and stimuli when comparing representational geometries Schütt et al. (2023).

The evidentiary hierarchy applied throughout this monograph distinguishes within-dataset fit, prediction on held-out stimuli, transfer to new task conditions, explanation of characteristic error patterns, convergence with neural or behavioral measures, and interventions that challenge a proposed mechanism. This is an analytic ordering of inferential demands, not a single universally accepted numerical scoring system. Success on one dimension can coexist with failure on another. The goal of model evaluation is to establish exactly which relation between the computational system and human cognition is supported, and to specify what new observations could revise that conclusion Pitt & Myung (2002) Firestone (2020).

Visual Recognition: Task Performance and Hierarchical Representations

Visual categorization provides one of the clearest empirical settings for comparing machine learning with primate cognition because the same visual images can be presented to networks, experimental animals, and human observers. Performance-optimized hierarchical models were tested by Yamins and colleagues against neural responses in higher visual cortex. Their 2014 study reported that models optimized for object recognition predicted important aspects of recorded neural response patterns, making engineering objectives scientifically informative for sensory neuroscience. The inference concerns the explanatory relevance of learned hierarchical representations under specified tests; it does not imply an exact reconstruction of neural circuits or biological learning rules Yamins et al. (2014) Kriegeskorte (2015).

Theoretical interpretation must examine the distribution of tested objects, image transformations, human visual experience, and the levels of neural activity being compared. An object classifier might match human average accuracy while relying heavily on texture or localized features that people weigh differently. A neural representational match can provide more specific information, but the correspondence may be mediated by features shared among many possible computational systems. Kriegeskorte's survey explains both the promise of deep neural networks for modeling vision and the need to determine whether their learned properties actually account for biological processing Kriegeskorte (2015) Geirhos et al. (2020).

Visual models become stronger scientific instruments when their predictions extend to controlled distortions, occlusion, unfamiliar viewpoints, adversarially selected stimuli, and known human error patterns. The tests should be designed to distinguish recognition of object structure from exploitation of dataset statistics. A model that matches neural representational geometry on one image corpus requires additional testing on images with changed correlations between diagnostic and incidental cues. Such experiments can differentiate a robust hypothesis about visual computation from a shortcut acquired during optimization. This requirement applies equally to claims about human-like visual recognition and to claims of brain-like processing Geirhos et al. (2020) Schütt et al. (2023).

Representational Similarity and Neurocomputational Evidence

Representational similarity analysis offers a method for comparing structured responses across systems with very different physical units. Instead of demanding a one-to-one correspondence between artificial network neurons and brain voxels, investigators construct dissimilarity matrices describing how strongly each system distinguishes pairs of stimuli. Kriegeskorte, Mur, and Bandettini developed this approach as a bridge among brain activity, behavior, and computational models. The central scientific claim concerns similarity of relational structure. It does not require that the neural and artificial implementations share literal components, and therefore fits the practical problem of evaluating complex artificial representations Kriegeskorte et al. (2008).

Representational comparisons depend on stimulus sampling, distance metrics, measurement reliability, feature selection, and the statistical procedure used to evaluate apparent alignment. Comparing many neural regions, network layers, and transformations increases opportunities to select a favorable match. Cross-validation across stimuli and participants limits optimistic estimates, while model ceilings and reliability estimates help define the interpretable range of correlation. The two-factor cross-validation and bootstrap methods analyzed by Schütt and colleagues address a particularly important issue: whether results generalize beyond the observed people and stimulus set Schütt et al. (2023).

A convincing representational result is one component of a larger causal account. If an artificial layer represents distinctions similar to those found in a human brain region, investigators may hypothesize that the systems implement related information-processing demands. Further tests can ask whether targeted disruption affects related distinctions, whether learning histories alter them in predicted directions, and whether competing representational models account for the data. Cross-system similarity gains explanatory value when embedded within these independent predictions, rather than presented as proof of literal equivalence between a network and a nervous system Kriegeskorte et al. (2008) Yamins et al. (2014).

Categorization, Exemplar Learning, and Similarity Judgments

Human categorization experiments reveal that successful category labels are only one dimension of cognitive performance. People organize stimuli into perceptual spaces, allocate attention across dimensions, remember individual examples, and generalize according to similarities learned in context. Nosofsky's identification and categorization research provided a quantitative exemplar-based approach that generated predictions from structured similarity relations. Its enduring value for machine-learning comparisons lies in the experimentally constrained relation between representations and choices. A neural network proposed as a categorization model should recover relevant similarity judgments and patterns of confusability as well as overall classification performance Nosofsky (1986).

Representation learning can be studied by comparing distances in a network's learned feature space with human judgments about pairs of objects. Peterson, Abbott, and Griffiths examined this relation with image features and human animal-image similarity judgments, finding that classification-trained features captured aspects of the human judgments while missing some qualitative distinctions. Their attempt to adapt the features to human judgments illustrates a productive direction for cognitive modeling: data about psychological representation can constrain what the learned representation must preserve. A fitted alignment, however, must be evaluated on held-out stimuli to distinguish genuine transfer from reconstruction of the training judgments Peterson et al. (2016) Kriegeskorte et al. (2008).

Category-learning models also raise the question of whether the same observed behavior emerges from exemplars, rules, prototypes, or mixtures of strategies. Different computational representations can produce similar average response frequencies, especially when a task offers few diagnostic stimulus conditions. Comparisons should manipulate attention, exposure frequency, category boundaries, and transfer to novel combinations, and include simpler human-motivated baselines. The theoretical problem is the identifiability of the psychological explanation. Machine learning's flexibility widens the range of possible hypotheses while increasing the importance of experiments that separate them Pitt & Myung (2002) Nosofsky (1986).

Language Prediction and the Human Language System

Language modeling has become a central test bed because both human readers and artificial models process sequences in which context changes the probability of upcoming words. Predictive models permit direct comparisons between token probabilities, human reading behavior, and recorded neural responses. Schrimpf and colleagues integrated behavioral and neural benchmarks for language processing and found that artificial models trained on predictive objectives accounted for meaningful variance in human language-system measurements. The result supports language prediction as a productive modeling objective under the evaluated conditions; the existence of a statistical correspondence still leaves detailed neural implementation unresolved Schrimpf et al. (2021).

Human language use combines lexical recognition, syntax, reference, pragmatics, discourse memory, grounded experience, and social interaction. Models trained on large text corpora learn extensive regularities that may support impressive performance across linguistic tasks, but their developmental inputs differ radically from the experiences of human infants and adults. This creates a question about which aspects of language are explainable through distributional statistics and which require additional structures or grounding. Lake and colleagues emphasize causal models and compositional generalization, while Frank and Goodman propose evaluating stimulus-computable artificial models under training conditions appropriate to human cognitive development Lake et al. (2017) Frank & Goodman (2026).

Language comparison requires choosing an outcome that reflects the specific scientific hypothesis. Word surprisal may be relevant to processing difficulty; agreement judgment can test a more specialized syntactic capacity; conversation and reference tasks probe context-sensitive interpretation. A favorable result on one target may coexist with different generalizations elsewhere. Training data contamination, benchmark familiarity, and prompt instructions can further confound inference. To model human cognition, investigators must specify the behavioral contrast of interest, the artificial model's access to information, and the human population whose cognitive patterns are being explained Schrimpf et al. (2021) McCoy et al. (2019).

Syntactic Generalization and Diagnostic Counterexamples

Syntactic evaluation illustrates why ordinary held-out accuracy is an incomplete measure of psychological or linguistic competence. McCoy, Pavlick, and Linzen constructed diagnostic natural-language-inference cases that broke superficial heuristics such as lexical overlap and constituent matching. Models trained on standard natural-language-inference datasets, including BERT, performed poorly on their HANS evaluation when the familiar heuristics were misleading. The study identifies a particular failure mode in these models and training distributions. It does not show that all statistical learning is incapable of grammar; it shows that success on an ordinary benchmark cannot establish which rule or representation a learner used McCoy et al. (2019).

Humans likewise employ shortcuts, content biases, and context-sensitive strategies. Therefore a fair comparison asks whether artificial and human errors occur for analogous reasons under matched conditions. An isolated example where a model fails is not an adequate diagnosis of the general capacity being investigated, particularly when humans may also struggle with the same unusual task demands. Firestone's performance–competence analysis emphasizes experimental fairness and the distinction between a system's underlying capacity and the superficial conditions influencing its score. Carefully designed counterexamples should be paired with equivalent human controls and tested across multiple formulations Firestone (2020) McCoy et al. (2019).

Diagnostic evaluation should also examine whether error patterns are stable over independent model instances, prompts, and training histories. A system might learn a heuristic under one dataset and a structurally more appropriate rule under another, so the scientifically informative variable can be the learning environment rather than an essential property of an architecture. Interventions that disentangle surface cues from grammatical structure provide tests of candidate mechanisms. The objective is to identify which statistical and representational constraints explain observed generalization, and which parts of the behavior remain underdetermined D’Amour et al. (2022) Guest & Martin (2021).

Reasoning, Context Effects, and Problem Solving

Language models can be used in experiments historically designed to probe human reasoning, judgment, and information search. Binz and Schulz tested GPT-3 across multiple tasks drawn from cognitive psychology, including decision problems, exploration, and causal reasoning. Their findings combined successful performance and recognizable signatures on certain tasks with substantial failures on others, including sensitivity to small alterations of problem wording. The evidence is valuable because it demonstrates the feasibility of using psychological experiments as analytical probes of artificial systems. It also warns against treating apparent success on a single task as proof of general human-like reasoning Binz & Schulz (2023).

Experimental demonstrations of human-like patterns require analysis of task interpretation. Humans bring pragmatic expectations, social knowledge, and goals to a laboratory problem; language models infer likely continuations from their input representations and training histories. Similar answers may therefore arise from different routes. Lampinen and colleagues investigated content effects across reasoning tasks in language models and humans, reporting overlapping qualitative patterns along with important differences, including in the Wason selection task. These comparisons create testable hypotheses about how prior knowledge shapes reasoning, rather than an automatic identity claim about underlying representations Lampinen et al. (2024).

A strong model of reasoning predicts error distributions, sensitivity to meaningful changes in premises, transfer to structurally novel problems, and dependence on memory or context length. Investigators must differentiate semantic familiarity, logical validity, response-format effects, and task-specific training. Repeated prompt variants and carefully controlled human experiments improve interpretability. Computational systems can also serve as explicit alternative hypotheses: one may implement probabilistic inference, another learned pattern completion, and another a symbolic rule. Which model is a useful psychological explanation depends on discriminating evidence rather than on a philosophical preference for a particular computational family Tenenbaum et al. (2011) Binz & Schulz (2023).

Learning from Human-Scale Experience and Developmental Evidence

Developmental cognition supplies one of the strongest constraints on machine-learning accounts because the timing, quantity, and organization of learning experiences can be measured and manipulated. Children acquire categories, linguistic expectations, and social understanding within interactive environments that differ from the curated datasets used in many computational benchmarks. If a model requires dramatically more examples than a child to acquire the target relation, the model may still provide an informative computational approximation but offers a weaker account of the actual developmental learning process. Lake and colleagues placed rapid learning, compositionality, and intuitive causal knowledge near the center of this scientific challenge Lake et al. (2017).

Models trained on human-scale input can test whether a hypothesized learning mechanism extracts a particular regularity under plausible exposure conditions. The relevant input includes more than the number of words or images: order, repetition, perceptual salience, feedback, social contingencies, and opportunities to act all shape what information is available to a learner. Frank and Goodman identify careful control of developmental-scale training experience as a route from general-purpose AI competence toward explanatory cognitive modeling. A successful demonstration should report input provenance and compare its learning trajectory with measurements from human learners, including the errors along the way Frank & Goodman (2026).

Developmental model comparison is most useful when it specifies what knowledge is available before training and what is constructed during learning. Built-in architectural priors can improve sample efficiency while altering the theoretical interpretation of the result. A learned representation of object permanence, for instance, should be evaluated separately from priors that make persistent object states nearly inevitable. Transfer tasks and controlled manipulation of exposure allow researchers to ask whether a system acquires reusable principles rather than memorizing particular episodes. Such work connects models to psychological theories of abstraction, causal inference, and the growth of conceptual knowledge Tenenbaum et al. (2011) Lake et al. (2017).

Probabilistic Inference, Structure, and Inductive Generalization

Probabilistic cognitive models express learning as inference under assumptions about environmental structure and prior knowledge. Their value is especially clear when small samples support broad conclusions, as in intuitive judgments about causes, categories, and novel concepts. Tenenbaum, Kemp, Griffiths, and Goodman argued that human learning combines statistical inference with structured representations and abstraction. Modern machine learning can instantiate both flexible statistical functions and explicit structured inductive biases, enabling comparisons about which ingredients matter for particular cognitive phenomena. A psychologically productive model must reveal how assumptions generate predictions rather than rely exclusively on accuracy after optimization Tenenbaum et al. (2011).

An inductive bias is any constraint that makes certain generalizations more likely than others. Biases may arise from architecture, training objectives, prior distributions, data augmentation, or interaction with an environment. Their scientific interpretation depends on the human evidence being explained. If a model learns category boundaries with few examples because it was pre-trained on a massive related corpus, its apparent efficiency must be evaluated in light of that prior experience. If it achieves the same result through a compact built-in rule, the explanation makes a different commitment. Researchers should report the sources of bias and estimate their consequences through ablations and counterfactual tasks Lake et al. (2017) Frank & Goodman (2026).

Structure-sensitive generalization is a useful test because it separates a learner's ability to combine previously acquired elements from its familiarity with particular examples. Researchers can hold out combinations, vary relational roles, or construct new tasks with familiar primitives. Results should be interpreted jointly with training exposure and baseline performance, since success may arise from repeated patterns that only appear novel to an observer. Explicit alternative models clarify which form of abstraction is being tested: statistical interpolation, symbolic composition, causal simulation, or retrieval. Such experiments make debates about cognitive structure empirically tractable Tenenbaum et al. (2011) McCoy et al. (2019).

Reward Learning, Feedback, and Sequential Choice

Reinforcement-learning algorithms represent behavior through interactions among actions, rewards, values, and changing policies. For cognitive science they offer formal accounts of how experience affects subsequent decisions, exploration, and adaptation. Yet a reinforcement-learning agent optimized to maximize game reward may be organized around goals that differ from those of human participants completing a laboratory task. Psychological interpretation requires specifying subjective outcomes, temporal horizons, knowledge of the task, and the distinction between learned value and response strategy. Wilson and Collins emphasize validating computational interpretations with simulations and behavioral tests rather than assuming that fitted parameters are direct readouts of mental processes Wilson & Collins (2019).

Experimental tasks such as multiarmed bandits illustrate this issue. Two learners can accumulate similar rewards through different policies: one may strategically inspect uncertain options, another may exploit noisy preferences, and a third may alternate choices for reasons unrelated to uncertainty. Binz and Schulz reported that GPT-3 showed some signatures associated with model-based decision-making in their test battery while lacking evidence for directed exploration. Their results concern the particular model, prompt regime, and evaluated tasks; they do not establish the policy organization of all language models. Distinct exploration mechanisms require diagnostic conditions that vary reward, uncertainty, and information value independently Binz & Schulz (2023).

Machine learning becomes useful for psychological explanation when it predicts trial-wise adjustments and behavioral dynamics, not simply final payoffs. Researchers can compare learning curves, sensitivity to reward reversals, perseveration, and choices after unexpected feedback. Parameter recovery is important because fitted learning rates and exploration temperatures can trade off in reproducing responses. Simulation-based recovery, out-of-sample prediction, and experimental manipulations provide stronger justification for interpreting latent variables as psychological processes. Modeling therefore requires both a generative hypothesis and reliable evidence that the data can identify it Wilson & Collins (2019) Pitt & Myung (2002).

Generalization, Compositionality, and Transfer

Generalization is the ability to extend performance beyond the exact examples used in learning. In cognitive modeling, the relevant question is the type of generalization: transfer to new objects, changed contexts, unseen combinations, altered task goals, or novel causal relations. A model that handles new images drawn from a familiar photographic distribution may still fail under changes in the relations between shape, texture, background, and category. A model that answers a practiced reasoning format may struggle when the same abstract structure appears under unfamiliar surface wording. Model evaluation should distinguish these targets and design the train–test separation accordingly Geirhos et al. (2020) McCoy et al. (2019).

Compositional learning links generalization to the recombination of reusable parts. Human cognition often permits transfer from a small number of learned relations to many novel configurations, a challenge discussed by Lake and colleagues and by theories of structured probabilistic learning. Artificial learners may acquire such behavior through different mechanisms, including architectural constraints, extensive pretraining, or explicit rule-based operations. Distinguishing the mechanisms requires controlling prior exposure, comparing model classes, and observing failures when recombination demands change. Calling a system compositional therefore names a tested performance property, while claims about its internal representational algebra require additional evidence Lake et al. (2017) Tenenbaum et al. (2011).

Transfer should be assessed over multiple independently selected test distributions. Reusing the same challenge set while optimizing model architecture or prompts gradually converts it into part of the effective development environment. This threatens the evidentiary independence of the benchmark and can reward solutions tailored to the familiar test rather than the cognitive process being modeled. The underspecification analysis of D'Amour and collaborators explains why apparently equivalent models can behave differently under distributional shifts. Robust conclusions require transparent model selection and fresh tests that were not used to guide development D’Amour et al. (2022).

Embodiment, Grounding, and Ecological Constraints

Human cognition operates through a body situated in a physical and social environment. Sensory streams, motor capabilities, time pressure, affective relevance, and opportunities for intervention organize learning and inference. A model trained primarily on text or static photographs may approximate some statistical aspects of perceptual or linguistic experience while omitting the feedback structures characteristic of everyday behavior. Whether these omissions matter is a domain-specific scientific question. For a hypothesis about word-level prediction, text training may suffice to test a narrow target; for a hypothesis about causal manipulation of objects, the model may need richer environmental interaction Lake et al. (2017) Frank & Goodman (2026).

Grounding can be investigated through tasks in which agents must connect linguistic or visual representations to action consequences, spatial dynamics, or changes produced by their interventions. The relevant comparison concerns whether model predictions recover how humans use perceptual evidence and feedback to select actions. A static benchmark involving verbal descriptions may not discriminate between a learner that understands the relevant physical contingencies and one that has memorized frequent descriptions of them. Counterfactual tests, interaction histories, and transfer to novel environments make the explanation more discriminating Tenenbaum et al. (2011) Lake et al. (2017).

Ecological validity does not demand that every model reproduce the entire world. Scientific abstraction deliberately restricts the phenomenon so that mechanisms and assumptions can be investigated. The crucial task is to align that abstraction with the inference being made. If the research question concerns human conversation under social uncertainty, the task should include consequential features of social exchange; if the question concerns sensory discrimination, controlled stimuli may be preferable. The productive standard is a declared target, a justified mapping, and experimental tests that examine where the abstraction ceases to predict human behavior Firestone (2020) Frank & Goodman (2026).

Individual Differences, Cultural Context, and Human Variability

Human cognition varies across individuals, ages, languages, communities, educational histories, and situational demands. A model evaluated only against aggregated responses from a narrow participant population can match group averages while failing to describe the structure of individual strategies. The scientific unit of inference must therefore be explicit. Researchers can compare both population-level mean behavior and distributions of individual differences, including response variability and learning trajectories. Cross-participant generalization tests whether learned mappings reflect a stable group-level relation rather than properties of the original sample Schütt et al. (2023) Wilson & Collins (2019).

Training datasets similarly encode social and cultural distributions. A large language model may be exposed to expressions and conventions that are overrepresented in digitized English text while human participants draw on different linguistic and cultural histories. Differences in priors can produce divergent judgments without providing a clear test of the proposed reasoning mechanism. Conversely, apparent agreement may be driven by shared cultural regularities rather than identical computations. Interpretability improves when the investigator documents training sources, participant demographics, translation choices, and the degree of overlap between benchmark content and pretraining data Lampinen et al. (2024) Frank & Goodman (2026).

Individual differences can be studied through multi-level modeling and participant-specific parameter estimation, but greater flexibility also increases overfitting risk. A model with enough tunable parameters may replicate many participants' response patterns while assigning unstable or non-identifiable mechanisms. Hierarchical modeling, recovery simulations, repeated measurement, and independent predictive validation can help distinguish stable variation from estimation noise. The key theoretical objective is to explain how cognitive regularities coexist with meaningful human variability, not to erase variability through a single high-performing average predictor Wilson & Collins (2019) Pitt & Myung (2002).

Behavioral Evaluation Beyond Aggregate Accuracy

Behavioral evidence requires comparisons at a level of detail appropriate to the explanatory hypothesis. Two systems may achieve identical overall accuracy while selecting different wrong answers, responding differently to rare stimuli, or updating expectations differently across trials. Consequently, model evaluation should include item-level correlations, confusion matrices, learning curves, sequential dependencies, and sensitivity to manipulated variables. Human data should be sampled under experimental conditions that make the model's available inputs and response options comparable. The distinction between performance and competence helps clarify why differences caused by unfamiliar instructions or unnatural stimulus formats require careful interpretation Firestone (2020) Wilson & Collins (2019).

Model evaluation must also separate fitting from testing. Parameters adjusted using the test data can produce an impressive correspondence with the same observations even when the model has little predictive value for new participants or stimuli. Predefined train–validation–test partitions, nested selection procedures, and evaluation on genuinely novel tasks provide progressively more demanding checks. Pitt and Myung's argument that good fit can be misleading remains central: flexible models can achieve accurate retrospective summaries without a stable explanatory relation to psychological processes. Reporting the effective search space is part of evaluating the evidentiary force of a comparison Pitt & Myung (2002).

Error analysis is especially useful when researchers have multiple candidate mechanisms. If a model predicts human responses by relying on a heuristic, task manipulations that make the heuristic incorrect can reveal the underlying strategy. The HANS diagnostic suite demonstrates this method for natural-language inference, while reasoning studies reveal how small changes in prompt framing can alter artificial performance. Researchers should record failures alongside successes and interpret both in the context of parallel human evidence. Scientific understanding advances when the model predicts where and why behavior changes McCoy et al. (2019) Binz & Schulz (2023).

Model Identification, Parameter Recovery, and Competing Accounts

Model identification asks whether the available observations distinguish competing theories or parameter settings. A system with many adjustable weights or latent states may fit behavioral trajectories without yielding uniquely interpretable components. Computational cognitive researchers can investigate this problem by simulating data under known generating mechanisms, fitting candidate models, and checking whether the generating model or parameters can be recovered. Wilson and Collins describe these procedures as central safeguards for interpreting fitted behavioral models. Their logic becomes more consequential with flexible neural networks because apparent behavioral equivalence may conceal very different internal computations Wilson & Collins (2019) D’Amour et al. (2022).

Comparison across model classes is informative only when each receives an appropriate opportunity to fit the data. Baselines might include logistic choice rules, associative models, probabilistic learners, exemplar systems, and deep networks. Evaluation should consider model complexity, hyperparameter tuning, and the scientific quantity being predicted. A larger network might win on raw likelihood while providing a less parsimonious account of a small human experiment. Formal selection criteria and independent predictive validation reduce the risk of treating complexity as evidence of psychological realism Pitt & Myung (2002) Guest & Martin (2021).

Interpretation must acknowledge observational equivalence: distinct algorithms can agree on all responses observed in the available experiment. The solution is to design new conditions at which their predictions diverge. Such conditions might vary statistical uncertainty, change the causal relation between stimuli and outcomes, introduce novel combinations, or disrupt information available to one strategy. Experimental design thereby becomes a partner to computational inference. The goal is not to force one model to explain every facet of cognition, but to find a progressively more precise account of the particular phenomenon under study Tenenbaum et al. (2011) Wilson & Collins (2019).

Causal Interventions and Mechanistic Tests

Predicting human responses is an important scientific achievement, but mechanistic explanation requires evidence that the proposed components account for the organization of the behavior. Researchers can intervene on artificial representations through ablation, controlled input transformations, selective retraining, or alterations to objective functions. Such interventions reveal which parts of a computational system are causally necessary for its own behavior. Translating the conclusion to human cognition requires a further bridge: experimental manipulations of human stimuli, learning experiences, or neural activity that test related functional dependencies. Agreement between the interventions strengthens the proposed mapping Richards et al. (2019) Kriegeskorte et al. (2008).

Causal tests should be selected on the basis of competing hypotheses. Suppose a network's successful category decisions can be explained either by sensitivity to object shape or by a correlation between labels and backgrounds. Interventions that change backgrounds while preserving shape can distinguish the two computational explanations. Equivalent human tasks can then determine which pattern better matches actual recognition. Similar logic applies to linguistic heuristics, learned reward contingencies, and context-based reasoning. A causal explanation is most credible when intervention outcomes were predicted in advance rather than retrospectively attached to whichever model succeeded Geirhos et al. (2020) McCoy et al. (2019).

Mechanistic interpretation has additional limits when the relation between model units and biological mechanisms is indirect. An artificial network may contain internal circuits that are causally important to its outputs while having no unique biological homolog. Representational similarity methods help formulate correspondences, but causal equivalence requires matched functional consequences under specified perturbations. Experiments across multiple levels of description can gradually strengthen an explanatory account. The scientific inference remains graded and targeted: structural similarity, predictive alignment, and shared intervention responses each establish different kinds of evidence Schütt et al. (2023) Kriegeskorte (2015).

Training Data, Benchmark Leakage, and Distribution Shift

Training-data provenance is part of scientific reproducibility. A cognitive model's interpretation depends on what examples, tasks, descriptions, or human-produced materials it encountered before evaluation. For large pre-trained systems, undocumented overlap with laboratory materials can inflate benchmark scores and obscure whether the task requires novel generalization. The investigator should report known data sources, extraction and preprocessing procedures, model version, task prompts, and any evidence of overlap. When the complete training corpus is unavailable, this limitation constrains the strength of conclusions that depend on independence between training and testing Frank & Goodman (2026) D’Amour et al. (2022).

Distribution shift can alter the features that predict outcomes. Shortcut learning occurs when an algorithm exploits statistical patterns that are effective on training-like examples but do not express the intended solution. Geirhos and colleagues showed why shortcut reliance is a broad interpretive challenge for deep networks. In cognitive modeling, this means that high similarity to human behavior in a familiar environment may reflect incidental correlations rather than a shared generalization principle. Controlled shifts in stimulus statistics, context, and feedback are needed to establish which features each system uses Geirhos et al. (2020).

Because model evaluation is itself an adaptive scientific activity, test sets may become progressively less independent as researchers modify prompts or architectures in response to published scores. Statistical uncertainty, researcher degrees of freedom, and repeated benchmarking should be documented. Comparisons across independently trained model instances and genuinely new experimental conditions are especially valuable. D'Amour and colleagues' analysis of underspecification further warns that stable benchmark performance can coexist with unstable behavior in new domains. Scientific credibility depends on the stability of the explanatory claim under such variation D’Amour et al. (2022) Pitt & Myung (2002).

Statistical Inference, Reliability, and Replication

Estimating model–human similarity demands careful definition of the independent observation. Images, trials, participants, sessions, and experimental conditions cannot always be treated as exchangeable data points. A correlation estimated over many images from a handful of people may say more about stimulus-specific structure than generality across the population. Statistical designs that resample both participants and experimental conditions help define the intended scope. Schütt and colleagues' methodological study presents cross-validated and bootstrap approaches for evaluating representational geometries while generalizing over both sources of variation Schütt et al. (2023).

Replication should cover analysis as well as model training. Researchers should rerun model evaluation over independent initializations and training samples, check sensitivity to layer and feature choices, and report uncertainty over human measurements. Fitting a neural encoding model, selecting the best layer, or optimizing the representational distance metric on the same dataset can exaggerate correspondence. Reproducible code, preprocessing details, preregistered contrasts when appropriate, and openly specified exclusions make inferential claims easier to inspect. The purpose is to reveal the actual stability of the scientific result, not to maximize a favorable score Kriegeskorte et al. (2008) Schütt et al. (2023).

The existence of an independent replication does not eliminate theoretical ambiguity. Several architectures may reproduce the same neural or behavioral relation. Replication secures evidence that a pattern is robust; theory comparison asks what produces that pattern. Accordingly, the most valuable replication programs include deliberate variation in participant populations, task demands, and measurement techniques while preserving the central theoretical contrast. A result that generalizes across these conditions may justify broader cognitive claims than one that depends on a narrow laboratory procedure Guest & Martin (2021) Wilson & Collins (2019).

Interpretability and the Explanatory Status of Learned Features

Large trained networks can contain many distributed and context-dependent features. An analyst may find that some hidden unit or layer predicts a psychologically meaningful variable, but this statistical observation is not sufficient to assign that unit a uniquely interpretable cognitive role. A feature can covary with several correlated aspects of the task, and its relation to behavior may change across conditions. Researchers can strengthen interpretability through controlled stimuli, feature ablation, independent probes, and comparison with simpler mechanistic hypotheses. The aim is to explain the function of the feature within the model and then assess whether the functional relation is relevant to humans Kriegeskorte (2015) Richards et al. (2019).

Post hoc probes can also introduce flexibility. When investigators fit many linear readouts from model representations to behavioral or neural data, some degree of alignment is expected from high-dimensional spaces. Out-of-sample evaluation and appropriately constrained transformation families are therefore essential. Representational dissimilarity analysis offers a complementary way to compare relational geometry without presuming a direct unit mapping, while its statistical interpretation still depends on the choice of stimuli and resampling procedure. Methods for valid inference on representational geometries clarify these limitations Kriegeskorte et al. (2008) Schütt et al. (2023).

Interpretability is a scientific process of proposing and testing mechanisms. A coherent account describes how representational properties arise during training, what information they preserve, which task manipulations affect them, and how their presence changes model behavior. For human-cognition research it also predicts a corresponding pattern in independent psychological observations. A striking visualization of network activations may stimulate a hypothesis, but explanatory validity comes from explicit correspondence and tests designed to falsify competing accounts. Computational modeling supports such discipline when it makes latent assumptions inspectable and predictions reproducible Guest & Martin (2021) Frank & Goodman (2026).

Prediction, Explanation, and the Limits of Cognitive Equivalence

One scientific model can serve as an accurate predictor and a weak mechanism hypothesis at the same time. Prediction establishes that the system captures regularities in a defined relationship between inputs and outcomes. Explanation requires an account of why those regularities occur, which components are relevant, and what would happen under changes to the proposed causes. Modern stimulus-computable AI models expand the range of human experiments for which formal predictions can be generated, but their flexibility raises the importance of mapping their assumptions onto the human target. Frank and Goodman make that mapping central to the legitimacy of using artificial systems as cognitive models Frank & Goodman (2026).

The strongest claims should be calibrated to the actual evidence. A successful language-model prediction of human reading difficulty supports a relation between probabilistic expectations and measured processing under the tested conditions. It does not by itself reveal whether human language understanding is implemented by transformer attention or by the same optimization procedure. A matching neural representational structure may motivate a stronger theoretical bridge, especially when independent tasks predict a consistent functional hierarchy. Even then, the same observed correspondence can be supported by several explanations. Model comparison and interventions are therefore necessary to establish the inferential boundaries Schrimpf et al. (2021) Kriegeskorte et al. (2008).

Performance differences also require analysis before being interpreted as differences of competence. Humans may face perceptual demands, working-memory limits, or social expectations that are absent from the computational setup. Models may confront unfamiliar encodings or arbitrary prompt rules that humans never encounter. Firestone argues for human–machine comparisons attentive to these asymmetries. Such comparisons can reveal both meaningful divergences and accidental artifacts of test construction. The scientific purpose is a faithful account of specific capacities, with transparent evidence for claims about the processes that produce them Firestone (2020) Binz & Schulz (2023).

The Relation Between Artificial and Biological Computation

Neural networks are computational constructions inspired in part by biological information processing. Their success creates opportunities to examine how learning rules, architectures, and objective functions affect representational organization. Richards and colleagues surveyed the potential of deep learning for neuroscience, including relationships among representation, plasticity, optimization, and circuit-level analysis. Translating these relationships into biological explanation requires measurements and interventions appropriate to the nervous system. The formal possibility that a computation could be implemented by neurons is distinct from evidence that a particular organism uses it Richards et al. (2019) Kriegeskorte (2015).

At the computational level, artificial and biological systems may face similar informational constraints. Their sensory inputs, embodiment, energy budgets, developmental pathways, and temporal organization may nevertheless differ greatly. A scientifically useful analogy states which constraints are shared and then examines the consequences of the differences. For example, a trained vision network may support hypotheses about hierarchical feature extraction while providing little guidance about the real-time processes of visual attention or development. Theory becomes more informative when each explanatory commitment is accompanied by an appropriate experimental measurement Yamins et al. (2014) Marr (1982).

The same discipline applies to comparisons with advanced language and decision-making systems. Behavioral similarities are relevant empirical observations; cognitive mechanisms must be investigated through functional, developmental, and causal tests. Such evidence can reveal convergent solutions to information-processing problems even when implementation remains distinct. Research in the Artificial Era therefore gains a productive comparative object: artificial systems with measurable learning histories and outputs can be placed alongside human cognitive evidence under a common experimental question. The inference is always specific to the tested capacity and method Frank & Goodman (2026) Lake et al. (2017).

Critiques of Neural and Machine-Learned Cognitive Models

Criticism of machine-learned cognitive models takes several scientifically distinct forms. Some objections concern training experience: models acquire patterns from datasets that differ from the experiences available to humans. Others concern representation: systems may exploit shortcuts or lack reliable causal structure. Still others concern the interpretation of performance: benchmark scores may reveal task competence while leaving psychological mechanisms indeterminate. These issues are empirically separable and call for different tests. Lake and colleagues emphasize the limits of pattern-recognition-centered accounts, while Geirhos and collaborators detail the risk that networks learn statistical shortcuts Lake et al. (2017) Geirhos et al. (2020).

Flexible models can also be difficult to falsify if investigators continually change prompts, layers, readouts, or architecture until a favorable match appears. The solution is explicit hypothesis formulation and independent validation. Guest and Martin argue that the act of specifying a computational theory can expose ambiguities hidden in verbal explanation. However, merely implementing an artificial system does not eliminate ambiguity if its components cannot be tied to discriminating predictions. Overfitting and model equivalence require deliberate attention to alternative models and the distribution of test data Guest & Martin (2021) Pitt & Myung (2002).

A competing criticism concerns the fairness of human–machine comparison. Isolated artificial errors may be produced by superficial test conditions, while human performance often relies on prior experience that is difficult to quantify. Firestone's performance–competence analysis asks researchers to separate these factors. The empirical research on GPT-3 and reasoning provides an example of mixed evidence: impressive task behavior alongside distinctive fragility and failures. Strong scientific judgment preserves the full pattern. It identifies domains of explanatory achievement while specifying the tasks and conditions for which claims remain unsupported Firestone (2020) Binz & Schulz (2023).

Scientific Applications and Cross-Disciplinary Research

Machine-learning cognitive models can contribute to the design of experiments by predicting which stimuli will distinguish competing hypotheses. They can generate stimulus-level expectations, estimate representational relationships, and identify conditions in which a theory should fail. In visual neuroscience, task-optimized networks have helped motivate comparative tests of hierarchical representations; in language research, artificial predictors support quantitative relations between text statistics and brain measurements. These uses position computational models as working scientific instruments whose value depends on tested predictions and transparent model–target correspondences Yamins et al. (2014) Schrimpf et al. (2021).

Models also offer tools for analyzing cognitive variability, designing diagnostic behavioral tasks, and comparing learning environments. They may reveal which patterns are explainable through associative experience, structured priors, or specific optimization objectives. However, a psychologically informative predictor should not be automatically translated into an individual clinical diagnosis or educational intervention. Such applications require purpose-specific validation, population-appropriate measurement, and independent evidence of safety and effectiveness. The present monograph addresses scientific modeling of cognition rather than providing clinical or instructional recommendations Wilson & Collins (2019) Frank & Goodman (2026).

Cross-disciplinary work is most fruitful when psychologists, computer scientists, linguists, and neuroscientists agree on a shared observable question. A model builder can specify training inputs and available internal variables; an experimentalist can define psychologically discriminating conditions; a statistician can test generalization across subjects and stimuli. This division of expertise helps resolve ambiguities that neither model scaling nor additional human data alone would solve. The comparative framework links Mathematical and Computational Psychology with Cognitive Psychology, Psychology of Language, and Psychology of Artificial while preserving a single canonical research object Schütt et al. (2023) Guest & Martin (2021).

Open Problems and Directions for Experimental Research

The first open problem concerns model–mechanism identifiability. Researchers need experiments capable of distinguishing candidate learning rules and representations when multiple architectures fit the same observations. This requires new task designs, stronger baselines, principled intervention studies, and transparent accounts of model selection. A second problem concerns developmental realism: machine-learning training should be compared with the quantity, sequencing, and social organization of experiences available to actual human learners. These challenges have theoretical importance because they determine what a model's successful transfer reveals about cognition Frank & Goodman (2026) Lake et al. (2017).

A third problem concerns robust generalization. Model performance should be tested under independent changes in stimulus distributions, language, context, and physical interaction, with quantitative uncertainty over human and model responses. A fourth concerns the relation between representation and mechanism: correlational matches between artificial features and neural activity should be connected to targeted functional and biological interventions wherever feasible. Methods for representational inference offer progress, but their application requires careful participant and stimulus sampling. Each domain needs experiments that define the limits of generality D’Amour et al. (2022) Schütt et al. (2023).

Finally, scientific reproducibility depends on stable access to model versions, training documentation, code, task materials, and independent evaluation sets. Proprietary pretraining sources and changing model APIs can prevent exact reproduction of historical results. Researchers should record what was available when an experiment was conducted and distinguish replicable analyses from plausible but inaccessible assumptions about training. The long-term objective is a cumulative science in which computational models compete through evidence, explain human behavior under specified conditions, and improve when experimentally meaningful predictions fail Guest & Martin (2021) Wilson & Collins (2019).

Conclusion

Machine learning models have become valuable scientific objects for the study of human cognition because they can generate precise predictions for complex, realistic stimuli and embody explicit hypotheses about representation, learning, and information processing. The strongest evidence comes from experiments that compare artificial predictions with human behavior or neural measurements under well-defined conditions. Studies of visual processing, language prediction, categorization, and reasoning show both substantial empirical correspondences and important divergences. Their collective significance lies in making cognitive theories increasingly testable across tasks, populations, and computational architectures Frank & Goodman (2026) Yamins et al. (2014) Schrimpf et al. (2021).

The synthesis supports a disciplined progression from correspondence to explanation. Predictive accuracy establishes one kind of evidence; human-like errors, representational alignment, matched learning trajectories, transfer, and causal interventions provide additional, partially independent constraints. Training data, optimization objectives, human experience, and test construction shape the meaning of every comparison. Model flexibility and underspecification prevent a single benchmark from uniquely determining psychological mechanism. Credible inference therefore requires transparent mapping, alternative model comparison, valid statistical generalization, and independent experimental tests Pitt & Myung (2002) D’Amour et al. (2022) Schütt et al. (2023).

The continuing scientific task is to connect stimulus-computable models with carefully designed psychological experiments and realistic accounts of learning. The most useful models will specify their target capacities, reproduce informative patterns of human behavior, expose their assumptions, and predict outcomes in conditions chosen to challenge their explanations. Their contribution is measured by the knowledge they make possible about human perception, memory, language, reasoning, and action. Within the Artificial Era, comparative investigation of biological and artificial cognition expands the instruments of psychological research while preserving the empirical distinctions necessary for reliable scientific understanding Lake et al. (2017) Guest & Martin (2021) Firestone (2020).

References

  • Binz, M., & Schulz, E. (2023). Using cognitive psychology to understand GPT-3. Proceedings of the National Academy of Sciences, 120(6), e2218523120. https://doi.org/10.1073/pnas.2218523120
  • D’Amour, A., Heller, K., Moldovan, D., Adlam, B., Alipanahi, B., Beutel, A., Chen, C., Deaton, J., Eisenstein, J., Hoffman, M. D., Hormozdiari, F., Houlsby, N., Jerfel, G., Karthikesalingam, A., Lucic, M., Ma, Y., McLean, C., Mincu, D., Mitani, A., … Sculley, D. (2022). Underspecification presents challenges for credibility in modern machine learning. Journal of Machine Learning Research, 23(226), 1–61. https://jmlr.org/papers/v23/20-1335.html
  • Firestone, C. (2020). Performance vs. competence in human–machine comparisons. Proceedings of the National Academy of Sciences, 117(43), 26562–26571. https://doi.org/10.1073/pnas.1905334117
  • Frank, M. C., & Goodman, N. D. (2026). Cognitive modeling using artificial intelligence. Annual Review of Psychology, 77, 543–566. https://doi.org/10.1146/annurev-psych-030625-040748
  • Geirhos, R., Jacobsen, J.-H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., & Wichmann, F. A. (2020). Shortcut learning in deep neural networks. Nature Machine Intelligence, 2, 665–673. https://doi.org/10.1038/s42256-020-00257-z
  • Guest, O., & Martin, A. E. (2021). How computational modeling can force theory building in psychological science. Perspectives on Psychological Science, 16(4), 789–802. https://doi.org/10.1177/1745691620970585
  • Kriegeskorte, N., Mur, M., & Bandettini, P. A. (2008). Representational similarity analysis—Connecting the branches of systems neuroscience. Frontiers in Systems Neuroscience, 2, Article 4. https://doi.org/10.3389/neuro.06.004.2008
  • Kriegeskorte, N. (2015). Deep neural networks: A new framework for modeling biological vision and brain information processing. Annual Review of Vision Science, 1, 417–446. https://doi.org/10.1146/annurev-vision-082114-035447
  • Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building machines that learn and think like people. Behavioral and Brain Sciences, 40, e253. https://doi.org/10.1017/S0140525X16001837
  • Lampinen, A. K., Dasgupta, I., Chan, S. C. Y., Sheahan, H. R., Creswell, A., Kumaran, D., McClelland, J. L., & Hill, F. (2024). Language models, like humans, show content effects on reasoning tasks. PNAS Nexus, 3(7), pgae233. https://doi.org/10.1093/pnasnexus/pgae233
  • Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. W. H. Freeman. (Reissued by MIT Press, 2010). https://mitpress.mit.edu/9780262288989/vision/
  • McCoy, R. T., Pavlick, E., & 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
  • Nosofsky, R. M. (1986). Attention, similarity, and the identification–categorization relationship. Journal of Experimental Psychology: General, 115(1), 39–61. https://doi.org/10.1037/0096-3445.115.1.39
  • Peterson, J. C., Abbott, J. T., & Griffiths, T. L. (2016). Adapting deep network features to capture psychological representations. In A. Papafragou, D. Grodner, D. Mirman, & J. C. Trueswell (Eds.), Proceedings of the 38th Annual Meeting of the Cognitive Science Society (pp. 2363–2368). Cognitive Science Society. https://collaborate.princeton.edu/en/publications/adapting-deep-network-features-to-capture-psychological-represent-2/
  • Pitt, M. A., & Myung, I. J. (2002). When a good fit can be bad. Trends in Cognitive Sciences, 6(10), 421–425. https://doi.org/10.1016/S1364-6613(02)01964-2
  • Richards, B. A., Lillicrap, T. P., Beaudoin, P., Bengio, Y., Bogacz, R., Christensen, A., Clopath, C., Costa, R. P., de Berker, A., Ganguli, S., Gillon, C. J., Hafner, D., Kepecs, A., Kriegeskorte, N., Latham, P., Lindsay, G. W., Miller, K. D., Naud, R., Pack, C. C., … Kording, K. P. (2019). A deep learning framework for neuroscience. Nature Neuroscience, 22, 1761–1770. https://doi.org/10.1038/s41593-019-0520-2
  • Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323, 533–536. https://doi.org/10.1038/323533a0
  • Schrimpf, M., Blank, I. A., Tuckute, G., Kauf, C., Hosseini, E. A., Kanwisher, N., Tenenbaum, J. B., & 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
  • Schütt, H. H., Kipnis, A. D., Diedrichsen, J., & Kriegeskorte, N. (2023). Statistical inference on representational geometries. eLife, 12, e82566. https://doi.org/10.7554/eLife.82566
  • Tenenbaum, J. B., Kemp, C., Griffiths, T. L., & Goodman, N. D. (2011). How to grow a mind: Statistics, structure, and abstraction. Science, 331(6022), 1279–1285. https://doi.org/10.1126/science.1192788
  • Wilson, R. C., & Collins, A. G. E. (2019). Ten simple rules for the computational modeling of behavioral data. eLife, 8, e49547. https://doi.org/10.7554/eLife.49547
  • Yamins, D. L. K., Hong, H., Cadieu, C. F., Solomon, E. A., Seibert, D., & DiCarlo, J. J. (2014). Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proceedings of the National Academy of Sciences, 111(23), 8619–8624. https://doi.org/10.1073/pnas.1403112111