RESEARCH DOSSIER
Species-Appropriate Comparative Cognitive Testing
Abstract. Species-appropriate comparative cognitive testing evaluates how well experimental assays identify psychological processes across organisms that differ in ecology, sensory access, motor affordances, motivation, rearing history, and learning opportunity. This fundamental research monograph reconstructs the historical emergence of the comparative assay problem, defines the relation among task performance, behavioral competence, construct validity, and evolutionary inference, and critically examines evidence from inhibitory-control tests, multisite primate memory research, ape social-cognitive experiments, and corvid physical-cognition tasks. It integrates experimental controls, repeatability and measurement-invariance analysis, participation and sampling biases, phylogenetic comparative methods, hierarchical statistics, and replication standards. Particular attention is given to the evolutionary self-control comparison of MacLean and colleagues, the ManyPrimates collaboration, research on motivation and selection into testing, and the 2026 second-order meta-analysis of statistical fragility in animal cognition. The synthesis finds that physically identical tasks can have unequal psychological requirements, whereas calibrated species-specific interfaces can support narrowly defined functional comparisons when construct-preserving manipulations are independently validated. It advances an evidence-based sequence for developing, testing, and reporting cross-taxon assays and delineates open problems in ecological fitness validation, latent comparability, and extensions to artificial cognitive systems. The work is a critical narrative research synthesis rather than an original experiment, registered systematic review, or new meta-analysis.
Keywords: comparative cognition; species-appropriate testing; comparative psychology; construct validity; ecological validity; measurement invariance; sensory calibration; motor affordances; motivation; animal cognition; inhibitory control; cognitive testing; ManyPrimates; phylogenetic methods; replication; repeatability; sampling bias; animal welfare; Artificial Era
The Scientific Object of Species-Appropriate Cognitive Testing
Species-appropriate comparative cognitive testing is the design, calibration, administration, and interpretation of behavioral assays that permit justified comparisons of psychological operations across organisms whose sensory worlds, motor repertoires, motivational systems, learning histories, and ecological problems differ. The scientific object is the inferential link between a task-defined performance and a specified cognitive construct. A valid test identifies the operations required to perceive the relevant events, retain information, select a response, and act on the environment, then demonstrates that alternative pathways to success or failure are sufficiently constrained. Species appropriateness is an empirical property of a design relative to particular populations and hypotheses; it must be investigated rather than presumed from a familiar apparatus Krasheninnikova et al. (2020) Rowe & Healy (2014).
The field addresses an apparent paradox of comparative measurement. A physically identical apparatus can impose unlike demands on different animals, while physically different apparatuses can instantiate closely related informational problems. An opaque tube followed by a transparent tube, for instance, may elicit reaching, pecking, trunk manipulation, or a locomotor detour. Each response places different demands on attention and motor coordination. Treating the resulting proportion of correct trials as a direct scale of general intelligence introduces construct-irrelevant variance unless the sensory and motor prerequisites are calibrated. Conversely, modifying the apparatus may increase ecological accessibility but alter the decision problem. The methodological task is to articulate and test which features must be invariant for comparison MacLean et al. (2014) Beran (2015).
This monograph uses a critical narrative and methodological synthesis of original comparative experiments, multisite collaborations, studies of sampling and repeatability, phylogenetic methods, and recent evidence reviews. It distinguishes direct task outcomes, evidence for a latent psychological capacity, evolutionary interpretations, and broader claims about species-level cognition. The scientific questions are whether tasks assess analogous operations, which ecological modifications preserve construct identity, how selection into testing changes inference, what replication establishes, and how cross-species data can support causal and phylogenetic hypotheses. Its contributions are an integrated analysis and a set of testable design criteria grounded in the cited literature, not an original experiment, registered systematic review, or new meta-analysis Farrar et al. (2021) Boogert et al. (2018) Sol et al. (2025).
Historical Formation of the Comparative Assay Problem
The comparative investigation of animal minds emerged from natural history and the study of learning in diverse environments. As experiments became central to psychology, an animal's response to an artificial problem became a tractable observation, but laboratory success acquired interpretations that extended beyond what the immediate task measured. The continuity hypothesis associated with Darwin motivated comparisons across living lineages while also raising enduring questions about anthropomorphic descriptions and the relationship between a human-centered category and a different organism's behavior. Modern comparative cognition retained experimental control while developing finer distinctions among learning, perception, attention, memory, inhibitory processes, and physical and social reasoning Shettleworth (2012) Tinbergen (1963).
Tinbergen's 1963 account of ethological explanation supplies a durable methodological discipline: the proximate mechanisms, development, functional consequences, and evolutionary history of a behavior are distinct questions with different evidentiary demands. A demonstration that an animal succeeds in retrieving food after a delay is evidence about performance under given conditions. Identifying the representational and motivational mechanisms requires interventions. Establishing how experience shapes the behavior requires developmental study, whereas phylogenetic reconstruction requires comparisons across related species and explicit historical models. Collapsing these explanatory levels makes both positive and negative cognitive claims stronger than their evidence warrants Tinbergen (1963) Felsenstein (1985).
Modern comparative work increasingly exposed the difficulties of transporting paradigms among taxa. Cognitive batteries used with apes, corvids, parrots, dogs, rodents, and other species created opportunities for convergent questions but also juxtaposed apparatus-specific constraints. The large comparative inhibition project of MacLean and colleagues offered a landmark in collaborative standardized testing; subsequent scrutiny of the cylinder task, its constructs, and site effects illustrated why standardization is a necessary research accomplishment rather than a conclusive validity guarantee. Large collaborations such as ManyPrimates formalized protocol sharing, multisite participation, preregistration, and cross-institutional data aggregation as responses to earlier limitations MacLean et al. (2014) Beran (2015) ManyPrimates (2019).
From Behavioral Performance to Cognitive Construct
Comparative assays require an explicit distinction among stimulus, behavior, score, and psychological explanation. Stimulus properties include accessible spectral and temporal characteristics, reward visibility, spatial arrangement, and environmental cues. Behavior is the organism's observable orientation, choice, latency, or action. The score is a defined summary of the observations, such as the first response, learning rate, success probability, or trial-to-trial adjustment. The cognitive construct is a theoretical account that relates variations in these observations to operations such as memory retention or interference control. Validity depends on the pattern of predictions across interventions and contexts rather than on the evocative name assigned to a task Rowe & Healy (2014) Boogert et al. (2018).
A single behavioral result generally admits multiple explanations. Apparent preference for an informative cue may reflect attentional salience, reinforcement history, spatial habits, or familiarity. Failure to transfer a rule after training may reflect perceptual confusion about a new display, an unlearned motor action, competing reward expectations, or reduced motivation. The appropriate response is to derive discriminating predictions, manipulate the relevant feature, and measure preconditions independently. For example, transparent-barrier tasks should separately establish that an individual perceives the barrier, can execute the detour, understands how the food can be reached, and remains sufficiently engaged across trials. Only then can a test phase meaningfully probe behavioral inhibition MacLean et al. (2014) Shaw (2017) van Horik & Madden (2016).
Construct alignment across species is a research hypothesis about functional organization. A delayed-choice task may assess location memory in different taxa if they share the relevant discrimination, encoding, retention, and response demands; equivalent delays in seconds do not necessarily create equivalent memory difficulty for organisms differing in activity rhythm and ecological time horizon. More robust interpretations use several conceptually connected tasks with different surface demands, repeated observations, and transparent reporting of discordant outcomes. Convergence among tasks strengthens a construct interpretation, while divergent results can reveal heterogeneous capacities or confounds. Reliability establishes consistency of measurement but does not, on its own, prove that the intended capacity is being measured Cauchoix et al. (2018) ManyPrimates (2019).
Ecological Validity, Functional Relevance, and Test Transport
Ecological validity concerns the extent to which the relations investigated by an assay correspond to meaningful information, actions, and consequences in the organism's ordinary life. A species that routinely retrieves cached food may display highly developed spatial routines, whereas another depends on tracking visual motion, olfactory information, or social cues. Translating these capacities into laboratory tests requires more than borrowing a human psychological label. The experimenter must specify the organism's natural perceptual opportunities, reward economy, relevant delays, and feasible actions. Field observation supplies candidate operations and meaningful outcome variables, while controlled experiments identify mechanisms and test alternatives Sol et al. (2025) Shaw (2017).
Ecological calibration should avoid the assumption that a familiar or naturalistic setting automatically improves measurement. A wild foraging problem may embed many uncontrolled opportunities for learning, attention allocation, and environmental variation. An artificial task can isolate a causal process more effectively if its prerequisites are demonstrated and the organism can interact with it readily. Conversely, successful acquisition of a novel laboratory operation can reveal flexibility not visible in ecological observation. The strongest design justifies the bridge between ecological competence and experimentally controlled information demands, often by contrasting a naturalistic assay, a simplified analogue, and an interference condition that separates target and alternative mechanisms Rowe & Healy (2014) Boogert et al. (2018).
Transporting a task to a distant lineage demands documented decisions about stimulus equivalence and apparatus modifications. Such modifications should be classified as changing delivery modality, response effector, incentive, available cues, or cognitive structure. A change from hand reaching to beak manipulation may preserve a detour rule but introduce different error opportunities; a change from visual to olfactory stimuli can preserve a discrimination relation while transforming attentional demands. Cross-species inference therefore benefits from both within-species validation and bridges across closely related taxa before broad phylogenetic claims. Krasheninnikova, Chow, and von Bayern specifically advocate comparisons at several taxonomic levels to address perceptual, motivational, morphological, and standardization problems Krasheninnikova et al. (2020) Felsenstein (1985).
An Evidence Hierarchy for Comparative Cognitive Claims
Scientific claims form a sequence from observed performance, through task competence, to construct-level inference and historical evolutionary explanation. A subject that repeatedly chooses a baited location after a delay has demonstrated a reproducible operation in an environment defined by the apparatus and training. To infer memory maintenance, investigators must address perceptual access to baiting, unintended cues, spatial strategies, reward preferences, and the distribution of responses across delays. To infer species-level characteristics, data must represent within-species variation, not simply the subset willing to participate. To infer an evolutionary cause, comparisons must then account for shared ancestry and competing ecological predictors ManyPrimates (2019) Farrar et al. (2021) Felsenstein (1985).
This hierarchy prevents category mistakes in both directions. Unexpected success by a bird or invertebrate can invalidate a narrow claim of exclusivity, yet it does not automatically demonstrate that the neural or representational processes are the same as in a primate. Unexpected failure can identify a limit of performance under tested conditions, yet it may remain unclear whether the limit resides in the target capacity or in task access. Howard and Barron argue that research on cognitive limits requires carefully interpreting negative results, addressing practical task limitations, and countering selective publication of positive results. Such reasoning gives null outcomes an explicit scientific role Howard & Barron (2024) Jelbert et al. (2014).
A strong evidence review records the exact unit of analysis for each inference: trial, individual, population, facility, species, or phylogenetic branch. Reporting a species mean from a single small captive sample does not justify speaking for an entire lineage. A comprehensive evaluation links each result to sampling frame, operational definition, primary and secondary outcomes, observation conditions, analytic uncertainty, and independence of replication. The evidence standard is not simply methodological uniformity; it is the ability to explain why a difference in measured performance is attributable to a cognitive operation and which alternatives remain plausible Farrar et al. (2021) Mizuno et al. (2026).
Sensory Access, Salience, and the Animal's Perceptual World
An assay begins before the first recorded choice: the organism must detect the relevant information. Color, contrast, spectral sensitivity, viewing angle, flicker, odor dispersal, acoustic frequency, and spatial resolution vary across taxa. The accessible stimulus must be described in the functional terms of the animal rather than solely by human observation. An object that looks conspicuous to an experimenter may fall outside the tested animal's effective perceptual contrast, while a scent trail invisible to the experimenter may resolve an ostensibly visual discrimination. The proper design includes sensory pretests and controls for redundant or inadvertent cues, with stimulus placement and illumination matched to the population's capacities Krasheninnikova et al. (2020) Rowe & Healy (2014).
Equivalence of discriminability is more informative than identity of physical stimulus. Where feasible, investigators should first estimate detection and discrimination functions, then compare animals under conditions that place them at comparable distances above their own perceptual thresholds. Such calibration helps distinguish a cognitive decision from the ability to register its inputs. It does not eliminate all modality differences because perceptual representations may differ in dynamics and ecological meaning. Within-species manipulations of intensity, duration, and cue availability can establish whether experimental success depends on a supposedly irrelevant perceptual advantage. Cross-modal adaptation should preserve the informational relation specified by the research hypothesis and document what additional pathways became available Krasheninnikova et al. (2020) Jelbert et al. (2014).
Perception also shapes social-cognitive assessment. Great apes had repeatedly failed some explicit false-belief tasks, whereas an anticipatory-looking paradigm revealed patterned eye movements indicating expectation about an agent's action under a false-belief scenario. Krupenye and colleagues' experiment did not simply make the same verbal problem easier; it measured a different response form and a more implicit level of processing. This case demonstrates the value and interpretive challenge of designing a behaviorally accessible assay. Complementary experiments, including subsequent use of self-experience in anticipatory-looking tests, are needed to specify precisely what information the animal used and which accounts remain viable Krupenye et al. (2016) Kano et al. (2019).
Motor Affordances, Apparatus Geometry, and Response Constraints
Motor execution is a component of task performance. Grip morphology, reach trajectory, beak shape, trunk dexterity, locomotor speed, posture, and an animal's willingness to contact unfamiliar objects can determine whether a nominally identical problem has the same response cost. In a transparent cylinder test the required movement is a learned detour, yet different species may encounter different physical obstacles or persistent response biases. Mechanical properties of the apparatus and accessibility of the reward must be evaluated separately from the purported inhibitory demand. A pretraining success criterion demonstrates that a maneuver is possible, but it cannot guarantee equivalent effort or response competition during the scored test MacLean et al. (2014) Beran (2015).
Motor confounds can be isolated through matched control conditions: an opaque version of the barrier, a transparent version with reduced direct reach temptation, a response modality suited to the species, and repeated demonstrations of successful reward acquisition. First-trial errors and improvement across trials convey different information. The former may reflect spontaneous impulse regulation or novelty; the latter may reflect adjustment to transparent material and discovery of a motor strategy. Latency alone is difficult to compare when animals have different movement speeds, approach styles, and reward assessment. A defensible model separates decision latency, execution latency, and unsuccessful contacts wherever these components can be reliably coded Shaw (2017) Farrar et al. (2021).
Comparative research on physical causality faces related demands. New Caledonian crows in the water-displacement paradigm made selective choices among some functionally different objects and tube conditions but failed more difficult versions of the test. The juxtaposition is scientifically productive: it identifies a graded pattern that supports restricted claims about sensitivity to causal properties while leaving open the mechanisms supporting successful choices. Apparatus-based success requires interpreting object manipulation, perceptual salience, associative history, and abstract physical representation together. Changing which properties are causally relevant across trials helps test whether performance transfers to novel physical relations Jelbert et al. (2014) Horner & Whiten (2005).
Reward Economy, Motivation, and Behavioral State
Food rewards make many assays practical, but their incentive value varies with satiation, diet, handling, circadian stage, metabolic needs, social context, stress, and opportunity to obtain food elsewhere. A task may become easier or harder because subjects are more willing to explore, persist, or tolerate delay rather than because a memory or inference mechanism changes. Experimental control includes calibration of reward preference and sufficient documentation of prior access to food. In field settings, body condition and seasonal resources require particular attention. Shaw found that body condition influenced performance in wild North Island robins' detour-reaching test, while experience affected a different motor task Shaw (2017) van Horik & Madden (2016).
Novel-problem performance is especially vulnerable to this ambiguity. Van Horik and Madden reported that motivational traits predicted success in novel operant foraging tasks more clearly than the cognitive attributes they sought to assess. The result raises a general interpretive issue: persistence can increase opportunities to discover a solution while creating an appearance of superior reasoning. Measuring the number of attempts, exploration time, contact frequency, and task-specific learning can identify this pathway. Researchers should report refusal, disengagement, and latency to approach in addition to correct solutions, and should avoid excluding reluctant animals without assessing whether that exclusion systematically distorts the target population van Horik & Madden (2016) van Horik et al. (2017).
Captivity and habitat may further shift motivational conditions. In a comparison of wild and captive gray mouse lemurs, Splinter, Fichtel, and Radespiel reported stronger performance among wild animals in two of three cognitive tasks and discussed differences in energy requirements and food motivation as plausible contributors. This does not establish a general cognitive superiority of wild over captive populations. It instead makes setting an active explanatory variable and highlights the need for matched rewards, individual histories, and naturalistic comparisons. A robust analysis records origin and current environment, tests interactions with task type, and resists converting a particular group difference into a universal hierarchy Splinter et al. (2025) Farrar et al. (2021).
Learning History, Rearing, Familiarity, and Strategy Formation
Experimental experience changes what an animal understands about objects, experimenters, barriers, rewards, and permissible actions. Zoo residents may complete dozens of cognitive studies, whereas wild individuals may first encounter a test apparatus during the focal experiment. Hand-reared animals may attend more readily to human social cues. Habituation and training provide access to task requirements but also introduce experience-dependent strategies. A valid cross-population comparison therefore requires an inventory of prior tests, reward histories, rearing conditions, socialization, and relevant human interaction. These factors should be measured as potential predictors rather than handled only through informal descriptions of animal background ManyPrimates (2019) van Horik et al. (2017).
Training criteria can themselves produce selection. Requiring animals to learn an opaque-cylinder detour before the transparent-cylinder test establishes a minimum motor operation, yet individuals unable or unwilling to complete the criterion are removed from the scored population. The resulting test sample is conditional on willingness, basic learning, and apparatus access. Comparing the successful subset across species without reporting training attrition conflates two processes: selection into measurement and the measured response after entry. A stronger approach retains training outcomes as data, models eligibility and nonparticipation, and distinguishes a task's feasibility in a species from the average performance among selected participants MacLean et al. (2014) Farrar et al. (2021).
Strategy may also change during the scored trials. An animal can learn to avoid a barrier, infer a relevant causal relation, exploit a subtle cue, or repeat a rewarded sequence without reproducing the theoretical mechanism imagined by the researcher. Probe trials that reverse stimulus–reward contingencies, remove predictive cues, or change task geometry can discriminate such pathways. Interpretation should attend to performance trajectories, not merely an aggregate success percentage. A first-trial preference, rapid acquisition, stable asymptotic performance, and successful transfer carry different implications about prior knowledge and learning Horner & Whiten (2005) Jelbert et al. (2014) Boogert et al. (2018).
Affective State, Welfare, and Voluntary Participation
Fear of novelty, arousal, social separation, fatigue, and stress alter approach behavior and exploration. These state variables may affect both measurement quality and animal welfare. Humane comparative testing requires opportunities to withdraw, transparent stopping criteria, enrichment-sensitive scheduling, and procedures tailored to each species' social and ecological requirements. Institutional animal ethics oversight provides the governing framework for research with live animals, while study-specific reporting should make clear how motivation and welfare were monitored. A subject that does not approach an apparatus has produced meaningful information about test accessibility, though this information should be interpreted without presuming an unobserved mental deficit van Horik et al. (2017) Farrar et al. (2021).
Voluntary participation creates a distinctive selection mechanism. Animals comfortable with a researcher, tolerant of novelty, or strongly motivated by the reward may appear disproportionately often in the dataset. When participation correlates with personality, sex, condition, or experience, the observed sample is not exchangeable with the full population. Van Horik and collaborators directly examined differential participation in cognitive testing and identified these predictors of engagement. Future designs can quantify participation probability, use repeated invitations, and report both the enrolled and the available populations. Statistical adjustment is possible only for recorded selection variables and depends on assumptions about unmeasured causes van Horik et al. (2017) Farrar et al. (2021).
Ethically sound apparatus design can improve measurement as well as welfare. Familiarization can reduce anxiety, choice-based engagement can make a task acceptable, and species-specific manipulanda can reduce accidental physical demands. Nevertheless, extensive familiarization may create new learning histories and change the construct under study. The investigator should explain these tradeoffs, retain treatment and training logs, and distinguish subject withdrawal from incorrect performance. A scientific dataset that treats inability to participate as zero ability commits a substantial measurement error; a dataset that silently discards all nonparticipants risks a different bias. Reporting both problems allows cumulative interpretation ManyPrimates (2019) Krasheninnikova et al. (2020).
Inhibitory Control as a Comparative Measurement Case
The 2014 cross-species study by MacLean and collaborators tested 567 individuals from 36 species with two tasks intended to measure self-control: the A-not-B task and the transparent-cylinder detour task. The study combined standardized behavioral procedures with phylogenetic comparative analysis and reported a relationship between absolute brain volume and performance across its tested taxa. Its scale made it a landmark demonstration that multispecies behavioral data could be assembled under coordinated protocols. Yet the label self-control contains several operations, and the tasks principally impose interference between a direct or previously rewarded response and an alternative response. Treating a composite score as a measure of all forms of self-control would exceed the tested operational scope MacLean et al. (2014) Beran (2015).
The A-not-B task involves suppressing a response previously associated with reward after the baited location is changed. The cylinder task requires an animal to use an indirect access route despite the visible food positioned beyond a transparent barrier. The tasks differ in memory demand, cue visibility, motor planning, and learning history. MacLean and colleagues found a positive cross-species association between performance on the tasks among species tested in both, supporting some convergent measurement under the study conditions. However, convergence does not establish a single general capacity without examining species-specific patterns, alternate task measures, and potentially influential outliers. The study itself discusses anomalous results for an elephant on the A-not-B task MacLean et al. (2014).
Later critiques make the distinction between behavioral inhibition and intertemporal choice central. Beran noted that choosing a larger delayed reward has often been grouped with response inhibition under the broad term self-control, despite being operationally distinct. A successful detour may reveal control over a prepotent motor response, whereas delaying gratification requires tradeoffs between rewards distributed over time. Combining these results in a common score demands theoretical justification and evidence of measurement structure. Proper comparative analysis should define the narrower construct, evaluate the multiple tasks' correlation, and report how inferences change when particular tasks or species are removed Beran (2015) Cauchoix et al. (2018).
Replication, Site Effects, and the Cylinder Task Revisited
Replication in comparative cognition concerns more than returning to the same nominal protocol. When the sampling site, cohort, apparatus supplier, experimenter, previous training, or animal's environment changes, the study tests how robust a result is to these variations. Farrar, Voudouris, and Clayton examined the representativeness problem through resampling analysis and a cylinder-task case study, using data from the large multispecies comparison and additional observations. Their analysis demonstrates why small idiosyncratic samples can yield unstable estimates and why between-site differences belong in the scientific inference. Site replication should therefore report both fidelity to the core task and divergences in contextual conditions Farrar et al. (2021).
An independent laboratory may produce a lower score for a species without discovering a biological difference between its population and the original study's population. The causes might include testing history, unfamiliar apparatus, group composition, reward delivery, or the proportion of animals that entered the test. The appropriate analysis distinguishes changes in the measurement instrument from changes in the sampled organisms and setting. Variation across repeated tests can be estimated using hierarchical models with individual and site effects. Whenever sample size permits, design-stage stratification across sites is preferable to explaining conflicting results after their appearance Farrar et al. (2021) ManyPrimates (2019).
Multi-task validation can further test whether a behavioral label survives changes in apparatus. For inhibitory processes, a detour task can be paired with reversals, inhibition of an established response, or reward-delay decisions, while anticipating that the components may only partly overlap. A low correlation may mean the tasks measure different processes, or that noise and context dominate their scores. A high correlation may arise from a shared noncognitive factor such as activity level or willingness to engage. Interpreting either pattern requires independent measurements of task compliance, repeated performance, stimulus discriminability, and motivational state Boogert et al. (2018) Cauchoix et al. (2018).
Short-Term Memory and the ManyPrimates Multisite Model
ManyPrimates created an infrastructure for collaborative comparative research across primate facilities and applied it in an initial short-term memory study. The 2019 pilot included 176 individuals from 12 primate species at 11 sites across several continents. Animals observed the location of a hidden reward and were tested after delays of 0, 15, or 30 seconds. The study reported better performance at shorter delays and explicitly described limits on the phylogenetic conclusions possible with its species sample. The distinctive contribution is methodological: a common protocol, shared data collection, reproducible analytic practices, and explicit recognition that testing conditions may differ across sites ManyPrimates (2019).
Delay intervals illustrate the relation between standardization and species appropriateness. Using the same number of seconds gives the researcher a common physical unit, but a delay can be affected differently by attention, post-baiting interference, natural behavioral rhythm, and prior task experience. The resulting temporal gradient may identify retention within a species, while mean performance differences between species additionally require controls for motivation, position bias, and movement. An informative comparative design samples multiple intervals and estimates the shape of each species' performance function rather than comparing only one time point. Inclusion of a near-zero baseline is essential for separating failures of encoding from changes during retention ManyPrimates (2019) Rowe & Healy (2014).
The multisite model also creates new statistical responsibilities. Observations are nested within animals, animals within groups or facilities, and species within phylogenetic relationships. A facility with many subjects may influence pooled estimates disproportionately; a species represented at only one site is confounded with that site's circumstances. Preregistration can define the primary outcome, exclusions, fixed-versus-variable protocol features, and the hierarchical model before examining results. Open data and analysis code enable independent checks but do not substitute for representative sampling. ManyPrimates demonstrates a framework for acquiring broader evidence while leaving the ecological and construct-validity work to specific research programs ManyPrimates (2019) Farrar et al. (2021).
Physical Cognition, Tool Use, and Causal Transfer
Physical-cognition tasks often ask whether animals select or transform objects in ways consistent with causal understanding. The water-displacement experiments of Jelbert and colleagues presented New Caledonian crows with variations in object density, water versus sand, and apparatus geometry. The birds selected functionally advantageous alternatives in several conditions and failed others, including more demanding arrangements. These dissociations allow a more detailed scientific description than simply ranking the animals as possessing or lacking a general notion of causality. Performance may combine learned contingencies, perceptual cue use, and sensitivity to relations between objects and outcomes Jelbert et al. (2014).
Transparency manipulations have also been used to distinguish imitation from emulation in social learning. In experiments by Horner and Whiten, chimpanzees and children encountered tool-related sequences under conditions that differed in the visibility of causal structure. Results from such paradigms constrain how observers reproduce actions and how available causal information modifies copying. Their validity depends on the physical demands of manipulation, similarity of prior tool experience, reward understanding, and the social relation between model and participant. Cross-species comparison must therefore establish that an apparently inefficient copied action is similarly visible and executable before interpreting differences in selective imitation Horner & Whiten (2005).
The most diagnostic physical tasks introduce transfer probes in which a previously rewarded superficial feature is dissociated from the property that actually generates the outcome. If an animal continues to select an effective object when its appearance changes, the observation constrains simple stimulus-association accounts. If it fails, the result still requires checking whether the transformation creates a novel manipulation demand. Good task batteries vary perceptual surface, causal relation, and action requirements independently where possible. The scientific aim is to identify which combinations of information support successful performance and which alternatives can be excluded empirically Jelbert et al. (2014) Horner & Whiten (2005) Howard & Barron (2024).
Sampling Frames, Participation Bias, and Species-Level Inference
The population implied by a comparative conclusion should be identified before recruitment. A claim about captive zoo individuals concerns a population that has passed through capture or breeding, husbandry, survival, habituation, and participation filters. A claim about a species additionally assumes that variation across geographical ranges, life stages, sexes, and social ecologies has been addressed. Many studies instead obtain a convenience sample of individuals at one facility. Such samples can yield valuable experimental evidence about mechanisms, but the precision of the task score does not compensate for the absence of representative coverage. Selection into testing must therefore be documented as a separate process from success after testing begins Farrar et al. (2021) ManyPrimates (2019).
Selection is often structured by an animal's inclination to engage with unfamiliar equipment. Van Horik and colleagues found that participation in cognitive tests varied with measured personality, sex, body condition, and experience, demonstrating that exclusion can be systematically related to traits of scientific interest. A comparison of successful participants alone estimates conditional performance, and differences in training attrition can produce apparent species contrasts. Investigators should report the population available at the site, the individuals invited, initial participation, habituation completion, training eligibility, scored trials, withdrawals, and reasons for exclusion. Where feasible, repeating invitation opportunities and studying nonparticipants provides data about accessibility rather than silently turning selection into an unobserved cause van Horik et al. (2017) Boogert et al. (2018).
Sampling design should distribute observations among individuals, populations, sites, and taxa according to the target inference. Increasing repeated trials improves within-individual precision, but twenty trials from one individual do not supply twenty independent animals. Increasing individuals at one site does not supply a new ecological context; increasing species without balancing relatedness may still leave an evolutionary comparison narrow. The ManyPrimates pilot demonstrated feasible international collaboration while recognizing limits from species imbalance and the relatively small taxonomic sample. Design reports should separate trial counts, individual sample sizes, site coverage, species coverage, and effective phylogenetic contrasts. Hierarchical models can acknowledge clustering, though they cannot recover unsampled habitats or missing forms of variation automatically ManyPrimates (2019) Felsenstein (1985).
Repeatability, Reliability, and Latent Cognitive Structure
A task score must be sufficiently stable for the intended comparison while remaining sensitive to processes that genuinely change. Test–retest reliability examines consistency across occasions under specified conditions; interobserver reliability concerns the agreement of behavioral coding; internal structure concerns the relations among items and subtasks. In animal cognition, genuine learning, motivational variation, and ecological context may influence what a repeated assessment means. Repeatability is consequently a property of a population, procedure, time interval, and analytic model rather than a permanent characteristic of the animal. Researchers should partition within-individual variability from between-individual differences and should report uncertainty around estimates Cauchoix et al. (2018) Boogert et al. (2018).
Cauchoix and colleagues conducted a meta-analysis of the repeatability of individual cognitive performance across comparative research. The synthesis found repeatable individual differences overall, but estimates differed by cognitive domain, testing conditions, and analysis; the evidence did not warrant treating all cognition scores as interchangeable markers of a unitary trait. A stable score can reflect consistent exploration, sensory preference, or apparatus familiarity alongside the target ability. Likewise, a noisy score can obscure meaningful individual variation. Reliability should therefore be triangulated with controls for motivation and motor competence, and with assays that access the same putative operation through different surfaces Cauchoix et al. (2018) van Horik & Madden (2016).
An empirically defensible latent construct can be investigated through multitrait–multimethod reasoning. If several tasks designed to measure inhibitory control correlate after accounting for persistence, sensory access, and site, that pattern supports a common source of variance; if they do not, the construct may be heterogeneous or the measures unreliable. Across species, measurement invariance must be evaluated at the level supported by the data: whether task parameters, response functions, and latent–observed relationships are sufficiently similar for the intended comparison. Full equality of raw scores is rarely justified merely by shared apparatus. Partial or bounded invariance may support within-species contrasts and cautiously specified between-species differences Rowe & Healy (2014) Krasheninnikova et al. (2020).
Phylogenetic Nonindependence and Evolutionary Claims
Species comparisons confront a historical dependency: related species share aspects of their ancestry and cannot be treated as independent observations drawn from unrelated evolutionary experiments. Felsenstein's independent-contrasts framework made that dependency explicit. Modern studies may use phylogenetic generalized least squares, tree-structured random effects, or related comparative models, but every method inherits uncertainty about topology, branch lengths, trait measurement, and the assumed mode of evolution. Statistical correction cannot establish that a score validly measures the same psychological construct across taxa. Construct alignment is a prerequisite of phylogenetic estimation, not a product of the resulting significance test Felsenstein (1985) MacLean et al. (2014).
Cross-species associations between cognition and ecological predictors can have multiple causal explanations. A relation between performance and brain size, for example, can reflect correlated life history, motor demands, social opportunities, energy budgets, sampling structure, or the choice of tasks; an estimated phylogenetic relationship is not automatically a historical adaptation. MacLean and colleagues found a relationship between performance on two inhibitory-control tasks and absolute brain volume among their tested species, providing a tractable comparative hypothesis. Interpreting it requires sensitivity analyses for the specific task, taxon coverage, shared ancestry, body-size covariates, and selection into participation, together with independent tests in ecologically relevant settings MacLean et al. (2014) Farrar et al. (2021).
Evolutionary explanation ultimately asks how a functional trait varies, is inherited or developmentally transmitted, and contributes to outcomes related to survival and reproduction in appropriate environments. Sol and colleagues examined why direct evidence linking cognitive traits to fitness remains incomplete despite the intuitive appeal of adaptive cognition. They identify the need for well-characterized traits, heritable variation, long-term observations, and explicit ecological selection pressures. An individual that solves a laboratory puzzle rapidly has demonstrated a task result; interpreting the result as a fitness-enhancing evolutionary adaptation requires a further evidentiary program. Comparative testing supports that program when the measured process has an established connection to the organism's ecological decision problems Sol et al. (2025) Tinbergen (1963).
Design-Based Identification, Hierarchical Estimation, and Uncertainty
An experiment should state the target estimand: for example, the change in successful responses caused by increased delay within a defined population, or the difference in interference cost between two conditions after calibration of perceptual and motor prerequisites. Trial-level outcomes can be modeled with binomial or other response-appropriate likelihoods while including repeated observations for individuals and clustering by site. Learning across trials requires time or trial-order terms. When species are compared, evolutionary dependency may require a separate phylogenetic structure. The number of free model parameters must be justified by actual independent units and their variability, not by an inflated count of repeated trials Farrar et al. (2021) ManyPrimates (2019).
Task manipulations gain explanatory power when factors can be crossed. A delay-memory assay can manipulate delay and intervening distraction while checking a zero-delay encoding baseline. A physical-cognition assay can orthogonally vary object surface properties and their causal efficacy. An inhibition assay can separate previous-response learning from the immediate temptation of a visible reward. These structures allow competing mechanisms to predict different response patterns; they are preferable to asserting a capacity from above-chance performance in one condition. Randomization, counterbalancing, blinded video coding where feasible, preregistered exclusions, and analysis of first-trial versus acquired performance reduce opportunities for accidental cueing and post hoc interpretation Jelbert et al. (2014) MacLean et al. (2014) Horner & Whiten (2005).
Scientific reporting should include effect estimates and confidence or credible intervals, performance distributions, response trajectories, missingness, training attrition, and the number of independent species and facilities. The distinction between exploratory modeling and prespecified hypothesis tests should be explicit. Outlier analyses must acknowledge that removing a surprising species may remove biologically meaningful evidence; nevertheless, a large phylogenetic association should be tested for sensitivity to individual taxa. Study conclusions should specify the observed population and tested task. A broad claim about animal cognition can be justified only by evidence accumulated across comparable designs, independent laboratories, and appropriately diverse taxa Farrar et al. (2021) Mizuno et al. (2026).
Evidence Synthesis, Statistical Fragility, and Publication Bias
Evidence synthesis is essential because a single cognitive assay rarely resolves a broad theoretical question. Reviews should separate original empirical designs, replication attempts, meta-analytic summaries, and conceptual accounts. Effect sizes must be interpreted with regard to the task structure, outcome definition, species, laboratory conditions, and independence of samples. Pooling associations under one label can conceal differences between detour-reaching, delay tolerance, and learned reversal, even when all are described as self-control. A structured evidence map can classify which construct each experiment tested, which alternatives its controls addressed, and which populations remain absent Boogert et al. (2018) Cauchoix et al. (2018).
A recent empirical meta-research assessment makes the need for careful synthesis especially clear. Mizuno, Pollo, Lagisz, Guillette, Soma, and Nakagawa published a preregistered second-order meta-analysis in September 2026 covering 28 published meta-analyses, 48 eligible models, and 5,740 effect-size observations. Their estimated primary-study statistical power depended on the assumed underlying effect and ranged from 9.0% to 17.4% across three empirical reference specifications. They also examined magnitude and sign error. Because the underlying evidence was concentrated in rodent studies of learning and memory, the results diagnose weaknesses in the analyzed literature and cannot be treated as a precise universal estimate for every comparative species or cognitive domain Mizuno et al. (2026).
Selective publication can inflate the apparent support for striking interspecies contrasts and underrepresent conditions in which an animal fails to show the predicted capacity. Howard and Barron emphasize that negative evidence deserves carefully designed tests and interpretation. Preregistration of primary outcomes, archived protocols, registered reports, well-described null results, data and code preservation, and prospective multicenter replication can reduce avoidable distortions. These reforms should be paired with construct review; transparency about a poorly aligned assay cannot transform its score into a valid comparative measure. An evidence synthesis should therefore state both the statistical uncertainty of results and the conceptual uncertainty about what the procedures actually measure Howard & Barron (2024) Mizuno et al. (2026) ManyPrimates (2019).
A Calibrated Comparative Assay: Design and Validation Sequence
Species-appropriate measurement can be organized as a sequence of distinct validity decisions. Investigators first define the construct and the predicted pattern of behavior without naming any specific apparatus. They next identify ecological circumstances in which the operation matters and describe the perceptual inputs and actions required to express it. A pilot establishes whether each population can discriminate the stimuli, execute the response, recognize the reward contingency, and tolerate the testing situation. The substantive experiment then manipulates the hypothesized cognitive relation while holding validated prerequisites stable. Transfer probes determine whether the acquired solution generalizes when surface cues or action demands change. The report documents each stage and states where equivalent operation remains an assumption Krasheninnikova et al. (2020) Rowe & Healy (2014).
Calibration should include both positive and negative controls. A near-trivial condition estimates basic engagement and perception, while a manipulation designed to remove diagnostic information tests whether success depended on the intended cue. For example, a comparative memory paradigm can use immediate retrieval to establish encoding, longer intervals to estimate retention, and a condition with an irrelevant but salient competitor to test distraction. The same species may require a different response interface to access the task; where interfaces change, researchers should validate their consequences before interpreting between-species differences. A factorial design with control conditions across each taxon helps distinguish biological differences from measurement artifacts ManyPrimates (2019) Jelbert et al. (2014).
The strongest calibration strategy uses bridging designs: overlapping versions of the assay administered within selected taxa before extension to a broader phylogenetic sample. A bridging population can demonstrate whether replacing a motor effector or reward modality preserves individual rank order or the effect of a cognitive manipulation. Such evidence provides a concrete basis for comparison while making its boundaries visible. If the bridge fails, the experimenter may identify two valid within-species tasks that cannot yet support a shared quantitative scale. Reporting that result improves the science by preventing unwarranted equivalence claims. Ecological flexibility and measurement standardization become connected, testable design decisions Cauchoix et al. (2018) Krasheninnikova et al. (2020).
Preregistration, Open Materials, and Reproducible Multitaxon Protocols
A comparative preregistration should identify its research question, intended populations and sites, the construct-to-task mapping, primary outcomes, training criteria, exclusion rules, plausible confounds, allocation of trials, sample-size rationale, and planned analysis. It should predefine which protocol elements must remain fixed to preserve the information structure and which can be modified according to a species-specific calibration rule. Researchers can preregister separate confirmatory and exploratory components; exploratory changes should be dated and justified rather than concealed. This practice allows a novel apparatus improvement to enter the scientific record without rewriting an original prediction after the data are known ManyPrimates (2019) Farrar et al. (2021).
Open materials extend beyond data tables and scripts. An assay is more fully reproducible when a research archive includes measured apparatus dimensions, stimulus-generation files, lighting or auditory specifications, reward descriptions, video examples of successful and unsuccessful trials, administration instructions, training logs, equipment variations, and the coding scheme. Such records allow later investigators to evaluate whether an apparent replication used equivalent sensory and motor demands. For cross-taxon work, exceptions should be recorded in a site-by-species protocol matrix. This degree of procedural transparency helps separate biological heterogeneity from differences introduced by the testing infrastructure ManyPrimates (2019) Farrar et al. (2021).
Collaborative studies must plan a division between central coordination and local expertise. A centralized protocol can define core causal contrasts, data dictionaries, and statistical procedures, while local teams supply anatomical, behavioral, husbandry, and ecological knowledge. Joint pilot reviews can flag situations in which one species cannot express the expected response, so that a meaningful adaptation is made before full data collection. Site effects should be modeled and interpreted as part of the evidence. The ManyPrimates example illustrates this collaborative potential, while the 2026 meta-meta-analysis by Mizuno and colleagues underscores the value of better-designed, more transparent and more diverse evidence beyond strongly represented laboratory rodent paradigms ManyPrimates (2019) Mizuno et al. (2026).
Critical Comparisons: What Species Differences Can Establish
A species contrast is informative only relative to the matched operations and populations. A higher average success rate can show more effective performance at a specific task. It can support a stronger explanation when experiments demonstrate equivalent perceptual access, feasible motor responses, comparable incentive value, and adequate sampling. It can identify an evolved difference only within an appropriate comparative framework that addresses historical dependency and independent ecological evidence. These levels should be stated explicitly because each demands additional observations. Cognitive testing is most informative when the scientific question is narrow enough that several plausible causal accounts lead to different measurable predictions Felsenstein (1985) Sol et al. (2025).
Failures are equally important to interpret in terms of their diagnostic structure. If a species performs at chance on every tested delay, even when the reward is immediately available, the memory inference is undermined by failure of the encoding baseline. If it succeeds with a simple tool but not a visually transformed tool, the contrast might reflect a limit in transfer, an unfamiliar grip, or reduced salience. If a subject declines to engage, the investigator has measured participation under the available conditions, not an absence of cognition. Scientific conclusions should follow this conditional structure, distinguishing the limits of tested performance from the limits of the organism's capacities Howard & Barron (2024) Jelbert et al. (2014).
Some debates arise from incompatible conceptual vocabularies rather than from contradictory observations. Studies may describe the same behavioral pattern as intelligence, executive control, cognitive flexibility, rule learning, or associative adjustment. A comparative evidence table should identify what changed experimentally, which behavior was counted, which alternatives were controlled, and which construct each interpretation invokes. Such cross-study alignment can reveal where apparently conflicting conclusions actually concern distinct tasks and where claims of shared underlying operations remain underdetermined. This analytical discipline is especially important when comparing phylogenetically distant species whose successful behavior is superficially similar Shettleworth (2012) Boogert et al. (2018).
Cognition, Artificial Systems, and the Boundary of Functional Comparison
The Artificial Era extends comparative psychological methodology to systems whose sensing, action, memory, and developmental histories may be engineered rather than biological. The legitimate scientific bridge is operational: define the information available to a system, its admissible actions, the relevant task contingencies, and the causal manipulations intended to identify a process. A language model, a robot, and a food-caching bird can each succeed at a delayed-selection benchmark, yet the interface, training history, memory substrate, and reward structure can differ radically. Functional analogy is a hypothesis about transferable relations between inputs, internal state or process, and outputs; it should be evaluated through controlled intervention and cross-context transfer Rowe & Healy (2014) Krasheninnikova et al. (2020).
Species-appropriate testing provides a general lesson for cross-system psychology: identical textual or visual prompts need not impose equal demands on differently constructed systems. For an artificial agent, contamination of training data, prompt sensitivity, tool availability, persistent memory, and access to external retrieval are potential pathways to success that must be explicitly controlled. For a biological organism, perception, motivation, and movement place additional constraints on the test. An informative comparison can match a causal role while varying the delivery interface, but it should report differences in learning regime, feedback, and opportunity for adaptation. Similar performance is evidence of an outcome, and additional tests are required to infer similar mechanisms or psychological organizations ManyPrimates (2019) Boogert et al. (2018).
An emerging program for Psychology of Artificial can apply the same hierarchy used in comparative animal work: behavioral competence, construct validation, reliability, generalization, causal intervention, and longitudinal stability. This program is a research proposal in the intellectual horizon of Artificial Era and From Homo to Artificial. It does not transform methodological analogy into proof of consciousness, affective experience, or biological homology. Carefully delimited cross-system tests may instead establish where artificial and biological systems implement related functions, where their response structures diverge, and which experimental manipulations reveal those differences Shettleworth (2012) Sol et al. (2025).
Unresolved Problems and the Next Research Program
The first outstanding problem is measurement invariance across distant taxa. A practical research program should develop calibrated item banks with multiple sensory interfaces, define the informational relations invariant across those versions, and collect bridge data in species capable of performing more than one version. Experimental manipulation of surface features can test whether a common relation governs performance. If consistency appears only under one apparatus, a broad cross-species latent scale remains questionable. The research output should include validated conditions and explicit noncomparabilities, not simply species ranks Krasheninnikova et al. (2020) Rowe & Healy (2014).
The second challenge is to disentangle within-individual change from enduring between-individual variation and cross-species differences. Longitudinal, multisite cohorts can estimate how repeated training, maturation, physiological state, and environmental variation alter scores. Such designs must keep a record of prior experimental exposure, voluntary participation, and opportunities for learning. Preplanned analyses can decompose individual, family, site, species, and phylogenetic variance, although their estimability depends on adequate sampling at every level. New task batteries should be chosen for independent sensitivity to the target operations rather than their convenience or popularity in earlier laboratories Cauchoix et al. (2018) Farrar et al. (2021).
The third challenge is the transition from experimental competence to evolutionary explanation. A laboratory result can motivate ecological field measures, but causal relations to survival, resource acquisition, social reproduction, or fitness must be tested with their own longitudinal methods. The synthesis by Sol and colleagues emphasizes that studies still have limited information about heritable cognitive variation and its sustained consequences. Future programs should combine a carefully validated assay with field outcomes and phylogenetic designs while reporting the uncertainty each method cannot resolve. The 2026 meta-meta-analytic findings further justify larger, transparent, multi-taxon replication programs with explicit statistical power and bias assessments Sol et al. (2025) Mizuno et al. (2026).
Conclusion
Species-appropriate comparative cognitive testing is a discipline of scientific inference. Its central requirement is that a behavioral score be connected to an explicitly defined psychological operation under conditions the tested organism can perceive, enact, and sustain. The comparative value of a task arises from calibrated construct alignment, experimentally constrained alternative explanations, and transparent population coverage. Physically identical apparatuses do not guarantee psychological equivalence, and species-adjusted apparatuses can be compared when the relevant informational and causal relations have been demonstrated Krasheninnikova et al. (2020) Rowe & Healy (2014).
Research on self-control, delayed memory, social prediction, physical causality, and foraging demonstrates both the productivity and the interpretive difficulties of comparative assays. The MacLean multispecies project and ManyPrimates collaboration expanded the scope of data collection, while subsequent work on sampling, motivation, task repeatability, and the fragility of cumulative statistical evidence identified major requirements for stronger conclusions. Scientific integration depends on preserving the distinctions among observed performance, cognitive explanation, and evolutionary history, then specifying what new experiments could connect those levels MacLean et al. (2014) ManyPrimates (2019) Mizuno et al. (2026).
The resulting research agenda is concrete: preregister construct-level predictions, validate sensory and motor prerequisites, record motivation and selection into testing, use multiple diagnostic tasks and bridge conditions, model individuals and species as structured samples, and publish failures and replications alongside positive findings. Cross-system inquiry in the Artificial Era inherits the same evidentiary responsibilities when comparing organisms with artificial agents. A mature comparative psychology gains explanatory precision through designs that are appropriate to the systems investigated and sufficiently explicit to support cumulative independent verification Farrar et al. (2021) Sol et al. (2025).
References
- Beran, M. J. (2015). The comparative science of “self-control”: What are we talking about? Frontiers in Psychology, 6, Article 51. https://doi.org/10.3389/fpsyg.2015.00051
- Boogert, N. J., Madden, J. R., Morand-Ferron, J., & Thornton, A. (2018). Measuring and understanding individual differences in cognition. Philosophical Transactions of the Royal Society B: Biological Sciences, 373(1756), 20170280. https://doi.org/10.1098/rstb.2017.0280
- Cauchoix, M., Chow, P. K. Y., van Horik, J. O., et al. (2018). The repeatability of cognitive performance: A meta-analysis. Philosophical Transactions of the Royal Society B: Biological Sciences, 373(1756), 20170281. https://doi.org/10.1098/rstb.2017.0281
- Farrar, B. G., Voudouris, K., & Clayton, N. S. (2021). Replications, comparisons, sampling and the problem of representativeness in animal cognition research. Animal Behavior and Cognition, 8(2), 273–295. https://doi.org/10.26451/abc.08.02.14.2021
- Felsenstein, J. (1985). Phylogenies and the comparative method. The American Naturalist, 125(1), 1–15. https://doi.org/10.1086/284325
- Horner, V., & Whiten, A. (2005). Causal knowledge and imitation/emulation switching in chimpanzees (Pan troglodytes) and children (Homo sapiens). Animal Cognition, 8, 164–181. https://doi.org/10.1007/s10071-004-0239-6
- Howard, S. R., & Barron, A. B. (2024). Understanding the limits to animal cognition. Current Biology, 34(7), R294–R300. https://doi.org/10.1016/j.cub.2024.02.043
- Jelbert, S. A., Taylor, A. H., Cheke, L. G., Clayton, N. S., & Gray, R. D. (2014). Using the Aesop's fable paradigm to investigate causal understanding of water displacement by New Caledonian crows. PLOS ONE, 9(3), e92895. https://doi.org/10.1371/journal.pone.0092895
- Kano, F., Krupenye, C., Hirata, S., Tomonaga, M., & Call, J. (2019). Great apes use self-experience to anticipate an agent's action in a false-belief test. Proceedings of the National Academy of Sciences, 116(42), 20904–20909. https://doi.org/10.1073/pnas.1910095116
- Krasheninnikova, A., Chow, P. K. Y., & von Bayern, A. M. P. (2020). Comparative cognition: Practical shortcomings and some potential ways forward. Canadian Journal of Experimental Psychology, 74(3), 160–169. https://doi.org/10.1037/cep0000204
- 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
- MacLean, E. L., et al. (2014). The evolution of self-control. Proceedings of the National Academy of Sciences, 111(20), E2140–E2148. https://doi.org/10.1073/pnas.1323533111
- Many Primates, Altschul, D. M., Beran, M. J., Bohn, M., et al. (2019). Establishing an infrastructure for collaboration in primate cognition research. PLOS ONE, 14(10), e0223675. https://doi.org/10.1371/journal.pone.0223675
- Mizuno, A., Pollo, P., Lagisz, M., Guillette, L., Soma, M., & Nakagawa, S. (2026). Statistical fragility in the meta-analytic evidence base of animal cognition: A meta-meta-analytic reappraisal. Biology Open, advance online publication, bio.062841. https://doi.org/10.1242/bio.062841
- Rowe, C., & Healy, S. D. (2014). Measuring variation in cognition. Behavioral Ecology, 25(6), 1287–1292. https://doi.org/10.1093/beheco/aru090
- Shaw, R. C. (2017). Testing cognition in the wild: Factors affecting performance and individual consistency in two measures of avian cognition. Behavioural Processes, 134, 31–36. (Published online in 2016.) https://doi.org/10.1016/j.beproc.2016.06.004
- Shettleworth, S. J. (2012). Modularity, comparative cognition and human uniqueness. Philosophical Transactions of the Royal Society B: Biological Sciences, 367(1603), 2794–2802. https://doi.org/10.1098/rstb.2012.0211
- Sol, D., Bateman-Neubert, A., Noguer, L., & Taylor, A. H. (2025). The evolutionary puzzle of cognition: Challenges and insights from individual-based studies. Philosophical Transactions of the Royal Society B: Biological Sciences, 380(1929), 20240123. https://doi.org/10.1098/rstb.2024.0123
- Splinter, F., Fichtel, C., & Radespiel, U. (2025). Cognitive performance in wild and captive grey mouse lemurs (Microcebus murinus). International Journal of Primatology, 46, 644–663. https://doi.org/10.1007/s10764-025-00490-6
- Tinbergen, N. (1963). On aims and methods of ethology. Zeitschrift für Tierpsychologie, 20(4), 410–433. https://doi.org/10.1111/j.1439-0310.1963.tb01161.x
- van Horik, J. O., & Madden, J. R. (2016). A problem with problem solving: Motivational traits, but not cognition, predict success on novel operant foraging tasks. Animal Behaviour, 114, 189–198. https://doi.org/10.1016/j.anbehav.2016.02.006
- van Horik, J. O., Langley, E. J. G., Whiteside, M. A., & Madden, J. R. (2017). Differential participation in cognitive tests is driven by personality, sex, body condition and experience. Behavioural Processes, 134, 22–30. https://doi.org/10.1016/j.beproc.2016.07.001
Social-Cognitive Tests and Response-Mode Equivalence
Comparative tests of social understanding expose the limits of relying on one response modality. If a task demands explicit selection among boxes, a subject may fail because it is reluctant to approach, cannot exploit a human experimenter's instruction, or lacks the relevant understanding. Anticipatory looking measures where the animal predicts another agent will act and reduces some requirements associated with explicit choice. In the 2016 great-ape false-belief study, eye tracking detected anticipatory looking consistent with an agent's outdated information. The inference is specific to the stimulus conditions and predictive behavior; it does not by itself settle all questions about explicit belief attribution Krupenye et al. (2016).
Follow-up experiments must separate agent-directed anticipation from learned stimulus contingencies. Kano and colleagues subsequently examined the relevance of apes' own experience for performance on a false-belief anticipation task. Such tests investigate whether the response depends on information about the protagonist's perceptual history, on a simpler associative cue, or on a broader representational operation. The methodological advance is to change a feature predicted to affect one theoretical account and observe whether performance changes accordingly. Identical gaze metrics across taxa may still require different baselines because visual exploration and social attention differ among individuals and species Kano et al. (2019) Krupenye et al. (2016).
Cross-species social testing also must assess whether the social partner is natural to the organism's interaction system. A human pointing gesture can serve as a meaningful cue for some domesticated or extensively socialized animals while remaining novel or threatening to another population. Substituting a conspecific signal can increase ecological validity, yet may change the underlying question from understanding a human communicator to processing species-specific social information. A good research program studies both conditions with explicit hypotheses about which representational relation remains the same, and regards response-mode calibration as part of the evidentiary chain Krasheninnikova et al. (2020) Tinbergen (1963).