Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Human–Artificial Health Interaction, Decision Support, and Safety

Abstract. Human–artificial health interaction is the situated process through which patients, clinicians, caregivers, institutions, and computational systems exchange information, interpret evidence, and coordinate health-related decisions. This monograph examines the psychological organization of that process, with particular attention to calibrated reliance, decision delegation, communication, automation bias, professional accountability, equity, privacy, accessibility, and patient safety. It integrates foundational human-factors scholarship, original experimental studies, prospective clinical trials, independent validation, reporting standards, and health-governance guidance through a critical narrative synthesis. The evidence shows that performance measured for a model in isolation is only one component of the safety and effectiveness of a clinical decision-support arrangement. Human response to suggestions, quality of the underlying data, timing and interface of advice, clinical workflow, and organizational capacity to detect and correct failures jointly determine outcomes. Results vary across tasks and study designs: experimental evidence documents susceptibility to misleading advice, some simulations demonstrate gains from well-designed support, and prospective screening studies show benefits in defined workflows. This synthesis establishes a research framework centered on the observable decisions of human–artificial teams, the consequences of those decisions for patients, and the continuing verification of systems in their actual settings of use.

Keywords: human–artificial health interaction; clinical decision support; human factors; automation bias; calibrated reliance; patient safety; diagnostic reasoning; AI-assisted care; trust; clinical AI; health equity; patient autonomy; model monitoring; Artificial Era.

Scientific Overview and Research Questions

Human–artificial health interaction arises whenever information or recommendations generated by a computational system influence the understanding, choice, action, or experience of a person involved in health care. Its most consequential expression is a decision pathway: a patient reports a concern, information is gathered, an artificial system produces an estimate or explanation, a clinician or patient interprets the output, and a subsequent action changes the probability or distribution of health outcomes. Each stage introduces a distinguishable source of reliability, uncertainty, and potential error.

The research object is the interactive arrangement as a whole. A model can produce a statistically accurate risk score while its use causes inappropriate action; a less accurate model can sometimes help a professional identify neglected evidence; and an apparently fluent dialogue can conceal serious omissions in urgent-care guidance. The field therefore requires joint measurement of the human decision-maker, the artificial system, their interface, the institution that governs their use, and the person whose interests are affected. The human side includes beliefs, cognitive workload, attention, vulnerability, health literacy, professional expertise, and willingness to seek clarification. The artificial side includes training distribution, input validity, outputs, uncertainty, interface behavior, version changes, and failure modes.

Four questions organize this investigation. Under which conditions does artificial advice improve actual clinical or patient-centered decisions? How are trust and reliance calibrated to a tool's demonstrated competence on a specific task? Which interactions among cognitive bias, workflow constraints, data selection, social inequality, and governance create safety hazards? Which experimental and implementation designs provide credible evidence of patient benefit and permit timely detection of harm? These questions connect health psychology with cognitive psychology, human factors, implementation science, medical informatics, clinical epidemiology, and the psychology of human–artificial relations.

The scope includes clinician-facing triage and diagnosis, predictive alerts, screening, patient-facing conversational information, shared decision-making, and caregiving contexts. It addresses therapeutic dialogue only where the communication mechanism bears on safety, informed choice, or responsibility; psychotherapy efficacy is a separate clinical research object. General wellness reminders and ordinary telehealth delivery are treated as adjacent technologies, while the distinctive object here is artificial participation in health-related interpretation or judgment.

Scientific Method and Evidence Hierarchy

This monograph is a critical narrative synthesis of selected original experiments, randomized evaluations, external validation studies, theoretical human-factors work, reporting guidelines, and official policy documents available by November 3, 2026. It has no registered systematic-search protocol, PRISMA flow diagram, exhaustive screening denominator, pooled effect estimate, or newly collected participant data. Source selection prioritizes original empirical evidence and established methodological guidance that directly tests or defines the interactive mechanism. The synthesis uses reporting frameworks to appraise evidence, rather than falsely claiming that the narrative review itself follows a trial or diagnostic-accuracy reporting protocol.

Evidence is stratified by what it establishes. Retrospective test-set comparisons evaluate technical performance under a fixed data distribution. Reader studies test changes in professional behavior under controlled information conditions. Randomized vignette studies estimate short-term causal effects on simulated judgments. Live pragmatic trials can evaluate the effects of a configured workflow on detection, workload, errors, and sometimes patient outcomes. Observational surveys describe attitudes and associations, while independent external validation tests generalization beyond a developer's original environment. Governance standards articulate duties and safeguards; their adoption is not evidence that a product is safe.

The distinction matters because many published AI claims concern a model's sensitivity, specificity, or examination score rather than a measurable patient outcome. The reporting framework Vasey, B., Nagendran, M., Campbell, B., et al (2022) explicitly places human factors and real clinical use within early evaluation. Its 17 AI-specific and ten generic reporting items emphasize the intervention in context. Liu, X., Cruz Rivera, S., Moher, D., Calvert, M. J., & Denni (2020), Collins, G. S., Moons, K. G. M., Dhiman, P., Riley, R. D., B (2024), and Sounderajah, V., Guni, A., Liu, X., et al (2025) address different kinds of trial, prediction-model, and diagnostic-accuracy reports. Their coexistence demonstrates why no single statistic or reporting checklist can establish complete clinical safety.

Historical Foundations of Clinical Decision Assistance

The development of artificial decision support follows several intersecting histories: probabilistic inference and clinical diagnosis; computerized medical records; expert-system knowledge representation; medical imaging and predictive modeling; cognitive engineering; and more recent generative language models. Across these traditions, a recurring technical ambition is to provide useful information at the moment of decision. The corresponding psychological challenge is to establish when information is noticed, questioned, accepted, ignored, or misunderstood.

The science of automation supplied an early framework for interpreting those difficulties. Parasuraman, R., & Riley, V (1997) distinguished use, misuse, disuse, and abuse of automation, linking inappropriate reliance to monitoring failures and the location of responsibility. Skitka, L. J., Mosier, K. L., & Burdick, M (1999) experimentally analyzed errors of omission and commission in a simulated flight task with an imperfect decision aid. Their results arose outside medicine, so they establish a general cognitive mechanism rather than a direct clinical effect. The transfer to medicine requires task-specific testing because medical decisions have distinct evidence, base rates, consequences, and accountability structures.

The twenty-first-century expansion of statistical learning transformed the objects available for assistance. Rajkomar, A., Dean, J., & Kohane, I (2019) described the opportunities and methodological problems of machine learning in medicine. Imaging models, electronic-record risk scores, and language systems operate through different architectures and should be evaluated in relation to their distinct intended uses. The technical history is therefore also a history of changing human assignments: from reading a prompt in a workstation to reviewing an image ranking, interpreting an alert, or negotiating an apparently conversational recommendation.

Key Constructs and Operational Definitions

A clinical decision-support system is a configured technology that supplies patient-specific or situation-specific information, estimates, options, or recommendations relevant to a health-related choice. Artificial decision support is a subset in which a computational model participates in producing the output. A deployment is the particular combination of model version, data interface, threshold, user group, setting, workflow, and escalation policy. The unit of evaluation can be the individual output, the user decision, the care episode, or the population outcome; those units must be identified separately.

Reliance is an observable behavioral response: using advice, changing a decision after advice, declining to seek additional information, or delegating a step. Trust is an attitude or expectation concerning competence, integrity, predictability, or benevolence of a system or its deploying organization. Trust and reliance correlate in some contexts, but they have different measures and consequences. A person may comply with a distrusted alert because institutional procedure requires it, or praise a system without using its advice.

Calibrated reliance is the alignment of acceptance and rejection behavior with the system's demonstrated competence in the relevant circumstances. Overreliance occurs when a user gives inappropriate weight to erroneous output; underreliance occurs when useful output is rejected or overlooked. Automation bias describes systematic judgment error associated with over-weighting automated advice or failing to detect what an automated system did not flag. It is empirically assessed by comparing behavior with and without advice and by manipulating whether the advice is correct.

Safety refers to the avoidance or minimization of preventable harm associated with the use of a system, measured against a specified task and baseline. Reliability refers to consistent performance under stated conditions; accuracy refers to correspondence with a chosen reference standard. These properties are connected but require different evidence. Clinical utility concerns whether use changes decisions, workflow, or outcomes in ways relevant to care. A high test-set score establishes neither actual utility nor an acceptable harm distribution.

Accountability is the assignment of responsibility for model design, validation, authorization, communication, clinical decisions, auditing, incident response, and compensation or remedy. These responsibilities often belong to different people and institutions. They should be specified through governance rather than inferred from whether a user clicked 'accept.'

The Human–Artificial Health Decision as a Distributed Process

A health decision has a temporal and institutional structure. Patients determine what they disclose, interpret possible consequences, and supply information that may be incomplete. Clinicians integrate history, examination, probabilities, professional norms, practical constraints, and patient preferences. Artificial systems process selected input representations and return outputs based on specified objectives and training histories. Organizations determine thresholds, staffing, supervision, permissions, interface placement, and escalation pathways.

A defensible research model follows the decision sequence: available evidence; human and artificial initial estimates; presentation of support; verification or contestation; final choice; implementation; feedback; patient consequences. The value of support is measured through the change from the counterfactual pathway, ideally with randomized allocation or appropriately designed causal analysis. Errors may occur because the input omits an essential feature, the model generalizes poorly, the user misreads output confidence, a warning arrives at the wrong time, or the organization lacks the capacity to act on a correct flag.

The object is thus sociotechnical without losing psychological specificity. Each human transition can be observed through decision logs, response times, attention, recall, confidence ratings, explanation comprehension, counterfactual choices, patient interviews, and validated measures of perceived control. Each technical transition requires provenance and version traceability. A complete evaluation links these levels to real outcomes while documenting uncertainty about causation.

Relationship to the Artificial Era

The historical horizon of the Artificial Era directs attention to sustained coordination between Homo and Artificial systems in consequential institutions. In this research, that horizon defines a field of testable human–artificial relations: the redistribution of attention, memory, information selection, and decision influence. It does not establish clinical competence, sentience, independent medical authority, or therapeutic effectiveness merely by naming an artificial participant.

The psychology of artificial participation can investigate whether users attribute intention or understanding to language interfaces, how these attributions affect disclosure, and how patients and professionals distinguish the system's displayed confidence from empirically measured reliability. It can also examine who becomes accountable when advice is produced across multiple human and machine steps. Empirical results on these questions must remain distinct from philosophical claims concerning the broader historical status of Artificial.

The Clinical Problem of Model Performance Versus Team Performance

Statistical classification studies ask whether scores align with outcomes. Human–artificial interaction studies ask whether people use scores effectively. A system can surpass an individual reader on a retrospective test set and still fail to improve the performance of a human assisted by it. Model-only comparisons usually leave out persuasion, anchoring, workload, incomplete access to context, and the cost of additional actions.

The distinction is illustrated by two complementary bodies of evidence. McKinney, S. M., Sieniek, M., Godbole, V., et al (2020) evaluated an imaging system in international breast-screening datasets, an important technical achievement whose clinical use still required prospective pathway evaluation. Tschandl, P., Rinner, C., Apalla, Z., et al (2020) tested human–computer collaboration for skin-cancer recognition and found gains in specified experimental decision settings, with effects dependent on representation of the support and clinician expertise. Their findings concern defined visual tasks, not general medical competence.

The fundamental synthesis is that model performance, human–model team performance, and patient benefit form three different inferential targets. A research program that moves from one to the next must carry forward the assumptions and limitations of each preceding stage.

Trust as a Behavioral and Institutional Mechanism

Lee, J. D., & See, K. A (2004) formulated trust in automation in terms of appropriate reliance under complexity and uncertainty. In health care, trust has additional institutional dimensions: patients may accept a recommendation because it comes through a trusted clinician or health system; clinicians may rely on embedded scores because a hospital has selected and authorized a vendor; and both may struggle to distinguish official deployment from independent scientific validation.

Trust is also object-specific. A professional may trust an algorithm to identify a radiological feature while doubting its recommendations for management. A patient may trust an interface to clarify terminology while declining to let it determine urgent care. Research should ask: trust in which capacity, for which task, on whose evidence, and at what cost of a mistake? Broad satisfaction questions often obscure the behavioral threshold at which trust becomes consequential.

The survey reported by Nong, P., & Platt, J (2025) examined US adults' trust in health systems to use AI responsibly. Its cross-sectional design supports analysis of associated attitudes, not causal claims about what a particular disclosure practice would achieve. A rigorous psychology of trust combines survey measures with observable responses to correct and incorrect advice, longitudinal changes in trust after failures, and patient understanding of the system's role.

The Boundary Between Assistance and Delegation

Assistance is an input to a human choice; delegation transfers a defined subtask or decision authority under conditions of authorization and oversight. These categories vary by design. An alert that is ignored has little direct influence; an automatically generated order can act before a clinician reviews it. A patient-facing conversational system can de facto become a decision authority if a user treats its fluent explanation as permission to delay care. The psychological meaning of delegation depends on actual behavior, not a disclaimer or a vendor's description.

A comprehensive evaluation identifies who authorizes a system to recommend, who is expected to verify, which actions can occur without review, when escalation is required, and how errors are corrected. Clinical responsibility is preserved through an explicit chain of roles and procedures. The appropriate degree of autonomy must be justified against intended use, reliable evidence, patient interests, and applicable regulation. The Food and Drug Administration (2026) distinguishes specified non-device clinical decision-support functions from software functions that remain medical devices; the regulatory determination depends on functional criteria rather than a generic 'AI' label.

Reliance Calibration and Confidence Formation

Reliance calibration can be studied through the relation between a support system's actual correctness and a person's propensity to follow it. A simple design randomizes exposure to advice, independently classifies advice as correct or erroneous against a reference standard, and observes acceptance, verification, and final accuracy. More informative designs include cases with different prevalence, evidence quality, ambiguity, and potential harm. Confidence ratings before and after advice help distinguish an informed update from an unexamined shift in belief.

The psychology of reliance includes experience, learning, situational pressure, prior success, professional status, and the meaning attached to a confident recommendation. A clinician facing a high workload may choose the most salient suggested action even while reporting low trust. A patient may accept an explanation because it is available immediately and uses familiar vocabulary. Both responses are conditional adaptations to circumstances. They should be tested against the system's conditional accuracy and the user's opportunity to inspect contrary evidence.

Appropriate calibration involves knowing when to withhold reliance. Clear presentation of uncertainty, task limitation, and failure patterns can support this capacity, but the effect of such design choices is empirical and sometimes mixed. Human observers can mistake visual precision, probability formatting, or lengthy rationales for valid evidence. The measure that matters is the change in decisions on cases where the system is correct and on cases where it is mistaken. A tool that increases both acceptance of correct and acceptance of erroneous advice may produce net benefit in one setting and unacceptable harms in another.

Automation Bias and Error Propagation

Automation bias includes omission errors, when people overlook an event that a tool fails to flag, and commission errors, when they follow an incorrect suggestion despite evidence supporting another choice. Skitka, L. J., Mosier, K. L., & Burdick, M (1999) established these forms in nonclinical simulation; the modern healthcare question is whether similar error structures occur for clinical evidence, imaging, and conversational recommendations. Gaube, S., Suresh, H., Raue, M., Merritt, A., Berkowitz, S (2021) provided direct evidence in a chest-radiograph decision study: inaccurate advice worsened diagnostic accuracy regardless of whether participants believed it came from AI or a human expert.

The finding resists a simplistic explanation that professionals are automatically captivated by machines. Source labeling affected some judgments of advice quality; task expertise and objective correctness also mattered. In practice, an incorrect computer output can be persuasive because it appears in the regular workflow, points attention toward a plausible interpretation, or is congruent with a user's preliminary hypothesis. Likewise, a correct warning may be dismissed after repeated false alarms. The critical variable is the interaction between evidence quality and the recipient's decision behavior.

Research should measure overreliance separately from mere agreement. When physician and system independently reach the same correct conclusion, agreement does not demonstrate influence. The strongest experimental comparisons require pre-advice answers and randomized information exposure, along with trials in which the model is deliberately wrong or uncertain. Ethical simulation permits such testing without exposing actual patients to contrived harmful recommendations.

Anchoring, Confirmation, and Order Effects

Advice influences not only the final selected option but the path of subsequent reasoning. An early diagnostic suggestion can become an anchor around which a user organizes later findings. A suggested label can direct attention to confirmatory evidence, while disconfirming symptoms receive less scrutiny. Pre-existing clinical expectations and artificial output can reinforce one another, producing a strong feeling of coherence unsupported by a complete differential diagnosis.

The timing of artificial advice is therefore a manipulable feature of intervention design. Researchers can present advice before an initial judgment, after an independent assessment, or after a structured review of discriminating evidence. The appropriate sequence varies by task, time pressure, and user expertise. Delayed disclosure may protect independent reasoning while imposing time cost; early advice may be valuable when a rare emergency requires immediate attention. Testing should record errors, escalation time, cognitive workload, and patient-relevant downstream actions rather than treating any reduction of influence as automatically desirable.

A fully specified protocol distinguishes between bias induced by advice and bias present before exposure. It uses blinded adjudication, prespecified hypotheses, and analysis of discordant cases. These measures support causal interpretation and reduce selective presentation of memorable mistakes.

Cognitive Workload, Interruptions, and Alert Fatigue

Clinical work unfolds within competing tasks, interrupted attention, changing priorities, and often inadequate time. A decision aid adds information, but also adds a demand to interpret, verify, and potentially act. If the output arrives too often or is too nonspecific, it can increase interruption burden and train users to ignore signals. An apparent improvement in algorithm sensitivity may fail to improve care when each true event is accompanied by numerous low-value warnings.

Wong, A., Otles, E., Donnelly, J. P., Krumm, A., McCullough (2021) independently evaluated a deployed sepsis prediction model and found performance and alert-burden problems relative to the model's claimed potential. Their retrospective validation illustrates why an alert threshold should be judged in a setting where prevalence, existing practice, and available response capacity are known. It does not imply that all sepsis algorithms or early-warning configurations perform identically.

The psychology of alarm exposure includes habituation, competing priorities, perceived urgency, uncertainty over who owns a warning, and the accumulated memory of false alarms. These mechanisms interact with staffing and workflow. The safer design question is not simply how many alerts the algorithm generates, but how the entire organization recognizes, prioritizes, assigns, and resolves them. Experimental process tracing and interrupted-time-series evaluations can help separate user attention failures from organizational inability to respond.

Expertise, Training, and the Shape of Assistance

Clinical expertise can both improve and complicate human–AI collaboration. Experts may identify subtle model errors and adjust a recommendation to fit the patient's context. They may also face stronger commitment to an initial diagnosis or undervalue unfamiliar forms of evidence. Less experienced users may gain more from well-targeted support, as Tschandl, P., Rinner, C., Apalla, Z., et al (2020) observed within their skin-lesion recognition experiments. The result is task-specific and should be investigated in other fields before generalization.

Training interventions need to convey more than menu functions. The relevant competence is to interpret output probability, know common failure conditions, recognize out-of-distribution cases, inspect source information, and escalate disagreement. One proposed measurable outcome is the ability to identify instances when the system is likely to be wrong. Another is the capacity to explain to a patient what role the system played in a recommendation.

Experience with general conversational AI is not equivalent to expertise with a regulated diagnostic system. The 2024 and 2025 randomized physician studies by Goh, E., Gallo, R., Hom, J., Strong, E., Weng, Y., Kerman, H (2024) and Goh, E., Gallo, R. J., Strong, E., et al (2025) underscore that even within clinical vignettes, observed benefits depend on the type of reasoning tested, the model, and the interaction task. Training research should therefore identify task categories, learning curves, and degradation after infrequent use.

Mental Models of Artificial Advisers

A mental model is the user's working representation of how a system produces outputs, the conditions under which it performs well, and how it may fail. In artificial health interaction, users encounter systems whose surface behavior can hide their internal limits. A probability may be interpreted as certainty about an individual; a language model may be assumed to have accessed an electronic medical record when it has not; an explanation may be read as a verified causal account even when it is generated from statistical associations.

Mental models are observable. Researchers can ask users to predict a system's behavior on altered inputs, describe what evidence informed the recommendation, estimate accuracy in unusual cases, or identify when a tool was unavailable. These tests establish whether explanations actually improve operational understanding.

The general explainable-AI tradition includes Ribeiro, M. T., Singh, S., & Guestrin, C (2016) local interpretability method. Its contribution concerns ways of representing model behavior, not evidence that a specific explanation format makes a medical deployment safe. Explanation fidelity, human comprehension, and downstream decisions should be evaluated independently. Explanations that are attractive, concise, or technically sophisticated can still fail to support appropriate reliance.

Uncertainty, Probability, and Risk Communication

Clinical predictions require interpretation of sensitivity, specificity, calibration, prevalence, and the cost of false-positive and false-negative actions. A risk score can be numerically precise while conveying substantial uncertainty about a person's eventual outcome. For rare conditions, a model with good discrimination may have a low positive predictive value when applied to the general population. A threshold determines which people trigger action, and this threshold should reflect downstream consequences and capacity.

Human–artificial interaction research asks what happens when probabilities are presented in natural language, percentages, risk categories, graphics, or normative action prompts. The language 'high confidence' may be understood differently by physicians and patients. The reliability of a model on a benchmark does not establish calibration of a language model's own self-reported confidence. Researchers should compare users' interpretation with reference frequencies and observe whether comprehension changes decisions.

Risk communication becomes ethically consequential when patients have limited literacy, face severe consequences, or are making decisions under fear and time pressure. A system's uncertainty should be communicated in terms useful for the choice at hand, including material unknowns and the conditions that would change the recommendation. The appropriate communication format is a research question; a single universal display cannot be assumed effective across tasks or populations.

Conversational Fluency and Illusions of Competence

Generative systems can supply fluent explanations, summarize records, translate terminology, and respond interactively to follow-up questions. Those capabilities can expand access to information while changing users' perception of reliability. A convincing narrative may combine correct general knowledge with an incorrect inference about an individual, an omitted emergency warning, or a fabricated source. The psychological hazard is heightened when the system's responsiveness resembles interpersonal understanding.

A clinically important distinction is the difference between perceived helpfulness and verified decision quality. Users may prefer a conversational answer to a terse institutional instruction while receiving less reliable information. Quality evaluation should therefore involve both user comprehension and independent clinical content review, including dangerous omissions, overconfident statements, and situations requiring prompt human assessment.

Yau, J. Y.-S., Saadat, S., Hsu, E., Murphy, L. S.-L., Roh, J (2024) experimentally compared chatbot responses to patient questions about emergency care and reported clinically problematic advice and difficult reading levels in that selected test. Their finite set of questions and systems does not yield a population-wide error rate for all future models. It establishes the importance of application-specific safety testing rather than assuming that natural-language ability constitutes emergency-care competence.

Patient Autonomy and Shared Decision-Making

Patients participate in care as persons with values, histories, preferences, resources, and varying degrees of information access. An artificial explanation changes the informational environment of a clinical encounter and may change who appears to possess authority. A decision remains meaningfully shared when the patient can understand the alternatives and their consequences, ask questions, communicate preferences, and participate without coercive framing.

An AI-generated ranking can assist discussion when it is tied to documented evidence and allows correction of missing information. It can undermine participation when it is presented as a final verdict, conceals uncertainty, or introduces subtle framing that discourages disagreement. Autonomy is therefore an observable aspect of interaction, not simply a checkbox for obtaining consent to technology.

The World Health Organization (2021) identifies human autonomy, well-being, transparency, accountability, inclusiveness, and sustainability as core principles in AI health governance. These are normative principles. Evaluation must additionally test whether patients understand an AI system's role, are offered relevant choices, and can obtain accessible human review.

Disclosure, Authority, and the Therapeutic Encounter

Disclosure of AI participation matters because patients and clinicians may otherwise misunderstand the source, status, or evidential basis of a recommendation. An effective disclosure specifies the particular task performed, the degree of human review, salient limitations, and a route to challenge or correct an output. Telling a patient only that 'AI was used' communicates little about the decision.

The provider–patient relationship contributes continuity and accountability beyond information transfer. AI support may free time for discussion or create extra demands to explain model outputs. Both possibilities should be assessed with observation, consultation duration, patient comprehension, adherence to agreed plans, and qualitative accounts of how people experience the encounter.

Research should examine whether source disclosure changes trust or reliance across different tasks without assuming a uniform direction. In the experiment by Gaube, S., Suresh, H., Raue, M., Merritt, A., Berkowitz, S (2021), advice-quality judgments differed by source label and expertise, while inaccurate advice impaired diagnosis across labels. The implication is that transparency must accompany substantive reliability and active oversight.

Accessibility, Disability, and Linguistic Inclusion

A system is accessible when intended users can perceive, understand, and operate it under real conditions. Health contexts include people using assistive technologies, older adults with sensory changes, people in acute distress, individuals with limited English proficiency, and users whose cognition or communication is temporarily affected by illness. Accessibility affects accuracy because misunderstanding input questions or output recommendations can directly change health-related behavior.

Evaluation should include representative users, speech and screen-reader compatibility where applicable, text comprehension, translation fidelity, color-independent warning presentation, and the practical possibility of reaching a qualified professional. A platform that works in a controlled laboratory may fail during a noisy emergency, on a shared phone, or with intermittent internet access.

The analytical separation between model fairness and interface accessibility is useful. A model may be statistically comparable across subgroups while its interface excludes some users. Conversely, an accessible interface may present predictions derived from unrepresentative clinical data. Both are necessary dimensions of evaluating human–artificial health interaction.

Privacy, Disclosure, and Information Boundaries

Clinical interactions involve unusually sensitive information: symptoms, diagnoses, medication histories, family circumstances, psychological disclosures, and identifiers. The patient often cannot observe where those data are stored, whether a conversation is used to improve a model, or who can retrieve it later. Privacy is therefore both a technical governance issue and a psychological determinant of willingness to seek care and disclose relevant details.

Safe information design begins with purpose limitation, access control, transparency about retention, security monitoring, and evaluation of data flows through external services. Those safeguards must match actual architecture and relevant law. The existence of a privacy policy is not evidence that users understand it. Researchers can assess disclosure decisions under varying explanations, observed comprehension, and willingness to seek human assistance.

Privacy harms may compound clinical harms: a person who withholds information for fear of exposure may receive less appropriate advice. Thus, privacy and clinical decision quality should be analyzed together rather than in isolated compliance documents.

Relational Continuity and Repeated Use

One-off vignette studies capture a brief interaction; many health uses extend over months or years. Repeated contact may teach a user the system's strengths and weaknesses, but it may also normalize errors, create dependency on a convenient conversational channel, or displace a trusted clinician without deliberate choice. The clinical meaning of repeated use depends on whether the tool remains stable, remembers relevant context reliably, and supports accountable escalation.

Longitudinal designs should monitor changes in confidence, the user's independent problem-solving, reliance after a visible error, discontinuation, and changes in help-seeking. They should differentiate system updates from user learning: otherwise a change in behavior may reflect an altered model rather than improved human calibration.

The longer horizon of human–artificial relations is relevant precisely because care is ongoing. A safe interactive arrangement must withstand changing health states, changing models, departures of professionals, and transitions between institutions.

Evidence from Imaging and Pattern-Recognition Research

Medical imaging offers a relatively structured setting in which models and clinicians can be compared against defined reference standards. Even here, a pathology label or consensus reader judgment may be imperfect, and performance can change when case prevalence and device characteristics differ. Retrospective image classification, blinded reader tests, and prospective clinical screening are progressively different evaluations of a decision-support arrangement.

The systematic review and meta-analysis by Liu et al. (2019)30123-2) examined deep-learning performance in medical imaging compared with healthcare professionals and highlighted limitations in the available evidence. A comparison of aggregate sensitivity or specificity within a selected dataset cannot demonstrate a reduction in actual complications, avoidable procedures, or patient anxiety. Moreover, studies with restricted disease spectrum or highly curated images may overstate transportability.

McKinney, S. M., Sieniek, M., Godbole, V., et al (2020) evaluated a breast-cancer screening system against mammography datasets from the United Kingdom and United States. Its results are important for technical benchmark development and cross-setting evaluation, but a screening service is an organized pathway: image acquisition, recall, biopsy, follow-up, communication, and quality assurance. The downstream decision process must be assessed before translating standalone model estimates into claims about population benefit.

Tschandl, P., Rinner, C., Apalla, Z., et al (2020) studied the addition of algorithmic support to skin-lesion decisions. Accuracy improved under good-quality support, and gains depended on user expertise and the representation of the AI information. The design isolates interaction effects more directly than a model-only benchmark, yet it remains bounded by the lesions, participants, and workflows studied. A clinically deployable service also requires assessment of missed malignancies, referral burden, patient outcomes, accessibility, and long-term learning.

Human–AI Diagnostic Reasoning in Randomized Experiments

The randomized trial by Goh, E., Gallo, R., Hom, J., Strong, E., Weng, Y., Kerman, H (2024) recruited 50 physicians from internal medicine, family medicine, and emergency medicine and compared access to a large language model plus conventional resources against conventional resources in clinical vignettes. The median per-case diagnostic reasoning score was 76% in the assisted group and 74% in the comparison group. The adjusted difference was 2 percentage points, with a 95% confidence interval spanning −4 to 8 points and P = .60. The estimated difference therefore did not establish an improvement in physician reasoning within the study's conditions.

An exploratory comparison also assessed the model alone and found higher performance than the conventional-resources physician group on the study's rubric. This result should not be converted into a claim that autonomous AI is generally safer than physicians. The model-only comparison was exploratory, the tasks were selected vignettes, and clinical action requires information gathering, responsibility for changing patient status, communication, and follow-up. Most importantly, the lack of a benefit from access to a relatively capable system is itself a human–artificial interaction finding: availability does not ensure effective incorporation of evidence.

The trial illustrates the necessity of distinguishing an experimental intention-to-treat contrast from speculation about why collaboration failed to improve performance. Possible explanations include inadequate interface integration, difficulties selecting questions, anchoring, varying user familiarity, and time constraints. Determining which is causal requires new factorial or process-tracing experiments rather than inferring a single mechanism from a null treatment effect.

Human–AI Management Reasoning and Task Dependence

A second randomized physician trial by Goh, E., Gallo, R. J., Strong, E., et al (2025) tested a different outcome: management reasoning across five complex, expert-developed cases. Ninety-two physicians were randomized to GPT-4 with conventional resources or conventional resources alone. Assisted physicians had higher rubric scores, with an adjusted mean difference of 6.5 percentage points (95% CI, 2.7 to 10.2; P < .001), and spent more time per case, with an estimated mean increase of 119.3 seconds (95% CI, 17.4 to 221.2; P = .02). The source includes a subsequent publisher correction; bibliographic and numerical interpretation should follow the corrected record.

The contrast with the 2024 diagnostic reasoning study is scientifically productive. It shows that 'AI assistance in medicine' is an insufficient unit of evidence. Diagnostic classification and management planning require different information, different transformations of uncertainty, and different scoring rules. A trial that demonstrates improved management scores in a simulation does not prove lower mortality, fewer adverse drug events, or better outcomes in routine practice.

The increased time cost is also meaningful. Better rubric performance can coexist with a larger cognitive and operational burden. Future research should quantify when this tradeoff is worthwhile, whether assistance is maintained under real clinical pressure, and whether benefits accrue to all patient subgroups. Inference from simulated tasks should remain explicitly marked as simulation evidence.

Incorrect Advice and the Mechanics of Influence

Gaube, S., Suresh, H., Raue, M., Merritt, A., Berkowitz, S (2021) examined physicians evaluating chest radiographs with advice labeled as either human or AI generated. The advice itself came from human experts and included intentionally inaccurate recommendations. Radiologists as a group rated information believed to originate from AI as lower in quality; less task-expert participants did not show the same overall pattern. Inaccurate advice nonetheless reduced diagnostic accuracy regardless of its supposed source.

The design distinguishes label effects from content effects. A physician's skepticism about a category of adviser does not guarantee immunity from misleading recommendations. Incorrect advice can alter attention and the set of options considered before the user consciously evaluates the adviser as reliable or unreliable. It is therefore important to measure both subjective confidence and the actual deviation from the reference decision.

The study does not directly test all current generative systems or continuous hospital workflows. Its controlled manipulation nevertheless supports a general experimental principle: intentionally seeded errors within ethically safe simulation are essential for assessing whether users can recognize and resist erroneous support. Testing only correctly answered examples cannot measure overreliance.

Prospective Breast Screening as a Workflow Experiment

The Mammography Screening with Artificial Intelligence (MASAI) program in Sweden offers an unusually important progression from technical performance evaluation to prospective pathway evidence. Lång et al. (2023)00298-X) reported a randomized, population-based safety analysis comparing AI-supported screen reading with the standard double-reading workflow. The trial explicitly treated the entire screen-reading arrangement as the intervention, making it more relevant to real care than isolated retrospective image comparison.

A subsequent randomized screening-accuracy report by Hernström et al. (2025)00267-X) examined screening performance and characteristics of detected cancers. Its conclusions need to be read in relation to the actual Swedish screening protocol, the selected AI-assisted reading configuration, participating centers, and national baseline practices. A result in one service design does not establish that any AI tool or any lower-staffed adaptation will preserve the same balance of detection and recall.

Most recently, Gommers et al. (2026)02464-X) reported interval-cancer, sensitivity, and specificity outcomes from the MASAI trial in The Lancet, providing a more mature source of evidence about screening consequences. Interval cancer is an essential outcome because increased screen-detected findings alone may not establish net benefit. Even this evidence cannot be extrapolated uncritically to another country, population, device, screening interval, staffing protocol, or model version.

The cumulative lesson is procedural: clinical value becomes more credible as a sequence of studies examines model accuracy, human workflow, prospective screening decisions, and longitudinal patient-relevant endpoints. Each transition changes the evidential question.

Independent Validation of a Deployed Sepsis Model

Predictive alerts are attractive because they seem capable of directing scarce clinical attention toward impending deterioration. However, an alert is only useful if it accurately identifies clinically consequential risk, arrives at an actionable time, avoids overwhelming users, and triggers an effective response. Wong, A., Otles, E., Donnelly, J. P., Krumm, A., McCullough (2021) externally validated a proprietary sepsis prediction model in an academic health system, with 27,697 patients and 38,455 hospitalizations.

The study reported a hospitalization-level area under the receiver operating characteristic curve of 0.63 (95% CI, 0.62–0.64). At the examined alert threshold, alerts covered a substantial number of hospitalizations while missing many cases. The paper was corrected in August 2021 for a numerical error concerning the number needed to evaluate, so readers should use the corrected version for threshold-specific estimates. This study is a concrete warning about assuming that broad deployment constitutes independent scientific validation.

The finding is observational and retrospective. It does not itself randomize hospital staff to alerts or quantify the causal effect on patient mortality. Its direct contribution is to reveal shortcomings in discrimination, calibration, and potential alert burden in the evaluated setting. A later implementation trial would be required to determine whether redesigned thresholds, presentation, clinical response teams, or alternative models improve outcomes.

Equity, Proxy Targets, and Hidden Objectives

Artificial health systems often learn to predict a convenient available label, and that label may encode historical inequity. A system asked to identify patients with high health needs may be trained on prior spending, service use, or referral rates. These quantities reflect both medical need and unequal access to care. Optimization can reproduce differences in institutional treatment even when protected characteristics are excluded from explicit input fields.

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S (2019) analyzed racial bias in an algorithm used for population health management and showed how a cost-related target distorted identification of health needs. The study demonstrates that statistical accuracy against a chosen target and fairness with respect to the clinical objective can diverge. Model fairness analysis must interrogate what is being predicted and how that target was generated.

The psychological consequences extend beyond numerical allocation. Patients who experience repeated institutional neglect may trust subsequent automated recommendations less; clinicians may take model rankings as official legitimization of unequal attention. These propositions require their own observational and experimental tests. Their plausibility follows from institutional trust research but cannot be equated with the numerical findings of the original algorithmic bias study.

Patients' Trust and Public Acceptance

Nong, P., & Platt, J (2025) analyzed a national survey of US adults concerning trust in health systems' responsible AI use. The published analysis described generally low trust and an association with broader trust in health care institutions. Cross-sectional survey data reveal attitudes at one historical moment; they cannot determine which communication or governance interventions would causally increase justified trust.

Patient acceptance is not the appropriate ultimate endpoint when acceptance would expose people to an unsafe system. A more defensible measure is informed, appropriately conditional acceptance paired with freedom to question output and access qualified care. An individual's preference to consult AI may reflect convenience, lower cost, embarrassment, geographic constraints, or a prior adverse healthcare experience. Research needs to identify these causes rather than presuming that any adoption is evidence of therapeutic benefit.

Trustworthiness is an institutional achievement evidenced by transparent procedures, accountable review, reliable data handling, and response to identified errors. Survey trust scores may improve when those properties are communicated, but communication alone should not substitute for demonstrating them.

Patient-Facing Emergency Advice

Patient-facing systems alter the decision pathway because the user may lack the clinical expertise needed to detect an error. The experimental comparison by Yau, J. Y.-S., Saadat, S., Hsu, E., Murphy, L. S.-L., Roh, J (2024) evaluated responses from four large language model chatbots to selected emergency medicine questions, using clinician assessments across multiple domains. The investigators reported that dangerous information appeared in a nontrivial portion of tested responses and that reading difficulty sometimes exceeded what would be appropriate for the intended public audience.

The study's sample of questions, models, time of testing, grading rubric, and conversational configurations constrain generalization. The resulting percentage must not be treated as the present-day incidence of dangerous advice across all models. Its substantive contribution is the demonstration that polished expression, answer completeness, appropriate triage, and verifiable sourcing are separate characteristics.

Safety evaluations must include omission scenarios: failure to advise urgent assessment when warranted, false reassurance, inappropriate specificity, and ambiguity about whether a response is informational or an individualized medical instruction. These scenarios are evaluated through prespecified expert criteria, not by user satisfaction alone.

Comparing Evidence Across Clinical Tasks

The empirical literature yields a pattern of conditional outcomes. Imaging reader studies may show improved classification within a bounded task. A randomized diagnostic-vignette trial may fail to show benefits from access to an LLM. Another randomized trial may improve management-reasoning scores in simulation while increasing time. Prospective population screening can demonstrate pathway effects under a precise workflow. Independent validation may expose weak performance of a deployed algorithm. These findings describe different systems, tasks, populations, and outcomes and are consistent once the inferential targets are separated.

This comparison supports a research hierarchy with three questions at every level: what was the exact intervention, what was the independently measured endpoint, and what alternative explanation remains plausible? For a technical benchmark, the principal threats include selection of cases and reference standards. For a reader study, they include expertise distribution and simulated context. For a trial, they include contamination, workflow fidelity, and incomplete patient follow-up. For a deployment, they include model drift, changing prevalence, alert saturation, and operational response capacity.

The evidence justifies investment in carefully evaluated human–artificial systems for particular indications. It also justifies withholding broader claims about autonomous diagnosis or universal safety until studies address those stronger propositions directly.

The Inference Gap Between Detection and Health Outcomes

Detection is not equivalent to effective care. More alerts can increase false alarms; higher diagnostic sensitivity may cause unnecessary procedures; improved screen-reading productivity may not translate into improved access if staffing is constrained elsewhere. Health outcomes result from an entire pathway of interventions, decisions, communication, and follow-up.

The inferential gap is most visible when studies report AUROC, accuracy, or benchmark scores as if they established patient survival, reduced disability, or equity. These metrics are valuable measurements, but they require explicit links to patient-relevant endpoints and appropriate causal designs. Vasey, B., Nagendran, M., Campbell, B., et al (2022) describe the need for early live evaluation, while Liu, X., Cruz Rivera, S., Moher, D., Calvert, M. J., & Denni (2020) specify additional reporting expectations for trials involving AI.

Where randomized trials are infeasible, designs such as prospective interrupted time series, staggered rollouts, matched comparisons, and target-trial emulation may improve evidence, provided their confounding assumptions and implementation changes are documented. None is automatically equivalent to randomized evidence, and the study design should be matched to the proposed clinical claim.

Clinical Safety as a Multilayer Property

Clinical safety is achieved through the interaction of technical reliability, cognitive practices, institutional responsibility, and patient protections. Each layer can compensate for failures in another, yet a nominal safeguard is useful only if it operates under the conditions in which harm might occur. A warning instructing users to verify a model does not provide effective verification when the necessary source data are hidden or the workflow allows only seconds for a decision.

A multilayer analysis asks where each potential error originates, how it propagates, who can recognize it, and which controls prevent or reduce its consequences. A data acquisition error may change the input representation. A predictive error may follow from population shift. An interface error may hide uncertainty or emphasize a misleading result. A human decision error may occur through anchoring, omission, or fatigue. An organizational failure may leave a valid alert unanswered. These mechanisms may form a chain, making incident analysis more informative than assigning blame to the last person to act.

The roadmap by Wiens, J., Saria, S., Sendak, M., et al (2019) emphasizes responsible development and translation of machine learning for health care. The actionable research implication is to treat data collection, evaluation, implementation, and monitoring as connected activities. No single validation certificate removes the need to detect unforeseen interactions after deployment.

Intended Use and Regulatory Classification

The intended use of a health-related AI function defines the clinical task, patient population, type of user, operational setting, and nature of the output. A patient information tool, a clinician-facing recommendation system, and an autonomous software action can have sharply different safety implications even when they share a model family.

The Food and Drug Administration (2026) clarifies criteria under which certain clinical decision-support software functions are excluded from the statutory definition of a medical device, while functions meeting device criteria continue under relevant medical-device policies. A general statement that a tool is 'decision support' does not settle classification. Functional criteria, including the intended users and ability to independently review a recommendation's basis, determine the relevant regulatory analysis.

The scientific question is distinct from a regulatory classification. A legally marketed device can still require post-deployment monitoring in a changed institution; a function outside device regulation can still cause psychological or informational harms. A comprehensive research assessment documents the actual product configuration, local deployment, users, and applicable jurisdiction.

Dataset Shift, Monitoring, and Version Control

Model performance depends on a distribution of patients, conditions, measurement processes, and clinical practices. Distribution shift occurs when deployment inputs or relationships between predictors and outcomes change relative to development data. New imaging hardware, a changed electronic record field, altered disease prevalence, shifts in referral pathways, or new patient demographics can degrade performance without a visible software failure.

The clinical discussion by Finlayson, S. G., Subbaswamy, A., Singh, K., Bowers, J., Kup (2021) emphasizes recognizing changes that may impair model reliability in practice. Monitoring should include the population encountered, missingness and data mappings, output distributions, discrimination and calibration where labels permit, important subgroup outcomes, and evidence of harmful recommendations. Outcome delays and incomplete feedback make this a genuine scientific challenge rather than a routine dashboard exercise.

Version control matters for generative and predictive systems alike. The model identifier, relevant prompting configuration, data interface, and decision threshold should be recoverable for any consequential output. Changing a model can invalidate earlier estimates of error rates or reliance calibration. A safer governance model ties validation claims to a specified version and provides a defined process for recalibration, rollback, and reauthorization.

Accountability Across the Decision Chain

The party that develops a model controls its training and some testing, but may not control how an institution integrates it. A healthcare organization controls procurement, access, workflow, and incident response, while clinicians control many patient-specific actions and patients exercise their own agency. Legal responsibility varies by jurisdiction and contractual arrangement; scientific accountability can be analyzed through observable tasks and audit trails.

A defensible accountability map identifies who approves intended use, who assesses local validity, who monitors errors, who communicates material limitations, who reviews disagreement, who can stop the system, and how people affected by a harmful action obtain an explanation and remedy. These are complementary duties and should not be collapsed into an assertion that 'the human remains in control.'

Control must be operational. If a clinician cannot inspect the basis of a recommendation or the institution makes refusal practically impossible, formal responsibility may exceed the person's capacity. Conversely, an ungoverned patient-facing application may escape institutional oversight while significantly affecting care-seeking. Human–artificial health psychology studies how people actually experience and exercise those roles.

Explainability, Contestability, and Practical Review

Explanations can serve different functions: establishing which data informed an output, summarizing model behavior, communicating uncertainty, supporting correction, or justifying a clinical action. These functions should be evaluated separately. Technical feature attributions may describe correlations without identifying causal clinical mechanisms. A coherent generated rationale may rationalize an answer rather than faithfully disclose its derivation.

Ribeiro, M. T., Singh, S., & Guestrin, C (2016) contributed a general method for explaining classifiers, but healthcare requires independent checks of fidelity and decision impact. An effective review interface gives the user access to relevant underlying observations, permits correction of erroneous data, and shows when the model's intended use does not match the current case.

Contestability adds a social and institutional dimension: someone affected by a recommendation must have a viable route to challenge it. This includes language access, an accountable reviewer, documented disagreement, and the possibility of revising a decision. Research can measure how often discrepancies are detected, how quickly corrections occur, and whether access to challenge differs across populations.

Risk Distribution, Equity, and Safety by Subgroup

An average performance statistic can conceal harms concentrated in a clinically meaningful subgroup. These differences may reflect training representation, data quality, referral pathways, thresholds, physiological variation, measurement bias, or institutional inequity. Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S (2019) demonstrates that even the outcome used to train a model may be systematically misaligned with care needs.

Safety assessment should prespecify clinically relevant subgroups and report event rates, denominator sizes, uncertainty, and the potential clinical consequences of observed differences. An apparent lack of disparity in a small sample is insufficient evidence of equality. Comparative evaluation should ask whether unequal access to follow-up, digital devices, specialists, and interpreters changes the practical consequences of the same algorithmic output.

A model may also shift burdens from professionals to patients: asking users to independently verify advice can disadvantage those with less time, money, language access, or medical knowledge. The ethical and psychological evaluation therefore concerns effective capacity to act, not merely formal availability of information.

Adverse Events, Near Misses, and Learning Systems

Clinical safety science learns from both actual injury and near misses. If a physician detects a wrong AI recommendation before action, the patient may be unharmed, but the event contains evidence about a potentially hazardous failure mode. Organizations need reporting channels that allow such events to be documented without suppressing information through professional fear or organizational reputation management.

An adverse-event record should describe the system version, relevant input conditions, output, interface context, user response, intended and actual action, patient outcome where applicable, and immediate mitigation. Root-cause inquiry can then distinguish technical defects from misunderstanding, poorly specified use, inadequate staffing, or missing escalation policies.

Near-miss data have limitations: reports are selected by what users notice, and events without obvious consequences may be underreported. They should complement structured audits of sampled outputs and prospective safety metrics. The objective is an adaptive governance process in which errors lead to documented improvement and re-evaluation.

Performance, Calibration, and Decision-Analytic Thresholds

Model discrimination describes how well scores distinguish cases with and without an outcome. Calibration describes the relation between predicted probabilities and observed outcome frequencies. Decision-analytic evaluation additionally considers the consequences of acting at a threshold. A highly discriminating model can still yield little net benefit if the recommended action is costly, the relevant event is rare, or the patient population differs from the validation sample.

The relationship is visible in Wong, A., Otles, E., Donnelly, J. P., Krumm, A., McCullough (2021), whose independent assessment combined discrimination, calibration, and comparison with existing clinical practice. A model that predicts events at an apparently useful level may still add little to the information already available to clinicians. Evaluations should predefine what decision the output is intended to change and assess alternatives such as ordinary clinical assessment, earlier human review, or a different alert threshold.

Sensitivity and specificity should be reported alongside prevalence, positive predictive value, confidence intervals, subgroup analyses, and patient-relevant costs of false positives and negatives. For many clinical uses, calibration is not simply a technical virtue: it directly affects whether a patient is reassured, monitored, referred, or subjected to intervention.

Organizational Implementation and Workforce Learning

The implementation of an AI system changes work distribution. Clinicians may spend less time on an initial classification but more time verifying inputs, resolving discordance, explaining a recommendation, or correcting records. Workload may shift from physicians to nurses, from specialists to generalists, or from institutions to patients. These effects are central to the psychological reality of the technology and require explicit measurement.

Clinical workflow trials should include training, adoption, fidelity, staffing, intervention reach, time cost, alert burden, turnover, and sustainability. A system evaluated during a carefully supervised trial may function differently after the research team departs. Longitudinal monitoring should address changing usage and whether the intended clinical response remains possible.

Institutional procurement should therefore be treated as a claim requiring evidence: the technology is expected to improve a defined decision or outcome for a defined population under defined conditions. Results should be independently audited when possible and renewed after material changes.

Global and Cross-Cultural Considerations

Models developed in high-resource health systems may operate on different clinical data, infrastructure, language conventions, disease prevalence, and access pathways when moved elsewhere. The same warning or risk score can induce different actions where confirmatory testing, emergency transportation, specialist care, or reliable connectivity are unavailable.

Cross-cultural research should examine shared decision norms, perceived institutional legitimacy, interpretation of risk, and acceptable data practices. Translation should be validated for medical meaning and pragmatic clarity, including dialect, health literacy, and context of use. Accessibility is a safety parameter because a correct output cannot reliably improve a decision if the intended user cannot comprehend it or act upon it.

Local evidence is essential. A favorable study conducted in one jurisdiction cannot substitute for evaluation of the model, workflow, and response capacity within a different care system. International comparisons should document these contextual changes rather than treating country as a simple subgroup label.

Reporting Standards and Evidence Transparency

Reporting standards are tools for preserving the interpretability of empirical evidence. Collins, G. S., Moons, K. G. M., Dhiman, P., Riley, R. D., B (2024) addresses prediction-model studies; Sounderajah, V., Guni, A., Liu, X., et al (2025) addresses diagnostic-accuracy studies; Liu, X., Cruz Rivera, S., Moher, D., Calvert, M. J., & Denni (2020) addresses trials; and Vasey, B., Nagendran, M., Campbell, B., et al (2022) addresses early live clinical evaluation.

A report should document exact model identification, intended users, participant characteristics, setting, data source, reference standard, user instructions, human–machine interaction, failure cases, statistical uncertainty, and implementation departures. Where proprietary barriers restrict independent replication, the limitation should be visible rather than hidden behind technical summaries.

The National Institute of Standards and Technology (2023), initially released in 2023 and under revision in 2026, provides a broader voluntary risk-management structure. Its value here is conceptual and procedural; conformity claims require concrete implementation evidence and do not replace clinical studies. The transparency objective is to let another investigator understand what was evaluated, reproduce relevant checks, and distinguish established performance from conjecture.

The Importance of Clinical Review

A scientific monograph about the psychology of decision support is different from an approved clinical practice guideline. Its conclusions may inform research and governance, while treatment decisions remain subject to qualified professional practice and current jurisdiction-specific standards. Any future edition that recommends a concrete diagnostic, triage, or treatment protocol should undergo review by appropriately qualified clinicians and verification against current regulatory and clinical guidance.

The need for external clinical review is especially strong when a manuscript interprets intervention outcomes or might influence immediate patient decisions. Editorial attribution should identify such reviewers only after review actually occurs. This manuscript makes no individualized therapeutic or prescribing recommendations; publication preparation should nevertheless include specialist scrutiny of contested clinical claims and corrected source records before final release.

A Measurement Architecture for Human–Artificial Health Interaction

A rigorous measurement program defines the unit, task, population, and counterfactual before selecting outcomes. At the system level, it measures discrimination, calibration, missing-data sensitivity, robustness, and drift. At the human level, it measures initial judgment, advice exposure, perceived credibility, comprehension, decision changes, and confidence calibration. At the team level, it measures final accuracy, error interception, appropriate escalation, time, and workflow burden. At the patient level, it measures benefit, harm, autonomy, understanding, access, and follow-up.

Measures should be linked through timestamps and anonymized or appropriately protected identifiers, allowing investigators to reconstruct how an output affected a decision. An outcome measured only at the model level cannot establish what a user actually saw. Likewise, an end-of-visit satisfaction survey cannot reveal whether a missed alarm altered clinical action. Longitudinal studies need to account for repeated observations nested in clinicians, patients, shifts, and institutions.

Reliance calibration can be summarized by the conditional probability of following advice when it is correct and when it is incorrect, alongside the counterfactual performance without advice. This should be interpreted with case difficulty and the independence of expert adjudication. A large gap between acceptance of correct and incorrect advice can be desirable, but any scalar summary also needs to reflect differing consequences of errors. High-stakes decisions require explicit severity weighting and uncertainty analysis.

Experimental Design for Causal Mechanisms

Randomized controlled experiments remain valuable for isolating human factors in a way that deployment observation cannot. A factorial study might manipulate advice accuracy, explanation format, timing, confidence presentation, and initial judgment requirement. It would randomly assign these features while using prespecified clinical vignettes and blinded grading. The resulting contrasts could test whether an apparent explanation benefit is actually driven by increased acceptance of every recommendation.

Such work must avoid selecting only high-performing AI cases. It should include wrong recommendations, ambiguous inputs, rare clinical presentations, and competing evidence. Ethics require that intentionally erroneous advice remain within simulation or otherwise controlled procedures that protect real patients. Participants should be stratified by expertise and prior tool exposure where justified, with sample sizes designed to estimate interaction effects.

Trial analysis should account for repeated ratings by clinicians and correlations among cases. Confidence intervals, uncertainty in the reference standard, and prespecified primary outcomes matter more than a long list of exploratory differences. Where the system uses a generative model, the prompt, version, tool access, randomness controls, and output provenance should be retained as part of the intervention definition.

Real-World Evaluation and Transportability

Clinical benefit requires observation of real decision sequences and their outcomes. Pragmatic randomization at the patient, clinician, shift, or institution level can test the effect of deploying a complete care pathway. Cluster trials must account for intracluster correlation and contamination from clinicians sharing strategies. Stepped-wedge designs need careful control for secular change. Nonrandomized evaluations require credible comparison groups and sensitivity analysis for confounding.

Transportability asks whether an effect established in one setting would persist in another. Important effect modifiers include baseline clinical skill, case prevalence, local workflow, laboratory availability, language, staffing, patient demographics, institutional trust, and model version. The MASAI reports02464-X) are especially valuable because they examine a defined real-world screening system. Their generalization remains conditioned on comparable practices and population characteristics.

Before broad adoption, research should replicate both technical and interaction findings across independent institutions. This entails documenting cases where performance degrades, not only successful replications. A favorable external AUROC result does not verify the training, user behavior, or downstream clinical consequences at a new site.

Longitudinal Evaluation of Learning and Dependency

Repeated interaction alters the human partner. A clinician may learn to distrust recurrent false positives, become adept at checking questionable outputs, or lose proficiency in an infrequently practiced task. A patient may use an assistant to prepare questions and gain understanding, or begin treating a convenient explanation as a substitute for clinical assessment. These possibilities require longitudinal observation rather than predictions drawn from initial preference.

Useful designs include repeated-measures studies, log-based analysis with patient privacy protections, randomized withdrawals where ethically justified, and comparison of performance after periods without access. Researchers should test independent competence, response to new model versions, time needed to recognize errors, willingness to seek human consultation, and patient experiences of control.

Long-term effects may differ by task. Cognitive offloading of administrative recall may free attention without impairing medical judgment, while delegation of differential diagnosis may alter the habits used to detect atypical cases. These are separate hypotheses; neither positive nor negative generalization is warranted without direct evidence.

Human–Artificial Communication as a Safety-Critical Interface

Conversational systems can provide an interactive route to medical information, but the conversation itself is a sequence of decisions. Each question changes the evidence the user provides, and each response can alter expectations, reassurance, worry, or help-seeking. A system that misunderstands a symptom or fails to ask a discriminating question may produce a plausible but hazardous conclusion.

A research design for conversational safety can score the completeness of information gathering, verification of uncertain symptoms, acknowledgment of limits, appropriate referral to qualified care, source fidelity, and the avoidance of contradictory or unsupported advice. Important test sets include rare presentations, incomplete histories, multilingual phrasing, crisis disclosure, and users who explicitly request certainty.

Human evaluations should distinguish clinical accuracy from friendliness, reading level, perceived empathy, and conversational coherence. Yau, J. Y.-S., Saadat, S., Hsu, E., Murphy, L. S.-L., Roh, J (2024) illustrate the need to analyze dangerous content and comprehension directly. Interface improvements are meaningful when they demonstrably reduce error rates or improve understanding without introducing new problems.

Privacy-Preserving Research and Reproducibility

The data needed for understanding human–artificial interaction may include sensitive clinical narratives, interaction transcripts, or decision logs. Research design should minimize collection, de-identify when feasible, restrict access according to legitimate purpose, and document retention and audit procedures. Strong privacy safeguards can complicate independent verification, requiring careful approaches to reproducible aggregate statistics, synthetic test cases, and approved secure analysis environments.

Reproducibility in this field involves both computational and behavioral elements. Reproducing a model output may be impossible if the model has changed, and reproducing the effect on clinicians may require similar instructions, expertise, workload, and local norms. Reports should distinguish exact repeatability of a frozen test from conceptual replication across new populations and models.

Conflicts of interest matter when researchers evaluate a product they developed or a healthcare organization has purchased. Disclosure of affiliations, funding, data access, and control over publication is part of evaluating evidence quality. Independent outcome adjudication and preregistration can reduce selective interpretation.

Critical Discussion of Competing Explanations

Many observed human–AI outcomes permit multiple explanations. Better decisions after assistance may arise from the model identifying overlooked evidence, from longer time spent on the task, from differential expertise, or from features of the experimental rubric. Worse decisions may arise from incorrect advice, a poor interface, case selection, lack of time for verification, or baseline difficulty. An explicit theory of interaction must provide experiments that distinguish among these mechanisms.

The comparison of Goh, E., Gallo, R., Hom, J., Strong, E., Weng, Y., Kerman, H (2024) with Goh, E., Gallo, R. J., Strong, E., et al (2025) illustrates this need. Their results differ, but so do the decision tasks and research designs. It would be inaccurate to average their point estimates as though they measured a homogeneous intervention. The appropriate inference is heterogeneity requiring more precise classification.

One unresolved question concerns the role of explanation. A detailed rationale may help a clinician inspect evidence, but it may also increase persuasive force when the conclusion is wrong. Another concerns authority cues: displaying a medical institutional logo may change compliance independently of correctness. A third concerns the interaction of clinical scarcity and algorithmic guidance: support may increase access where expertise is limited, while simultaneously concentrating risks among people least able to contest errors. Each is a tractable empirical program, not an established universal effect.

Evidence Synthesis Across Levels

The strongest defensible synthesis is conditional. Artificial systems can improve well-specified health decisions when their outputs are accurate, the interaction enables effective verification, the workflow supports appropriate response, and the deployment population resembles the evidence base. They can also misdirect clinicians and patients, propagate biases from proxy targets, create alert burden, or fail after dataset shift. The published literature contains evidence for both possibilities.

The field's primary analytical mistake is to conflate levels of evidence. An image classification benchmark does not measure collaborative accuracy; collaborative accuracy in a simulated task does not measure patient benefit; a favorable live trial in one institution does not prove indefinite reliability after software updates. A model's smooth conversational style does not establish safety. A policy statement does not establish implementation. Keeping these levels distinct produces a more durable science of human–artificial health systems.

The historical sequence of automation research, clinical machine-learning development, human-advice experiments, and prospective pathway trials points toward a unified research standard: every consequential artificial recommendation must be studied through the decisions it changes and the outcomes those changes produce. The independent evidence required rises with the severity and irreversibility of potential harm.

Priority Research Problems

Several unresolved problems have direct methodological implications. The first is a generalizable measure of appropriate reliance that accounts for differing costs of errors and distinguishes independent human agreement from AI-induced compliance. The second is reliable uncertainty communication that improves decisions when a system is wrong as well as when it is correct. The third is estimation of how model updates, interface changes, and clinician learning alter interaction over time.

The fourth is the measurement of differential harm across patient populations, including populations poorly represented in training data or unable to reach alternative care. The fifth is attribution of safety failures across developers, deployers, professionals, and organizations while maintaining effective patient remedies. The sixth is comparison of clinician-facing, caregiver-facing, and patient-facing uses under clearly defined authorization and escalation arrangements.

The seventh problem concerns the safety of persistent conversational engagement: how repeated exchanges change disclosure, belief, help-seeking, and dependence, and when those changes improve or worsen health. The eighth is cross-cultural transportability across languages, levels of care, and institutions. Finally, researchers need methods that connect technical version histories to patient-relevant outcomes while protecting sensitive data.

Implications for Health Psychology and the Artificial Era

Health psychology studies beliefs, coping, perception of risk, self-regulation, communication, and the behavioral organization of care. Artificial participation introduces a new source of information and influence into those processes. Its effects cannot be reduced to model architecture, because they are produced through decisions and relationships embedded in healthcare systems.

The historical movement from Homo to Artificial makes this domain especially consequential. Artificial advisers can become enduring elements of the environments through which people understand symptoms, negotiate medical uncertainty, and coordinate care. Their empirical evaluation must therefore attend simultaneously to cognition, communication, institutional power, and the changing capabilities and limits of computational systems.

This research perspective establishes a clear division of labor for future scholarship. Technical studies identify model characteristics; psychological studies identify how people interpret and act upon outputs; clinical trials evaluate consequences for care; and governance research examines whether the resulting institutions remain accountable and corrigible. These investigations converge on the same scientific object: health-related decisions produced through human–artificial interaction.

Conclusion

Human–artificial health interaction is an evidence-bearing decision process joining artificial outputs, human judgment, institutional procedures, and patient consequences. Its central constructs are reliance calibration, trust, decision delegation, attention, comprehension, accountability, and safety. Its principal research methods are controlled advice experiments, reader studies, prospective and pragmatic trials, independent external validation, psychometric and qualitative investigation, longitudinal monitoring, and interdisciplinary safety analysis.

The evidence establishes that benefits are possible in specific tasks and that important risks arise from misleading advice, poorly selected prediction targets, weak external performance, alert burden, opaque delegation, and inadequate governance. The randomized physician studies, collaborative imaging experiments, MASAI screening program, and independent sepsis model evaluation provide complementary observations across different levels of inference. Their findings support scientifically bounded conclusions rather than a single generalized verdict on healthcare AI.

A mature research program evaluates what the artificial system does, what a human understands and decides because of it, and what happens to patients as a result. It records the limits of each claim, verifies performance under actual conditions, and preserves routes for correction. Such inquiry supplies the empirical foundation for accountable human–artificial health relations during the Artificial Era.

References

  • Collins, G. S., Moons, K. G. M., Dhiman, P., Riley, R. D., Beam, A. L., Van Calster, B., Ghassemi, M., et al. (2024). TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ, 385, e078378. https://doi.org/10.1136/bmj-2023-078378
  • Finlayson, S. G., Subbaswamy, A., Singh, K., Bowers, J., Kupke, A., Zittrain, J., Kohane, I. S., & Saria, S. (2021). The clinician and dataset shift in artificial intelligence. New England Journal of Medicine, 385(3), 283–286. https://doi.org/10.1056/NEJMc2104626
  • Food and Drug Administration. (2026, January 29). Clinical decision support software: Guidance for industry and Food and Drug Administration staff. U.S. Department of Health and Human Services. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software
  • Gaube, S., Suresh, H., Raue, M., Merritt, A., Berkowitz, S. J., Lermer, E., Coughlin, J. F., Guttag, J. V., Colak, E., & Ghassemi, M. (2021). Do as AI say: Susceptibility in deployment of clinical decision-aids. npj Digital Medicine, 4, Article 31. https://doi.org/10.1038/s41746-021-00385-9
  • Goh, E., Gallo, R., Hom, J., Strong, E., Weng, Y., Kerman, H., Cool, J. A., Kanjee, Z., Parsons, A. S., Ahuja, N., Horvitz, E., Yang, D., Milstein, A., Olson, A. P. J., Rodman, A., & Chen, J. H. (2024). Large language model influence on diagnostic reasoning: A randomized clinical trial. JAMA Network Open, 7(10), e2440969. https://doi.org/10.1001/jamanetworkopen.2024.40969
  • Goh, E., Gallo, R. J., Strong, E., et al. (2025). GPT-4 assistance for improvement of physician performance on patient care tasks: A randomized controlled trial. Nature Medicine, 31, 1233–1238. https://doi.org/10.1038/s41591-024-03456-y
  • Gommers, J., Hernström, V., Josefsson, V., Sartor, H., Schmidt, D., Hjelmgren, A., Larsson, A.-M., Hofvind, S., Andersson, I., Rosso, A., Hagberg, O., & Lång, K. (2026). Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study: A randomised, controlled, non-inferiority, single-blinded, population-based, screening-accuracy trial. The Lancet, 407(10527), 505–514. https://doi.org/10.1016/S0140-6736(25)02464-X
  • Hernström, V., Josefsson, V., Sartor, H., Schmidt, D., Larsson, A.-M., Hofvind, S., Andersson, I., Rosso, A., Hagberg, O., & Lång, K. (2025). Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI): A randomised, controlled, parallel-group, non-inferiority, single-blinded, screening accuracy study. The Lancet Digital Health, 7(3), e175–e183. https://doi.org/10.1016/S2589-7500(24)00267-X
  • Lång, K., Josefsson, V., Larsson, A.-M., et al. (2023). Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): A clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study. The Lancet Oncology, 24(8), 936–944. https://doi.org/10.1016/S1470-2045(23)00298-X
  • Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
  • Liu, X., Cruz Rivera, S., Moher, D., Calvert, M. J., & Denniston, A. K. (2020). Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: The CONSORT-AI extension. Nature Medicine, 26, 1364–1374. https://doi.org/10.1038/s41591-020-1034-x
  • Liu, X., Faes, L., Kale, A. U., Wagner, S. K., Fu, D. J., Bruynseels, A., Mahendiran, T., Moraes, G., Shamdas, M., Kern, C., Ledsam, J. R., Schmid, M. K., Balaskas, K., Topol, E. J., Bachmann, L. M., Keane, P. A., & Denniston, A. K. (2019). A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: A systematic review and meta-analysis. The Lancet Digital Health, 1(6), e271–e297. https://doi.org/10.1016/S2589-7500(19)30123-2
  • McKinney, S. M., Sieniek, M., Godbole, V., et al. (2020). International evaluation of an AI system for breast cancer screening. Nature, 577, 89–94. https://doi.org/10.1038/s41586-019-1799-6
  • National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). U.S. Department of Commerce. https://www.nist.gov/itl/ai-risk-management-framework
  • Nong, P., & Platt, J. (2025). Patients’ trust in health systems to use artificial intelligence. JAMA Network Open, 8(2), e2460628. https://doi.org/10.1001/jamanetworkopen.2024.60628
  • Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453. https://doi.org/10.1126/science.aax2342
  • Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253. https://doi.org/10.1518/001872097778543886
  • Rajkomar, A., Dean, J., & Kohane, I. (2019). Machine learning in medicine. New England Journal of Medicine, 380(14), 1347–1358. https://doi.org/10.1056/NEJMra1814259
  • Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). "Why should I trust you?": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 1135–1144). Association for Computing Machinery. https://doi.org/10.1145/2939672.2939778
  • Skitka, L. J., Mosier, K. L., & Burdick, M. (1999). Does automation bias decision-making? International Journal of Human-Computer Studies, 51(5), 991–1006. https://doi.org/10.1006/ijhc.1999.0252
  • Sounderajah, V., Guni, A., Liu, X., et al. (2025). The STARD-AI reporting guideline for diagnostic accuracy studies using artificial intelligence. Nature Medicine, 31, 3283–3289. https://doi.org/10.1038/s41591-025-03953-8
  • Tschandl, P., Rinner, C., Apalla, Z., et al. (2020). Human–computer collaboration for skin cancer recognition. Nature Medicine, 26, 1229–1234. https://doi.org/10.1038/s41591-020-0942-0
  • Vasey, B., Nagendran, M., Campbell, B., et al. (2022). Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nature Medicine, 28, 924–933. https://doi.org/10.1038/s41591-022-01772-9
  • Wiens, J., Saria, S., Sendak, M., et al. (2019). Do no harm: A roadmap for responsible machine learning for health care. Nature Medicine, 25, 1337–1340. https://doi.org/10.1038/s41591-019-0548-6
  • Wong, A., Otles, E., Donnelly, J. P., Krumm, A., McCullough, J., DeTroyer-Cooley, O., Pestrue, J., Phillips, M., Konye, J., Penoza, C., Ghous, M., & Singh, K. (2021). External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 181(8), 1065–1070. https://doi.org/10.1001/jamainternmed.2021.2626
  • World Health Organization. (2021). Ethics and governance of artificial intelligence for health: WHO guidance. https://www.who.int/publications/i/item/9789240029200
  • World Health Organization. (2025). Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. https://www.who.int/publications/i/item/9789240084759
  • Yau, J. Y.-S., Saadat, S., Hsu, E., Murphy, L. S.-L., Roh, J. S., Suchard, J., Tapia, A., Wiechmann, W., & Langdorf, M. I. (2024). Accuracy of prospective assessments of 4 large language model chatbot responses to patient questions about emergency care: Experimental comparative study. Journal of Medical Internet Research, 26, e60291. https://doi.org/10.2196/60291