RESEARCH DOSSIER
Cross-System Functional Comparison
Abstract. Cross-system functional comparison investigates whether human, other biological, and engineered systems exhibit comparable discrimination, memory, adaptation, control, reasoning, and social prediction under explicitly specified experimental conditions. This fundamental research monograph provides a structured narrative synthesis of foundational psychology, comparative cognition, computational neuroscience, and the experimental study of artificial systems. It distinguishes observed task performance from latent capacity, algorithmic mechanism, physical implementation, and claims about subjective experience. Scientific evidence is evaluated through construct validity, measurement invariance, task calibration, causal perturbations, distribution shift, representational similarity, benchmark replication, and sensitivity to learning and training histories. Primary cases include honeybee relational discrimination, human–ape cognitive comparisons, comparative inhibitory control, deep reinforcement learning, artificial visual recognition, neural-model benchmarking, and psychological experiments with language models. The synthesis identifies conditions under which functional correspondences are defensible, explains why similar output scores do not establish equivalent mechanisms, and presents a reproducible cross-system comparison protocol. It concludes with priorities for longitudinal, mechanistic, and independently replicated research in the Artificial Era. This is a critical narrative scholarly review, not a preregistered systematic review, meta-analysis, or report of newly collected experimental data.
Keywords: cross-system functional comparison; comparative psychology; artificial cognition; comparative cognition; functional equivalence; measurement invariance; construct validity; experimental design; human–artificial comparison; psychological measurement; representational similarity; artificial intelligence; machine learning; transfer and generalization; cognitive control; Artificial Era
Introduction: The Scientific Domain of Cross-System Comparison
Cross-system functional comparison investigates whether, how, and under what conditions distinguishable biological and engineered systems perform comparable information-processing and regulatory functions. The comparison may involve humans, other animals, neural models, robotic controllers, and artificial language or decision systems. A scientific claim requires an identified function, a specified opportunity to perform it, an observable measurement, and a set of rival process explanations. The object of inquiry is therefore a relation among tasks, capacities, mechanisms, histories, and environments. Shared behavioral success supports a bounded statement of functional achievement; a claim about common representation, understanding, or psychological organization requires additional evidence. Marr (1982) Lake et al. (2017)
Four questions organize the research. Which function is being evaluated, and what observable evidence would establish it? Are test conditions sufficiently comparable to support a joint interpretation? Which competing mechanisms could produce the measured result? How does performance change after perturbation, transfer, delayed testing, resource constraints, or a change of environment? The questions apply to a child learning a classification, a crow discriminating objects, a trained network labeling images, and a language model solving written problems, but their answers must preserve each system's available inputs, actions, and learning history. Comparability is a research achievement established by design, rather than an assumption imposed by a common task label. Herrmann et al. (2007) Binz & Schulz (2023)
The fundamental analytical unit is a system-in-task configuration. Its description names the system version or organism and developmental stage, the stimulus distribution, available sensory and output channels, instructions, training and familiarization, external memory or tools, and the scoring rule. With these elements recorded, an investigator can separate the function under study from access to that function and from incidental features of the apparatus. A failure may indicate an inaccessible stimulus, ambiguous instruction, unfamiliar action, or mismatched incentive. A success may reflect a learned shortcut rather than the proposed abstraction. Comparative inquiry therefore requires multiple converging tests and explicit accounts of alternative routes to performance. Giurfa et al. (2001) Geirhos et al. (2020)
This monograph presents a critical narrative synthesis of established psychological, comparative, computational, and neuroscientific research. It reconstructs conceptual foundations, evaluates experimental paradigms and measurement validity, and develops practical principles for designing cross-system evidence. It reports neither newly collected observations nor a preregistered systematic-review or meta-analysis. The conclusions are tied to published findings and the explanatory reach of those findings. Within the Artificial Era, function-matched inquiry becomes especially significant because new engineered systems perform activities previously examined mainly in biological cognition. The research task is to identify precise commonalities and differences while preserving their theoretical and empirical scope. Kriegeskorte & Douglas (2018) Borsboom et al. (2004)
Historical Development of Comparative Functional Inquiry
Comparative psychological research developed around the problem of inferring capacities from behavior across organisms whose bodies, environments, and evolutionary histories differ. Animal learning and experimental psychology introduced controlled response measures, whereas ethological and ecological traditions emphasized the natural circumstances in which a capacity becomes useful. The gradual convergence of these approaches established a lasting methodological lesson: an outcome has to be interpreted alongside the organism's sensory access, motivational regime, developmental experience, and ecological opportunity. Contemporary comparative experiments retain this lesson even when the research question concerns abstract categories, executive control, or social inference. Herrmann et al. (2007) MacLean et al. (2014)
The twentieth-century study of human judgment, learning, and cognitive control added process models that linked observed errors and response patterns to candidate internal operations. Analyses of heuristic judgments showed that successful adaptation can coexist with systematic bias, while parallel distributed processing accounts demonstrated how automatic behavior and attentional control can emerge through graded pathway strengths. These contributions remain relevant when an artificial system and a person produce the same answer for different reasons. Outcome-level correspondence alone leaves learning trajectories, sensitivity to conflict, and internal organization unresolved. Tversky & Kahneman (1974) Cohen et al. (1990)
Computational cognitive science transformed the comparative problem by treating explicit models as candidates whose predictions could be tested against behavior. Marr's account distinguished the computational purpose of an operation, the representations and algorithms used to achieve it, and the physical implementation of those procedures. Its influence extends well beyond vision: researchers can now ask whether two systems address the same task, use similar decision rules, and embody comparable mechanisms as separate empirical questions. This separation permits productive functional comparison even when hardware or anatomy differs, while preventing an unwarranted leap from similar scores to shared internal processes. Marr (1982) Kriegeskorte & Douglas (2018)
Artificial agents supplied additional experimental cases. Reinforcement learning formalized how policies change through interaction with reward-bearing environments, deep networks yielded strong performance on perceptual benchmarks, and search-based systems achieved striking results in tightly defined games. Each success generated new comparisons with biological cognition and fresh questions concerning generalization, data demands, causal reasoning, and transfer outside the training environment. Comparative psychology thus expanded into a reciprocal research program: biological results suggest useful computational tests, while artificial models supply explicit mechanistic hypotheses for psychological investigation. Sutton & Barto (2018) Mnih et al. (2015) Silver et al. (2016) Hassabis et al. (2017)
Defining Function, Capacity, Performance, and Mechanism
A function is specified in relation to an input–output or state-transition problem, its success conditions, and the environment within which success matters. A capacity denotes a system's ability to perform that function across a stated class of conditions. Performance is the actual result obtained in a particular trial or sampling regime. A mechanism is an organized causal process that explains how performance is generated. These four terms support different levels of inference. A model can match a human accuracy score without matching human error dependencies, can match errors without matching learning, and can match both while employing a different physical realization. Marr (1982) Kriegeskorte et al. (2008)
Discrimination concerns distinguishing task-relevant conditions; retention concerns preserving usable information across delay or interference; adaptation concerns altering behavior under changed contingencies; control concerns selecting and regulating actions amid competition; generalization concerns applying learned relations to novel cases. Each construct requires an operational description of what varies and what remains fixed. For example, a delayed matching task couples perception, memory, response mapping, and reward history. It cannot independently establish a pure memory capacity unless the design controls those other determinants. The same reasoning applies to prompts that test an artificial model's apparent recall or abstraction. Giurfa et al. (2001) Putnick & Bornstein (2016)
Functional equivalence should be treated as a graded, domain-indexed conclusion. Two systems may have comparable accuracy for a specified class of images while diverging under image corruption; they may follow similar preference patterns in a decision task while differing in exploration; or they may retrieve a fact at equal rates while having radically different exposure histories. An exact match of all possible behaviors is unattainable in ordinary research. Studies instead define populations of tasks, confidence intervals for estimated differences, and planned tests where candidate explanations make divergent predictions. A comparison becomes explanatory when success and failure patterns constrain the mechanism space. Binz & Schulz (2023) Recht et al. (2019)
Several concepts remain analytically separate from functional comparison. Subjective experience, personhood, intention, welfare, and normative responsibility each raise additional theoretical and evidentiary problems. A report or output may be evidence within a design, but the ability to produce human-like sentences cannot alone settle whether a system has the same psychological state as a human speaker. Likewise, a bird's social anticipation or a machine's object recognition cannot be interpreted as a comprehensive inventory of minds. The present work establishes bounded functional inferences and explicitly tracks their relation to, rather than resolution of, adjacent attribution questions. Bender & Koller (2020) Tomasello et al. (2003)
Explanatory Levels and the Problem of Multiple Realization
Explanatory levels provide a disciplined vocabulary for identifying genuine similarity. At the computational level, investigators state the function and why that function solves a task. At the algorithmic level, they describe the representations, update rules, search procedures, and control architecture. At the implementation level, they identify the biological or engineered substrate and its constraints. A comparison at one level may be informative without entailing correspondence at another. When two agents solve an identical discrimination, agreement at the computational level is possible even when their sensory transduction, memory systems, or learning dynamics differ markedly. Marr (1982) Yamins & DiCarlo (2016)
The possibility of functionally similar behavior in physically distinct systems motivates research on multiple realization. Bird cognition supplies a biological illustration: corvid and parrot performance can invite comparison with primate abilities despite substantial differences in forebrain anatomy and evolutionary history. Reviews of avian cognition document sophisticated functions and highlight structural and functional correspondences among avian and mammalian neural systems. These findings support investigation of convergent capacities while requiring detailed neural, behavioral, and phylogenetic evidence before claiming mechanistic equivalence. They do not establish that every cognitive function arises from one common architecture. Güntürkün & Bugnyar (2016)
Computational neuroscience offers a more restrictive approach than surface behavioral comparison by asking whether candidate models predict patterns in neural and behavioral responses. Representational similarity analysis compares the geometry of response patterns without requiring a direct match between every neuron and model unit. Brain-Score-style benchmarking integrates multiple neural and behavioral tests, making it possible to discover where apparently successful artificial vision models approximate observed primate processing and where they diverge. Such correspondences strengthen a specified model's explanatory credentials, but predictive alignment remains dependent on measurement, stimulus sampling, and competing candidate models. Kriegeskorte et al. (2008) Schrimpf et al. (2020)
Explanatory adequacy consequently depends on the question being asked. An engineering evaluation may only need reliable task completion. A psychological theory may require human-like error structures, developmental learning curves, context sensitivity, and causal response to intervention. A neuroscientific model may additionally need neural response predictions and plausible organization of computations. The investigator should declare the intended level before reporting a comparison. Retrospectively promoting behavioral accuracy into neural or experiential equivalence creates a mismatch between evidence and conclusion. Cross-system functional inquiry is most productive when it allows useful partial correspondences while making unresolved levels explicit. Kriegeskorte & Douglas (2018) Lake et al. (2017)
Evidence Selection and Critical Review Method
The research strategy is a structured narrative review using historical primary works, controlled experiments, peer-reviewed methodological analyses, computational studies, and syntheses that clarify interpretive disputes. Sources were selected for relevance to functional constructs, direct comparison, measurement validity, causal identification, and counterexamples to superficial score equivalence. Major empirical cases include insect relational learning, human–ape social cognition, cross-species self-control, machine perception and transfer, language-model behavioral experiments, and neural-model benchmarking. The review does not claim exhaustive database coverage, duplicate screening, preregistration, risk-of-bias scoring across an enumerated corpus, or pooled effect estimation. Giurfa et al. (2001) Herrmann et al. (2007) Recht et al. (2019)
The evidential hierarchy is question dependent. Randomized perturbation and cross-condition transfer tests can discriminate competing process hypotheses in a particular apparatus. Cross-sectional species comparisons inform evolutionary questions when relevant confounding and shared ancestry are addressed. Model-to-brain comparisons test predictive correspondence under selected measurements. Psychometric invariance studies clarify whether observed scores have comparable meanings across groups. Each design has distinct validity threats. No single source type can carry every inference, and the strength of a conclusion depends on the closest available design that can test it. MacLean et al. (2014) Kriegeskorte et al. (2008) Chen (2007)
Positive findings are considered alongside distribution shift, task contamination, the effect of interface modifications, the possibility of shortcut learning, and replication problems. An impressive score on a benchmark becomes narrower evidence when similar new images or minor rewording alter performance. Conversely, a failure under a new interface is not automatically evidence of the absence of a function; investigators must show that the altered test preserves access to the relevant task. Reviewing competing explanations therefore involves specifying what evidence would distinguish them, and how much of a reported finding survives reasonable changes in measurement and sampling. Geirhos et al. (2020) Binz & Schulz (2023) Open Science Collaboration (2015)
Interpretive conclusions are classified as observed result, estimated association, candidate mechanism, methodological recommendation, or theoretical hypothesis. Where published abstracts alone provide the accessible evidence, this synthesis limits detail to what those abstracts support. Broad numerical generalizations are excluded unless the underlying sample and statistic are documented. The final research priorities follow directly from limitations identified in the experimental traditions: task calibration, shared constructs, causal interventions, longitudinal exposure histories, multi-environment sampling, transparent model versions, and independent confirmation. Borsboom et al. (2004) Putnick & Bornstein (2016) Lake et al. (2017)
Construct Validity in Cross-System Psychological Measurement
Cross-system comparison requires an account of what a score is supposed to measure. Construct validity concerns the interpretation supported by evidence connecting observed responses with the proposed psychological attribute. Cronbach and Meehl's classic formulation framed validation as a problem of theoretical relations among a construct, observable indicators, and competing accounts. Borsboom and colleagues later emphasized the causal relationship between an attribute and test outcomes. Taken together, these approaches demand that comparative researchers specify why a particular observation should change when a function changes, and why other sources of variation cannot explain the same effect. Cronbach & Meehl (1955) Borsboom et al. (2004)
Validity is conditional on the task, target system, and inferential claim. A language model's selection of an answer to a text vignette measures the probability of that answer under a configured prompt and generation regime. A human response in the same vignette additionally reflects comprehension of the instructions, beliefs about the research situation, response costs, and social expectations. Identical written questions may therefore generate superficially aligned data with different response processes. A valid comparison seeks convergent evidence from related tasks, discriminant evidence from theoretically distinct tasks, and interventions that distinguish candidate mechanisms. Binz & Schulz (2023) Bender & Koller (2020)
A test should be described in terms of target property, operational indicator, response requirements, scoring transformation, and expected sensitivity to manipulated variables. For discrimination, signal-to-noise conditions and response bias may be separately estimated. For memory, recognition and delayed reproduction may impose different reconstruction requirements. For control, switching costs can confound perceptual demands and motor or output latency. Investigators should avoid interpreting a composite score as a unified ability when its components have heterogeneous dependencies. The resulting validity argument should be made at the level actually supported by multiple measures. Davis & Goadrich (2006) Cohen et al. (1990)
Cross-system inference is especially vulnerable to anthropocentric task design. Humans can usually read instructions, point, type, and explain strategies; animals may require extensive stimulus familiarization and species-appropriate rewards; artificial agents may respond through tokens, action APIs, or simulated sensors. A score that penalizes one system's interface disadvantages may measure access rather than the intended function. Removing such disadvantages also changes the task, so equivalence requires a documented calibration strategy with independent checks. Validation is an ongoing empirical program extending beyond a single favorable performance comparison. Herrmann et al. (2007) MacLean et al. (2014)
Measurement Invariance and Differential Task Functioning
Measurement invariance examines whether a measurement model relates observed responses to a latent construct comparably across groups or occasions. In human psychometrics, researchers commonly examine increasingly restrictive hypotheses concerning shared factor structure, factor loadings, intercepts, and residual properties. Putnick and Bornstein's review shows both the importance and the inconsistent reporting of these tests. These techniques are valuable where different groups complete genuinely comparable indicators, but their assumptions cannot be carried unchanged from human questionnaire responses to robots, insects, or language models without a defensible measurement model. Putnick & Bornstein (2016)
For cross-system tasks, a parallel concept is functional measurement equivalence. A design should ask whether observed differences reflect the target capability under comparable demand, rather than unequal perceptual access, instruction parsing, training exposure, or action cost. Statistical invariance of a latent variable offers one possible test when the measurement structure supports it. Elsewhere researchers may need anchor stimuli, item response curves, matched difficulty functions, adaptive psychophysical calibration, and carefully evaluated response mappings. Measurement equivalence is thus a collection of evidentiary requirements, not a single threshold automatically available in every organism or system. Chen (2007) Borsboom et al. (2004)
Formal invariance testing also has limitations. Chen's simulation study demonstrated that fit-index changes depend on the pattern and proportion of noninvariant parameters. Mechanical reliance on a universal fit cutoff can therefore conceal substantively important differences. For cross-system designs, the limitations are amplified when item independence, latent structure, or common response scale is doubtful. An artificial model's confidence score, for example, need not be psychometrically interchangeable with a person's verbal confidence. Before estimating group means, researchers must justify the common scale or restrict their claims to task-level outcomes. Chen (2007) Putnick & Bornstein (2016)
When strict comparability fails, the failure can itself be scientifically informative. A task may reveal that biological and artificial systems encode uncertainty differently, show different response biases, or depend on distinct informational cues. Such differences do not invalidate all comparison; they refine the question to a level where reliable observations remain interpretable. Researchers can compare normalized detection curves, patterns of error across shared stimulus changes, or within-system treatment effects while acknowledging that latent quantities are not identical. The preferred outcome is an explicit comparison map specifying invariant, partially invariant, and system-specific components. Cronbach & Meehl (1955) Davis & Goadrich (2006)
Task Matching, Calibration, and Ecological Validity
Function-matched experiments begin with an experimentally stated problem rather than a copied stimulus. An image that a human can interpret immediately may require preprocessing for an artificial visual system; an object-displacement task suitable for a primate may be inaccessible to an insect whose perceptual ecology is different. Investigators should establish sensory and output accessibility before the main comparison, then match challenge structure, exposure, incentives, and feedback as far as scientific relevance permits. Because exact physical equality can produce functional inequality, the logic of matching must be justified with pilot measurements and control conditions. Giurfa et al. (2001) Herrmann et al. (2007)
Calibration can be organized in stages. First, identify a minimally demanding control that demonstrates that each system can perceive the stimuli and express a response. Second, map performance across difficulty levels to avoid floor and ceiling effects. Third, introduce the diagnostic manipulation intended to separate proposed processes, such as rule transfer, reward reversal, temporal delay, or altered contextual cues. Finally, assess whether the effect replicates with new examples, apparatus variations, and experimenters. This sequence creates a more interpretable comparison than testing all agents at a single arbitrary difficulty. MacLean et al. (2014) Putnick & Bornstein (2016)
Ecological validity asks whether the measured operation generalizes to environments of genuine functional importance. A honeybee learning an abstract matching rule in a laboratory offers evidence about relational discrimination, yet its natural use of vision and navigation motivates different ecological questions. A language model solving a written causal vignette may show fluency under textual descriptions without demonstrating successful action in a physical causal environment. A game-playing policy can optimize within a stable rule system while remaining untested under novel rules or objectives. Comparative claims should specify the world of tasks to which the measurements are intended to generalize. Giurfa et al. (2001) Binz & Schulz (2023) Silver et al. (2016)
Training and familiarity must be recorded as scientific covariates rather than relegated to procedural footnotes. A child learns through years of multisensory social experience, a captive animal through ecological exposure and laboratory shaping, and a model through pretraining corpora, fine-tuning, reinforcement signals, retrieval systems, and prompts. Equal test-time exposure does not equate total learning histories. Even fully matched final scores may reflect different data demands or learning efficiencies. Studies that compare sample efficiency should therefore declare what counts as a training example, including prior structure, inherited inductive biases, and externally supplied representations. Lake et al. (2017) Sutton & Barto (2018)
Experimental Controls, Perturbations, and Causal Identification
An informative cross-system experiment is designed around rival hypotheses. Suppose two agents classify familiar images correctly. One hypothesis attributes this performance to object-sensitive representations, and another to background texture or dataset artifacts. A diagnostic study varies background, color, viewpoint, occlusion, and image statistics while preserving object identity. An agent whose accuracy collapses under the manipulation provides evidence against the broader representation claim. Geirhos and colleagues' analysis of shortcut learning shows why machine learning systems often exploit predictive regularities that differ from the investigator's intended causal structure. Geirhos et al. (2020)
Perturbation logic extends to reasoning and decision making. A small alteration to a vignette can distinguish a stable relation from memorized lexical patterns; a reward reversal can test policy updating; removal of an external cue can test offloading dependence. Comparisons should preregister primary outcomes and specify whether a perturbation genuinely leaves the targeted function intact. A model may fail because a seemingly innocuous textual change produces an unusual input distribution, while a human may fail because a manipulation changes social interpretation. Causal claims gain strength when alternative explanations are challenged by multiple independent manipulations. Binz & Schulz (2023) Risko & Gilbert (2016)
The unit of randomization and the unit of analysis must align. In human experiments, participants contribute correlated trials. In animal studies, subjects may be nested within species, laboratories, or social groups. In artificial-system studies, repeated prompts sent to a deterministic configuration do not create independent system-level samples, and repeated stochastic outputs remain conditional on a fixed model and task. Appropriate designs separate individual, item, species, model, seed, and experimenter variation. Hierarchical models, bootstrap uncertainty over appropriate units, and out-of-sample validation can improve inference, provided assumptions and uncertainty are transparently reported. Putnick & Bornstein (2016) Open Science Collaboration (2015)
Intervention validity remains central. A procedure that changes one system's internal state or input pathway may have no analogous causal role in another. For example, perturbing a neural population and ablating a network component both change processing, but the units may have different functional organization and developmental dependence. Such interventions can nonetheless be compared at the level of response changes when the relevant explanatory mapping is specified. The analysis should identify the actual manipulated variable, the expected direction of each theory's prediction, and the extent of experimental control over competing pathways. Kriegeskorte & Douglas (2018) Schrimpf et al. (2020)
Metrics, Uncertainty, and Statistical Interpretation
Accuracy is useful when a task has unambiguous scored answers, but it conceals prevalence, response bias, confidence calibration, latency, and the pattern of errors. Detection tasks may require sensitivity and criterion estimates; asymmetric classification problems may favor precision–recall analysis over a single accuracy number. Davis and Goadrich showed how receiver operating characteristic and precision–recall curves relate while responding differently to class imbalance. A comparative protocol should select metrics corresponding to the scientific construct and disclose base rates, thresholds, and any optimization performed on evaluation data. Davis & Goadrich (2006)
Psychological tasks frequently produce multiple meaningful measures: error categories, reaction time distributions, speed–accuracy tradeoffs, individual adaptation trajectories, confidence judgments, and response consistency. Artificial systems have additional operational measures such as inference cost, context length, token consumption, retrieval dependence, and robustness across prompts. These metrics should be interpreted as costs and constraints within the tested architecture. A human response latency and a cloud model's wall-clock time are influenced by incompatible hardware, communication, and scheduling factors; raw numerical equality does not imply a common psychological speed construct. Cohen et al. (1990) Bommasani et al. (2023)
Statistical claims require uncertainty at the appropriate level. A difference observed across a small group of animals, a fixed set of human volunteers, or one machine-model checkpoint should not be generalized to entire species or model families without sampling justification. Confidence intervals for item effects and participant or model effects may diverge. Cross-species studies add phylogenetic dependence, while large benchmark collections introduce selection and evaluation reuse. The inferential population should be described alongside each estimate: stimuli, tasks, subjects, environments, software versions, and deployment conditions. MacLean et al. (2014) Recht et al. (2019)
Replication is particularly important where scoring rules, prompts, devices, or model weights can change. Independent studies should preserve enough methodological detail to repeat a comparison and probe whether an effect survives plausible variants. The Open Science Collaboration's multi-study replication project highlighted difficulties in reproducing published psychological effects, although its results cannot be mechanically transferred to all comparative or AI experiments. The general implication is a demand for transparently reported uncertainty and repeatability rather than treating a single significant contrast or leaderboard rank as a durable scientific property. Open Science Collaboration (2015) Geirhos et al. (2020)
Discrimination and Relational Abstraction Across Species
Perceptual discrimination is a natural starting point for cross-system comparison because its stimulus conditions can be systematically varied and its behavioral consequences can be measured. Yet the apparently simple question of whether an agent distinguishes two categories immediately becomes a question about sensory access, prior exposure, task instructions, and generalization. In an experiment on honeybees, Giurfa and colleagues trained subjects in matching-to-sample and nonmatching-to-sample procedures and tested transfer to stimuli not used during initial training. The resulting performance furnished evidence that insects can learn relational sameness and difference under these particular experimental conditions. It did not imply an unrestricted human-like concept of identity, and its interpretation depends upon the controls that ruled out memorization of the initially rewarded pictures. Giurfa et al. (2001)
A functional comparison with a human participant or image-processing model can hold the relational structure of the task constant while altering the display, training, and response mechanism to preserve accessibility. Success on novel matching items would provide evidence for transfer within the sampled relation, but the degree of abstraction still depends upon how far the novel items depart from the training distribution. Texture, color, spatial arrangement, number of objects, and reward histories may each supply alternative cues. Studies should therefore test multiple stimulus families and ask whether an inferred relation survives transformations designed to disrupt shortcuts without changing the relevant rule. This logic extends from elementary sensory discrimination to the evaluation of learned visual classifications in deep networks. Giurfa et al. (2001) Geirhos et al. (2020)
Comparisons between human children and other great apes show why an overall cognitive score cannot replace a task profile. Herrmann and colleagues administered a comparative test battery intended to distinguish physical from social cognitive skills. Their reported pattern concerned differences across particular skill domains and groups of participating subjects, rather than an absolute ranking of all psychological abilities. Differences in language comprehension, incentive structure, familiarity with human experimenters, and developmental setting remain important when interpreting apparent group differences. The strongest comparative inferences describe the component tasks and the hypotheses each was designed to discriminate. Herrmann et al. (2007)
The same requirement governs comparisons between artificial vision systems and biological observers. A classifier may attain high nominal image accuracy yet preferentially exploit local texture, background, or source-distribution regularities. A biological observer may use object shape, scene context, prior expectation, and active sampling in ways that differ from the tested model. An experimentally informative comparison maps both systems' errors and sensitivities across coordinated stimulus perturbations. Functional similarity is then inferred from the structure of convergent responses across a defined task family, not from one headline accuracy value. Geirhos et al. (2020) Recht et al. (2019)
Retention, Memory, and Prospective Requirements
Memory comparisons require a statement of the information to be retained, the delay or interference interval, the encoding opportunity, and the retrieval requirement. A biological organism's delayed choice is influenced by attention at encoding, motivational state, environmental cues, and available motor actions. An artificial system's apparent recall may instead rely upon persistent context, retrieval-augmented access, fixed parameters, or an external memory store. Both systems can supply correct delayed answers, but the causal operations producing those answers may differ substantially. Identical output accuracy therefore establishes a limited functional correspondence under the configured memory task, while process equivalence remains an independent hypothesis. Marr (1982) Risko & Gilbert (2016)
A rigorous comparative experiment distinguishes retention from immediate cue use and memory from inference at the moment of the test. Researchers can vary the retention interval, present misleading or interfering cues, alter context, and compare recognition with free reproduction. They should also specify whether the system had access to the correct answer through a database, preceding prompt, visible stimulus, or established task routine. External aids are part of the functional configuration and should be documented rather than silently removed from one system while remaining available to another. Such controls allow investigators to estimate the dependence of successful performance upon internal retention and environmental support. Risko & Gilbert (2016) Lake et al. (2017)
Prospective behavior raises additional questions about how information is used to guide later action. A system may select an option that would be advantageous in a future state, yet that observation alone cannot determine whether the system simulated the future, followed a cached rule, or reproduced an explicit recommendation learned during training. Competing accounts can be tested by changing the contingencies between an initial learning phase and a subsequent choice, or by presenting novel combinations of goals and constraints. A behavioral experiment gains explanatory force when the competing models predict different responses before the evidence is collected. Sutton & Barto (2018) Gershman et al. (2015)
In comparisons of human and artificial memory, storage capacity is only one dimension. Memory can be evaluated by what is encoded, what is forgotten, how uncertainty is expressed, whether recalled information can be recontextualized, and how retention contributes to planning. A model with a large context window might correctly reproduce material it was just supplied, whereas a child may recover an event from a socially and emotionally organized autobiographical history. A carefully defined task can compare retrieval function, but any claim that the same psychological memory system underlies the result requires longitudinal, mechanistic, and phenomenological evidence appropriate to that further claim. Lake et al. (2017) Kriegeskorte & Douglas (2018)
Executive Control, Inhibition, and Comparative Self-Regulation
Executive control research examines how an agent maintains task goals, resists competing responses, switches rules, and adjusts actions following conflict. In human experimental psychology, the Stroop task and computational accounts of its interference patterns have served as important tests of how automatic and controlled pathways interact. Cohen and colleagues' parallel distributed processing account connected observed effects to differences in pathway strength and contextual control. The model illustrates how a psychological construct can be approached through explicit process assumptions and contrasting trial conditions rather than inferred directly from a single response time or error rate. Cohen et al. (1990)
Comparative self-control research extends these questions across species. MacLean and collaborators reported a broad study relating performance in inhibitory-control tasks to comparative biological variables across multiple taxa. The scientific significance of such work lies partly in its attempt to apply comparable tasks to organisms with different ecological and sensory conditions. The resulting associations must be considered alongside species sampling, test familiarity, phylogenetic dependence, motivation, and task-specific motor affordances. Performance on an inhibition task describes a defined problem-solving situation and cannot function as an all-purpose measure of intelligence. MacLean et al. (2014)
Artificial agents add a useful causal dimension because their decision policies, context inputs, reward contingencies, or control modules can sometimes be modified directly. An artificial agent can suppress a high-value immediate response under a task-specific constraint, update a policy after negative feedback, or learn to allocate computational resources. Such behavior can be compared with biological inhibition at the level of choice structure and adaptive control. Its interpretation depends upon whether suppression arises through explicit hard-coded rules, policy optimization, an interaction history, or an inferred representation of future consequences. An engineered inhibition mechanism and a human executive-control process may produce similar actions while differing in organization and developmental origin. Sutton & Barto (2018) Mnih et al. (2015)
A strong design for control comparison uses several tasks that separate competing demands. Go/no-go performance can be paired with reversal learning, conflict monitoring, task switching, and sustained-goal conditions, with response interfaces calibrated for each system. Analyses should compare not only final accuracy but error recovery, temporal adaptation, susceptibility to distractors, and the transfer of learned control across contexts. If an agent succeeds only because task instructions explicitly prohibit a response, the result must be distinguished from generalized autonomous self-regulation. The cross-system question concerns the mapping from changing requirements to organized action under stated constraints. Cohen et al. (1990) Lake et al. (2017)
Ecology, Development, and Evolutionary Explanation
An organism's capacities emerge within developmental and ecological histories. Human children acquire language and cultural routines, other animals learn through species-specific activity and social opportunities, and artificial agents are trained on computationally curated data under engineered objectives. These histories shape the meaning of later test performance. A model's high one-shot classification score may reflect intensive prior exposure to related patterns, while an animal's gradual improvement may reflect unfamiliar equipment rather than low underlying discrimination capacity. Comparative experimental designs must therefore document exposure histories and compare learning curves whenever the research question involves acquisition rather than final competence. Herrmann et al. (2007) Lake et al. (2017)
Evolutionary comparison introduces questions concerning shared ancestry, convergence, and adaptation. Similar functions in distant taxa can motivate hypotheses about recurrent ecological challenges and different neural solutions. Evidence from birds underscores the value of comparing functional organization across distinct forebrain structures. Evolutionary explanations, however, require comparative phylogenetic and ecological support in addition to laboratory demonstrations of a capacity. Engineered systems have design and optimization histories rather than biological descent, so the term evolution has a different causal meaning unless an explicit evolutionary search process is used. Güntürkün & Bugnyar (2016) MacLean et al. (2014)
Developmental trajectory is itself a potential functional indicator. Two agents can produce similar mature behavior while one acquired it through gradual social learning and the other through large-scale supervised exposure followed by short task instructions. Comparing only their end states would ignore the distribution of examples, the timing of feedback, and the inductive assumptions enabling learning. Developmentally informed experiments examine sample efficiency, ordering effects, critical dependencies, and generalization after limited experience. Such designs can reveal both common learning principles and specific organizational differences. Lake et al. (2017) Sutton & Barto (2018)
Ecological and developmental findings also qualify simplistic interpretations of machine superiority or deficiency. A system may be optimized to perform a narrow benchmark exceedingly well while lacking robust competence under a different distribution or embodied setting. A biological organism may fail a contrived task yet succeed reliably in its natural environment. The proper comparison states the environment and the costs and benefits of success, then considers whether different systems have been given commensurate opportunities. The scientific result is a functional profile indexed to circumstances, not a timeless scalar ranking of biological and nonbiological forms of cognition. Recht et al. (2019) Giurfa et al. (2001)
Reinforcement Learning and Artificial Adaptive Function
Reinforcement learning supplies an explicit mathematical language for comparing adaptive behavior. In standard formulations an agent observes states or observations, chooses actions, receives rewards, and updates a policy or value function to improve expected cumulative outcomes. Sutton and Barto's treatment clarifies the distinction between a policy's behavioral result, the algorithm used to update it, and the structure of the environment in which the policy is learned. Those distinctions provide a tractable basis for asking whether biological and artificial systems adapt to comparable contingencies, how quickly their policies change, and how they respond when previously useful actions become disadvantageous. Sutton & Barto (2018)
Deep reinforcement learning introduced architectures combining learned representations with policy optimization. Mnih and colleagues demonstrated human-level or better performance on selected Atari games from visual inputs using a deep Q-network. The headline achievement applies to a defined collection of games and evaluation conditions; it cannot establish general human-level intelligence across domains. For functional comparison, the important findings concern the ability of a computational agent to learn temporally extended action policies from feedback, the dependence of learning on architecture and experience, and the possibility of comparing failure patterns or sample requirements with biological learning. Mnih et al. (2015)
Search-based artificial agents provide a different configuration. Silver and collaborators combined deep networks and tree search in a Go-playing system that achieved a substantial advance within the formal domain of that game. The system's competent moves show that effective decision making can arise from an engineered combination of learned evaluation and explicit search. A psychological comparison with expert human play should therefore separately investigate move selection, uncertainty, search allocation, expertise acquisition, and transfer to changed game rules. Similar moves do not show that human experts use the identical algorithm or that the artificial system shares the expert's broader practical life. Silver et al. (2016)
Adaptive function should be tested using learning curves, reversal conditions, delayed consequences, reward misspecification, and transfer to environments with altered contingencies. The researcher must also characterize the artificial agent's pretraining and shaped objectives, and the biological agent's prior reinforcement and motivational context. An apparently faster learner may rely upon greater hidden prior experience, a more favorable action interface, or a strong built-in policy prior. When these sources of advantage are documented, cross-system research can compare which forms of adaptation have been demonstrated and how much relevant experience each system required. Sutton & Barto (2018) Gershman et al. (2015)
Artificial Vision, Neural Benchmarking, and Representational Alignment
Artificial visual recognition offers some of the most mature opportunities for systematic brain–machine comparison. Goal-directed deep networks can be trained to classify large sets of images and then evaluated on previously unseen examples. Yamins and DiCarlo described how such models can serve as hypotheses about representations within sensory cortex. Neural model evaluation is demanding: researchers test how well features derived from the model predict observed neural responses and whether the same model explains patterns in animal or human behavior. A successful match on one neural dataset remains evidence about selected measurements, not a complete model of biological vision. Yamins & DiCarlo (2016)
Representational similarity analysis permits comparisons between response patterns measured in different systems. Rather than requiring an artificial unit to correspond to an individual neuron, it compares the relational geometry among activity patterns elicited by a set of stimuli. This method can identify whether two systems represent similar distinctions while allowing their internal units to be organized differently. Its inference depends upon stimulus sampling, choice of dissimilarity measure, noise properties, and the discriminative strength of alternative models. Representational similarity is therefore a valuable intermediate level between common output scores and claims of identical mechanism. Kriegeskorte et al. (2008)
Integrative benchmarking extends the approach by evaluating models against multiple neural and behavioral constraints. Schrimpf and colleagues argued for combining evidence across tasks and measurements to accelerate the search for mechanistic accounts of intelligent processing. A model that predicts neural measurements but fails a behavioral transfer test has a different explanatory profile from one that reproduces behavior yet poorly predicts measured neural responses. Combining such tests reveals selective correspondences and deviations. Researchers should preserve the distinction between a benchmark's predictive scoring method and a causal explanation of why the biological system produces its observed response. Schrimpf et al. (2020) Kriegeskorte & Douglas (2018)
Robustness testing demonstrates the limits of benchmark generalization. Recht and colleagues constructed a new ImageNet test set and observed that many classifiers performed worse than their original benchmark evaluations suggested. Geirhos and colleagues synthesized evidence that learned image classifications can depend on shortcuts rather than the intended visual object relations. These findings show that a model's nominal visual success is sensitive to evaluation distribution and available cues. A cross-system vision study can use carefully controlled stimulus changes to investigate whether model and observer rely upon similar information, then extend the result only to the sampled visual conditions. Recht et al. (2019) Geirhos et al. (2020)
Language Models as Objects of Behavioral Experiment
Language models can be studied experimentally through their distributions of responses to carefully constructed prompts. Binz and Schulz applied methods drawn from cognitive psychology to GPT-3, examining performance on canonical tasks and probing how linguistic presentation affects observed behavior. Their work demonstrated the methodological value of treating a language model as an experimental system whose responses can be manipulated, measured, and compared. The resulting behavioral patterns do not imply that the model's internal processes reproduce those of a human participant. They establish specific similarities and dissimilarities that can motivate more discriminating experiments. Binz & Schulz (2023)
Prompts are an integral component of the system-in-task configuration. Instructions, conversational history, examples in context, output format requirements, decoding parameters, and access to tools can each change the response. A comparison using one prompt risks treating an accidental interface choice as a property of the underlying model. Repeated testing with predefined prompt families, counterbalanced presentation, held-out problems, and declared model identifiers produces a more robust estimate of functional behavior. Changing model versions or providers without recording the substitution impairs replication and can confound historical comparisons. Binz & Schulz (2023) Bommasani et al. (2023)
A language model's success on a written psychology task also raises questions about construct interpretation. Textual selection among alternative judgments can reproduce some tendencies observed among human subjects, but the two systems differ in embodiment, response incentives, cumulative learning history, and social interaction with the investigator. Bender and Koller analyzed the distinction between linguistic form and the grounding needed for meaning-related claims. Their theoretical argument is relevant to experimental design because a score on a text-only task supports a claim about performance in that format, while grounding, action, and lived social understanding require separate observations and explanatory criteria. Bender & Koller (2020)
Evaluation should include more than correctness. Relevant measures comprise consistency across paraphrases, calibration under uncertainty, sensitivity to irrelevant framing, robustness on novel problems, response dependence on demonstrations, and information access through retrieval or tools. The Holistic Evaluation of Language Models program advocates evaluation along multiple dimensions rather than exclusive attention to aggregate accuracy. This broader approach is particularly useful when psychological constructs are borrowed from human research: it helps distinguish functional accomplishments from artifacts of the test interface and highlights the limits of an isolated benchmark result. Bommasani et al. (2023) Binz & Schulz (2023)
Reasoning, Compositional Generalization, and Causal Transfer
Reasoning comparisons are strengthened by tasks that require new combinations of familiar elements, rather than recall of known answer templates. A system may perform well on familiar logical puzzles because similar formulations were abundant in its training material. To evaluate compositional generalization, investigators can alter surface content, recombine relations, counterbalance the position of valid conclusions, and test previously unseen rule structures. Human subjects may also rely on learned heuristics, so a valid comparison examines errors, response strategies, learning trajectories, and sensitivity to task framing in both groups. Tversky & Kahneman (1974) Lake et al. (2017)
Causal transfer imposes a stronger requirement than predictive association. An agent may select an accurate outcome because it has learned a correlation between inputs and outputs. A causal claim asks whether the same system adjusts correctly when the generating structure changes, when an intervention alters an upstream variable, or when an action has a newly specified consequence. Such tests require explicit alternative models. A language model can verbally describe an intervention, while a robot or reinforcement learner may implement actions in an environment; the functional target and evidentiary reach differ unless experimenters provide a defensible mapping between them. Gershman et al. (2015) Sutton & Barto (2018)
Lake and colleagues identified compositionality, causal model building, intuitive physics, and learning-to-learn as important challenges for machine intelligence inspired by human cognition. Their framework is a research program rather than a claim that every biological learner already succeeds uniformly on all relevant tasks. It provides diagnostic criteria for constructing tests that go beyond surface matches. An experiment can assess whether a newly learned rule transfers to unfamiliar objects, whether a model uses information about interventions, and whether previously acquired structure supports efficient adaptation. The results should be reported for each tested domain, sample, and training regime. Lake et al. (2017)
A cross-system comparison of reasoning should distinguish normative accuracy, descriptive response pattern, and proposed causal process. Human heuristic effects need not imply a general defect in cognition; an artificial system's avoidance of one bias does not imply it possesses a more general form of rationality. Computational rationality approaches evaluate behavior relative to objectives and resource limitations, offering one way to interpret adaptation in different systems. Nevertheless, the applicable utility function, information set, and resource costs must be stated before a performance difference can be described as more or less rational. Tversky & Kahneman (1974) Gershman et al. (2015)
Benchmark Leakage, Distribution Shift, and Model Versioning
Benchmark evaluation is vulnerable to differences between the intended scientific test and the information already available to the tested system. A language model may have encountered the same problem or a close paraphrase during pretraining. A classifier can exploit correlations that recur between training and evaluation images. Human participants likewise arrive with learned task conventions and cultural exposure. These issues require careful interpretation because a high held-out score measures generalization only relative to the actual separation between training and evaluation distributions. Training-data disclosure, independent test creation, and carefully designed transformations are therefore essential components of comparative inference. Recht et al. (2019) Geirhos et al. (2020)
Distribution shift should be analyzed as a family of changes rather than a binary condition. Shifts may affect input appearance, category prevalence, sensory noise, reward contingencies, social interaction, or the relationship between cues and outcomes. A model's performance under these conditions can expose reliance on shortcuts, while biological responses can reveal perceptual or motivational dependencies. Researchers should report the shift-generating procedure and whether the altered inputs preserve the target construct. Otherwise an apparent generalization failure could reflect task corruption or loss of stimulus accessibility rather than a meaningful difference in functional organization. Geirhos et al. (2020) Herrmann et al. (2007)
Artificial-system versioning adds a reproducibility challenge absent from many classical psychological tasks. Model weights, tokenizer, prompt template, generation temperature, retrieval index, safety policies, external tools, and system instructions can all influence an experiment. A comparison may cease to be reproducible after a remote provider silently changes the model. Reporting the version and surrounding configuration, and preserving test materials when licensing and privacy allow, can make results more interpretable. Versioned model evaluations should also distinguish repeated draws from one configuration from independent replications across systems and research teams. Binz & Schulz (2023) Bommasani et al. (2023)
The methodological response is a layered evaluation protocol: predefine primary hypotheses; create independent and sufficiently novel tests; record the entire inference configuration; probe paraphrase and environmental robustness; analyze performance by task family; and report both favorable and unfavorable outcomes. Where data provenance is uncertain, the degree of possible leakage should be marked as a limitation of the inferred capacity. These practices align comparative psychological inquiry with broader efforts to improve experimental reproducibility and to prevent overgeneralization from performance on historically reused benchmark sets. Open Science Collaboration (2015) Recht et al. (2019)
A Reproducible Protocol for Cross-System Functional Comparison
A scientific comparison should begin with a functional hypothesis that could be falsified by observations. Investigators first define the target operation and its domain, such as discrimination of relational similarity after a change in stimulus appearance. Next they specify which systems will be compared, the input and output affordances of each, the relevant developmental or training histories, and the expected alternative mechanisms. Only after these decisions should a shared task battery and statistical analysis plan be established. Starting from a leaderboard or a striking example risks selecting a construct retrospectively to explain a result. Cronbach & Meehl (1955) Lake et al. (2017)
The second stage establishes access and calibration. Each system should demonstrate elementary perceptual or informational access to the materials, an intelligible response pathway, and performance above chance on controls requiring less of the target capacity. Researchers then sample several difficulty levels and determine whether the task is saturated by floor or ceiling effects. Calibrated controls and cross-system anchor items should be retained in the eventual comparison so that an observed difference can be interpreted relative to access and task difficulty. Statistical measurement invariance testing is appropriate only when the available data and construct model meet its assumptions. Putnick & Bornstein (2016) Chen (2007)
The third stage identifies mechanisms through intervention. Researchers compare novel and familiar stimuli, vary contextual cues, reverse contingencies, impose delays, remove information sources, and test transfer across tasks. The primary outcomes should include the pattern of failures alongside successes and, where meaningful, learning rate, confidence calibration, or behavioral variability. A robust analysis considers whether the manipulation has equivalent functional significance across participants or models. In this approach a common experimental label is the beginning of a comparability argument, while successful controls and discriminating results provide its evidence. Geirhos et al. (2020) Giurfa et al. (2001)
The fourth stage supports cumulative science. Analysis plans should define sampling units, uncertainty intervals, missing observations, multiple-comparison procedures, and the practical magnitude of effects. Investigators should preserve human task instructions, animal training protocols, artificial model identifiers, software configurations, stimuli, and scoring scripts to the extent ethically and legally possible. Independent replication should probe alternate stimulus sets, research groups, and model versions. Reports should label each conclusion at the level of task result, estimated functional generality, candidate mechanism, or broader theoretical inference. This permits future studies to extend rather than merely repeat a comparison. Open Science Collaboration (2015) Schrimpf et al. (2020)
Comparative Evidence Profiles and Interpretive Boundaries
The combined literature supports a central conclusion: different kinds of systems can demonstrate overlapping functionally defined achievements within specific experimental domains. Honeybees can transfer a relational matching rule in tested conditions; primates can exhibit differentiated profiles of social and physical problem solving; reinforcement-learning agents can learn high-performing game policies; and artificial vision systems can predict selected neural and behavioral responses. These are separately established findings from different evidentiary traditions. Their joint importance lies in the possibility of studying capacities across substrates, while their limits lie in the specificity of task, sampling, and inference. Giurfa et al. (2001) Herrmann et al. (2007) Mnih et al. (2015) Schrimpf et al. (2020)
Discrimination is most directly comparable when stimuli, sensory access, and response criteria can be controlled. Memory adds the problem of retention intervals, environmental cue use, and different forms of storage. Adaptation requires attention to prior training data and the dynamics of feedback. Control introduces competing actions and resource constraints. Social inference demands especially careful construction of alternatives to a proposed representation of another agent's mental states. These domains form a set of separable experimental programs. There is no single score that can validly collapse their performance into a universal ranking without an independently defended construct model. MacLean et al. (2014) Krupenye et al. (2016) Borsboom et al. (2004)
Mechanistic evidence can strengthen some comparisons beyond behavioral overlap. Neural response prediction and representational similarity analyses provide evidence that artificial models capture aspects of biological information processing. However, even a high prediction score does not establish shared learning history, complete functional organization, or psychological experience. Model–brain comparison is therefore strongest when multiple independent measurements converge and experimental perturbations discriminate the candidate computational explanations. A failure of alignment can be theoretically valuable because it identifies where the model's explanatory architecture requires revision. Kriegeskorte et al. (2008) Kriegeskorte & Douglas (2018)
Cross-system conclusions should be expressed as qualified empirical propositions specifying the task family and tested systems. For example, a model and a group of human observers may exhibit correlated mistakes on a particular image set, while diverging under altered texture cues. The resulting conclusion concerns a measurable pattern of response dependence. It should not be redescribed as an identity of perception, cognition, intentionality, or consciousness. Such distinctions are essential to preserving the inferential meaning of comparative results and identifying the additional experiments needed to support broader claims. Geirhos et al. (2020) Bender & Koller (2020)
Critical Discussion: Competing Models and Evidentiary Disputes
One persistent scientific dispute concerns whether shared competence is best explained through common computational principles, distinct learned shortcuts, or interactions with external support. Connectionist and reinforcement-learning approaches can reproduce numerous target behaviors using explicit architectures, yet a successful model can remain only one of several underdetermined explanations. Distinct mechanisms may generate matching observable response distributions on a limited task. Investigators therefore need new interventions, transfer challenges, and mechanistic predictions that discriminate among candidate accounts. A model's predictive ability is evidence of adequacy for the measured outcomes, not proof of explanatory uniqueness. Marr (1982) Kriegeskorte & Douglas (2018)
A second dispute concerns the reach of psychological vocabulary applied to artificial systems. Terms such as memory, planning, understanding, and attention are often used operationally in computational research, but the terms also belong to theoretical traditions about biological and social minds. The scientific solution is to define the specific operation, system configuration, and measured consequences each time the term is used. Functionally described recall may involve retrieval from an external database; functionally described planning may involve explicit tree search; conversationally appropriate responses may emerge from statistical language learning. Whether these operations warrant broader psychological attribution requires an argument that extends beyond the existence of a familiar output. Bender & Koller (2020) Silver et al. (2016)
A third dispute concerns comparability itself. Strictly identical physical procedures may systematically advantage one species or interface, while tailored tasks can introduce differences in difficulty and reward structure. Neither strategy automatically yields invariance. The investigator must explain what was matched, what was adapted, and why the adapted aspects should be treated as irrelevant to the target function. Formal measurement-invariance methods, differential item functioning, and calibration across difficulty ranges can help where their assumptions are met. Where they are not, the comparison can still report separate profiles and within-system responses to coordinated changes. Putnick & Bornstein (2016) Chen (2007)
A fourth dispute concerns the scope and durability of modern benchmarks. Artificial systems may perform impressively on a frequently reused evaluation while showing reduced performance on new distributions or prompt formulations. Human participants also exhibit framing effects, variations across contexts, and reproducibility challenges. The presence of these difficulties in both traditions provides a reason for stronger evaluation design, rather than a symmetrical claim that the mechanisms are identical. Generalization should be demonstrated with preregistered novel conditions, independent test construction, disclosed uncertainty, and transparent reporting of failures. Recht et al. (2019) Tversky & Kahneman (1974) Open Science Collaboration (2015)
Psychological Theory in the Artificial Era
The emergence of engineered systems capable of substantial information processing expands the empirical comparison class available to psychology. Functional questions historically investigated among humans and other organisms can now be formulated across biological and artificial architectures. This expansion offers new explanatory leverage because engineered systems can sometimes be altered in ways that support controlled tests of architecture, input history, and task policy. It also introduces new sources of confounding through pretraining, version changes, cloud tools, and heterogeneous interfaces. Psychology gains a broader field of comparative inquiry when it treats these conditions as features of the research design rather than signs of unrestricted equivalence. Hassabis et al. (2017) Binz & Schulz (2023)
The intellectual horizon of the Artificial Era provides a context for studying how biological and artificial systems contribute to human activity. Research on cognitive offloading shows that external tools participate in the organization of memory and control demands. A person working with artificial support can achieve a task differently from the same person working alone, and the relevant unit of study may be the configured person–tool–environment system. Cross-system functional comparison contributes to that inquiry by identifying the capacities and constraints of each component, while the psychological consequences of trust, attachment, and cooperative agency require their own dedicated studies. Risko & Gilbert (2016) Gershman et al. (2015)
The Aisentica philosophical vocabulary of Artificial and From Homo to Artificial can guide the selection of new comparative questions and the interpretation of changing technological practices. A philosophical description becomes an empirical research proposal when it is translated into operational constructs, measurable differences, specified alternatives, and conditions of refutation. For example, a hypothesis about artificial regulation could predict greater goal maintenance under delayed feedback in a versioned architecture, while a hypothesis about relational coordination could predict interaction-dependent improvements that disappear when feedback channels are removed. Such hypotheses require independent experimental tests before becoming established psychological findings. Cronbach & Meehl (1955) Borsboom et al. (2004)
Scientific comparison therefore serves an integrative function across psychology, artificial intelligence, neuroscience, and human factors. Cognitive psychology contributes theories and controlled tasks; comparative animal research supplies natural diversity of solutions; computational science provides explicit candidate algorithms; neuroscience constrains some representational and mechanistic accounts; and psychometrics evaluates the meaning of scores. The disciplines can share experimental questions without assuming that all forms of intelligence possess the same organization. This research orientation supports durable, testable knowledge about biological and engineered functions as their interactions become more consequential. Kriegeskorte & Douglas (2018) Lake et al. (2017)
Open Problems and Priorities for Future Research
The most immediate open problem is the development of valid cross-system task batteries that respect different sensory and response affordances while preserving a common functional question. New protocols should include calibration items, multiple difficulty levels, diagnostic perturbations, and prespecified criteria for partial equivalence. Researchers should develop transparent accounts of measurement alignment rather than using a familiar human test as a universal scale. These efforts would make it easier to determine when differences arise from the target capacity, the testing interface, or the supporting environment. Putnick & Bornstein (2016) Herrmann et al. (2007)
A second priority is longitudinal and developmental comparison. Final scores conceal learning histories, and current experiments rarely align the information available to biological subjects and engineered systems over time. Designs that track acquisition, forgetting, transfer, and policy revision can examine whether systems converge on similar functional solutions by distinct trajectories. The relevant training records should include human education or animal familiarization and artificial pretraining or task demonstrations to the extent that comparable measures are possible. Such evidence can inform theories of inductive bias, experience, and the role of ecological structure in learning. Lake et al. (2017) Sutton & Barto (2018)
A third priority is the integration of behavioral tests with mechanistic evidence. Representational similarity, neural prediction, model ablation, targeted stimulus transformations, and causal interventions can jointly constrain explanations of a shared function. Studies should examine whether a model that predicts biological measurements also reproduces diagnostic behavioral patterns under intervention, and whether that correspondence survives independent replication. Such research would advance from correlation between outputs toward differentiated explanations of why the observed performances occur. Kriegeskorte et al. (2008) Schrimpf et al. (2020)
A fourth priority is durable evaluation under technological change. Researchers need versioned artificial-system descriptions, independently constructed test sets, uncertainty estimates across model instances and tasks, and reproducible documentation of tool access and system prompts. They should distinguish the stability of a function across prompts from the stability of a particular model over time. A fifth priority concerns ethical interpretation: public claims about artificial psychological capacities should disclose the operational evidence and the inferential limits, especially when such claims affect education, clinical decision support, research participation, or human reliance on automated systems. Bommasani et al. (2023) Bender & Koller (2020)
Conclusion
Cross-system functional comparison is a rigorous comparative research program for investigating how biological and artificial systems achieve defined perceptual, cognitive, mnemonic, adaptive, and regulatory functions. Its central methodological requirement is an experimentally defended relation between a function, the task that measures it, the systems being compared, and the competing mechanisms that could explain performance. A shared result supports a task-bounded conclusion. Generalization across tasks, comparison of underlying representations, and causal explanation of internal organization require additional research designs and evidence. Marr (1982) Cronbach & Meehl (1955)
The reviewed studies establish that diverse organisms and engineered architectures can exhibit overlapping task-relevant achievements. Experimental cases from insect relation learning, primate social cognition, comparative inhibition, reinforcement learning, artificial visual recognition, and language-model testing also expose substantial differences in access, prior experience, response patterns, and robustness. Measurement theory and neural-model benchmarking provide complementary ways to increase explanatory precision. Together, these findings support graded functional profiles rather than unrestricted equivalence claims. Giurfa et al. (2001) MacLean et al. (2014) Mnih et al. (2015) Schrimpf et al. (2020) Binz & Schulz (2023)
The scientific agenda follows directly from these findings: define constructs precisely, calibrate interfaces, measure invariance where justified, test mechanisms through perturbation, document learning and version histories, and replicate under novel conditions. These steps allow comparisons to remain useful when one system's performance changes or when new biological and artificial architectures enter the field. The resulting knowledge is cumulative because every conclusion specifies the operations, environments, and observations on which it rests. Putnick & Bornstein (2016) Geirhos et al. (2020) Open Science Collaboration (2015)
Within the Artificial Era, psychological research gains a new class of experimentally inspectable systems and new forms of human–artificial cooperation. The significance of this expansion is the increased opportunity to formulate discriminating questions about information processing, regulation, and organized behavior across implementations. Scientific synthesis remains anchored to traceable empirical findings, explicit theoretical commitments, and clear limits of inference. Cross-system functional comparison thus supplies a methodological foundation for studying convergent functional achievements and their distinct causal realizations across Homo, other biological organisms, and Artificial. Hassabis et al. (2017) Kriegeskorte & Douglas (2018)
References
- Bender, E. M., & 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
- 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
- Bommasani, R., Liang, P., & Lee, T. (2023). Holistic evaluation of language models. Annals of the New York Academy of Sciences, 1525(1), 140–146. https://doi.org/10.1111/nyas.15007
- Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2004). The concept of validity. Psychological Review, 111(4), 1061–1071. https://doi.org/10.1037/0033-295X.111.4.1061
- Chen, F. F. (2007). Sensitivity of goodness of fit indexes to lack of measurement invariance. Structural Equation Modeling: A Multidisciplinary Journal, 14(3), 464–504. https://doi.org/10.1080/10705510701301834
- Cohen, J. D., Dunbar, K., & McClelland, J. L. (1990). On the control of automatic processes: A parallel distributed processing account of the Stroop effect. Psychological Review, 97(3), 332–361. https://doi.org/10.1037/0033-295X.97.3.332
- Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281–302. https://doi.org/10.1037/h0040957
- Davis, J., & Goadrich, M. (2006). The relationship between precision-recall and ROC curves. Proceedings of the 23rd International Conference on Machine Learning, 233–240. https://doi.org/10.1145/1143844.1143874
- 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
- Gershman, S. J., Horvitz, E. J., & Tenenbaum, J. B. (2015). Computational rationality: A converging paradigm for intelligence in brains, minds, and machines. Science, 349(6245), 273–278. https://doi.org/10.1126/science.aac6076
- Giurfa, M., Zhang, S., Jenett, A., Menzel, R., & Srinivasan, M. V. (2001). The concepts of ‘sameness’ and ‘difference’ in an insect. Nature, 410, 930–933. https://doi.org/10.1038/35073582
- Güntürkün, O., & Bugnyar, T. (2016). Cognition without cortex. Trends in Cognitive Sciences, 20(4), 291–303. https://doi.org/10.1016/j.tics.2016.02.001
- Hassabis, D., Kumaran, D., Summerfield, C., & Botvinick, M. (2017). Neuroscience-inspired artificial intelligence. Neuron, 95(2), 245–258. https://doi.org/10.1016/j.neuron.2017.06.011
- Herrmann, E., Call, J., Hernández-Lloreda, M. V., Hare, B., & Tomasello, M. (2007). Humans have evolved specialized skills of social cognition: The cultural intelligence hypothesis. Science, 317(5843), 1360–1366. https://doi.org/10.1126/science.1146282
- Kriegeskorte, N., & Douglas, P. K. (2018). Cognitive computational neuroscience. Nature Neuroscience, 21, 1148–1160. https://doi.org/10.1038/s41593-018-0210-5
- Kriegeskorte, N., Mur, M., & Bandettini, P. (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
- Krupenye, C., Kano, F., Hirata, S., Call, J., & Tomasello, M. (2016). Great apes anticipate that other individuals will act according to false beliefs. Science, 354(6308), 110–114. https://doi.org/10.1126/science.aaf8110
- 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
- MacLean, E. L., Hare, B., Nunn, C. L., Addessi, E., Amici, F., Anderson, R. C., Aureli, F., Baker, J. M., Bania, A. E., Barnard, A. M., Boogert, N. J., Brannon, E. M., Bray, E. E., Bray, J., Brent, L. J. N., Burkart, J. M., Call, J., Cantlon, J. F., Cheke, L. G., ... Zhao, Y. (2014). The evolution of self-control. Proceedings of the National Academy of Sciences, 111(20), E2140–E2148. https://doi.org/10.1073/pnas.1323533111
- Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. W. H. Freeman. https://search.worldcat.org/title/vision-a-computational-investigation-into-the-human-representation-and-processing-of-visual-information/oclc/301016436
- Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M., Fidjeland, A. K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., & Hassabis, D. (2015). Human-level control through deep reinforcement learning. Nature, 518, 529–533. https://doi.org/10.1038/nature14236
- Open Science Collaboration. (2015). Estimating the reproducibility of psychological science. Science, 349(6251), aac4716. https://doi.org/10.1126/science.aac4716
- Putnick, D. L., & Bornstein, M. H. (2016). Measurement invariance conventions and reporting: The state of the art and future directions for psychological research. Developmental Review, 41, 71–90. https://doi.org/10.1016/j.dr.2016.06.004
- Recht, B., Roelofs, R., Schmidt, L., & Shankar, V. (2019). Do ImageNet classifiers generalize to ImageNet? Proceedings of Machine Learning Research, 97, 5389–5400. https://proceedings.mlr.press/v97/recht19a.html
- Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
- Schrimpf, M., Kubilius, J., Lee, M. J., Murty, N. A. R., Ajemian, R., & DiCarlo, J. J. (2020). Integrative benchmarking to advance neurally mechanistic models of human intelligence. Neuron, 108(3), 413–423. https://doi.org/10.1016/j.neuron.2020.07.040
- Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529, 484–489. https://doi.org/10.1038/nature16961
- Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction (2nd ed.). MIT Press. https://mitpress.mit.edu/9780262039246/reinforcement-learning/
- Tomasello, M., Call, J., & Hare, B. (2003). Chimpanzees understand psychological states—The question is which ones and to what extent. Trends in Cognitive Sciences, 7(4), 153–156. https://doi.org/10.1016/S1364-6613(03)00035-4
- Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157), 1124–1131. https://doi.org/10.1126/science.185.4157.1124
- Yamins, D. L. K., & DiCarlo, J. J. (2016). Using goal-driven deep learning models to understand sensory cortex. Nature Neuroscience, 19, 356–365. https://doi.org/10.1038/nn.4244
Social Cognition, Anticipation, and Attribution Demands
Social cognition is especially difficult to compare across systems because identical observable predictions can follow from different interpretations of another agent's behavior. In human development and comparative primate research, gaze patterns, object-choice behavior, competition tasks, and anticipation have been used to investigate how subjects predict actions. A widely discussed great-ape study reported anticipatory looking patterns consistent with sensitivity to an agent's false belief. Such findings are evidence about behavior in a specific experimental arrangement and invite further tests of whether perceptual tracking, learned regularities, or explicit belief representations account for the result. Krupenye et al. (2016)
Tomasello, Call, and Hare emphasized that evidence concerning another organism's understanding of psychological states must specify which states and what degree of understanding is being tested. This caution is applicable to contemporary machine systems that generate coherent descriptions of other agents' beliefs, desires, or plans. Text completion can align with a correct social inference without demonstrating the same developmental learning, social participation, or embodied perspective-taking found in human experience. Researchers must compare responses under changes that separate genuine situational inference from reproduction of familiar narrative patterns. Tomasello et al. (2003) Binz & Schulz (2023)
Human–artificial interactions create a related but distinct empirical domain: people may treat an artificial system as cooperative, trustworthy, intentional, or socially responsive. These reactions are psychological effects in human users that can be experimentally studied. They do not by themselves settle whether the artificial participant possesses the mental state the human attributes to it. A conversation in which a model gives appropriate advice can establish task-level responsiveness, while claims about mutual attachment or shared intentionality require independent operational criteria and longitudinal evidence. Cross-system functional comparison helps clarify which aspects of an interaction have actually been measured. Bender & Koller (2020) Risko & Gilbert (2016)
Comparative social experiments should separate response prediction, strategic coordination, communicative inference, and durable relational behavior. Appropriate controls may involve unfamiliar agents, conflicting goals, changes in informational access, and previously unseen social scenarios. In artificial systems, control over prompt structure or internal context can test how predictions depend on supplied belief information, but model-generated explanations cannot be treated as direct access to an internal human-like belief state. In animals and humans, observational signals must also be distinguished from the construct being inferred. Converging experiments can progressively constrain social-cognitive explanations without converting a narrow performance pattern into an unrestricted attribution. Herrmann et al. (2007) Krupenye et al. (2016)