Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Human–Artificial Therapeutic Interaction and Clinical Accountability

Abstract. Human–artificial therapeutic interaction is a psychologically consequential sequence in which a person seeks mental-health-related support from an artificial conversational system and its replies acquire therapeutic meaning, influence behavior, or enter a clinical pathway. This fundamental research monograph integrates historical psychotherapy theory, social responses to computers, anthropomorphism, experimental comparisons of perceived empathy, randomized trials of structured and generative chatbots, meta-analytic outcomes, clinical crisis-safety evaluations, human-in-the-loop implementation studies, and professional and regulatory guidance. Using a critical narrative synthesis of original research and documented methodological standards through November 3, 2026, the analysis distinguishes perceived empathy and therapeutic role attribution from demonstrated treatment efficacy and clinical responsibility. Evidence supports symptom benefits from selected systems under specified trial conditions, while crisis-response studies document model- and interface-specific failure modes and the importance of executable human escalation. The synthesis organizes accountability around truthful disclosure, validated intended use, clinical governance, privacy, monitoring, version control, professional judgment, and completed handoffs. Persistent research problems include causal mechanisms, long-term outcomes, clinical comparators, rare adverse events, equitable generalization, dependence, and the reproducibility of oversight across changing artificial systems.

Keywords: human–artificial therapeutic interaction; clinical accountability; AI psychotherapy; therapeutic alliance; artificial empathy; informed consent; trust calibration; clinical escalation; AI safety; human-in-the-loop; psychological outcomes; Artificial Era

Introduction to Human–Artificial Therapeutic Interaction

Human–artificial therapeutic interaction is a psychologically consequential sequence of exchanges in which a person seeks or receives mental-health-related support from an artificial conversational system, and the system's responses acquire a therapeutic function through their content, interpersonal form, perceived authority, or connection with a clinical service. Its observable units include disclosure, system response, user interpretation, subsequent behavior, symptom course, risk identification, and the distribution of responsibility for action. The term covers a continuum from informational support and structured self-management to clinically supervised digital intervention. Each position on that continuum has a distinct evidentiary burden. A persuasive exchange can produce emotional relief without establishing that the system delivers a validated treatment, and a system embedded in a clinical workflow can influence care even when it has no independent professional status.

Clinical accountability is the traceable allocation of duties to design, validate, disclose, monitor, intervene, correct, and remedy within this interaction. It concerns the organizations and people who create, deploy, prescribe, recommend, supervise, and use the system; the circumstances under which a therapeutic representation is made; and the availability of effective action when its performance becomes unsafe. Accountability has empirical dimensions—whether failures are detected, human review actually occurs, and patients reach suitable care—and institutional dimensions involving consent, professional duties, safety engineering, data governance, and applicable law. The present research examines both dimensions within one scientific object. The adjacent subject of digital clinical systems concerns diagnostics, triage infrastructure, and wider service pathways; the adjacent subject of AI companionship concerns enduring nonclinical relationships. The present dossier asks how a bounded or continuing therapeutic exchange acquires clinical significance and how responsibility follows from that significance.

The central scientific questions are whether apparent empathy and therapeutic alliance can be measured independently from clinical efficacy; when transparent machine disclosure affects perceived trust and actual use; which interventions have demonstrated symptom benefits under experimental conditions; how unreliable crisis response arises; whether human oversight changes consequential outcomes; and what longitudinal and comparative research is required for accountable deployment. The synthesis addresses American clinical and research practice while treating international studies and standards as evidence relevant to generalization. Within the Artificial Era, artificial systems are studied as participants in a functional psychological exchange. Empirical claims about their capacities depend on measured system behavior and user outcomes, rather than on the philosophical identity ascribed to the system.

Research Method and Evidentiary Architecture

This monograph is a structured critical narrative synthesis of published original trials, comparative experiments, systematic reviews, clinical implementation studies, safety benchmarks, methodological research, and institutional guidance available through November 3, 2026. Studies were selected for the extent to which they identify an interactional construct, establish a comparison or measurable outcome, report human oversight or crisis behavior, or define a consequential ethical or clinical obligation. The review distinguishes direct experimental evidence, prospective observational evidence, laboratory and simulation benchmarks, systematic research synthesis, and normative guidance. It is an interpretive scholarly monograph rather than a protocol-registered systematic review, exhaustive database census, new meta-analysis, or original patient investigation. Its conclusions are correspondingly organized by strength and domain of evidence rather than numerical aggregation across incomparable designs.

A study's relevance does not make its findings interchangeable with those of another design. An experiment measuring whether readers judge a reply empathetic tests perception under controlled stimulus conditions. A randomized trial measuring symptom change after use of a named chatbot tests a specific product, population, comparator, and period. A safety benchmark presenting crisis prompts to general-purpose models measures the reliability of their responses to those prompts. An implementation study measuring contact after a flag tests the operations of a particular escalation pathway. None alone establishes the clinical performance of the other. Following Liu (2020) and Cruz Rivera (2020), complete appraisal asks how the model, interface, human interaction, version, integration, errors, and protocol are specified.

The framework has four linked levels of inference. Construct validity evaluates what an instrument or judgment captures. Causal validity evaluates whether assigning the system rather than a credible alternative changes an outcome. Transportability evaluates whether performance persists across populations, settings, languages, care systems, versions, and time. Accountability validity evaluates whether declared safeguards correspond to observable monitoring, reachable clinicians, timely decisions, and documented remediation. The later chapters return to these levels to prevent the common progression from credible conversation to unverified claims of safe psychotherapy.

The Historical Formation of the Problem

The scientific history begins with psychotherapy research on the interpersonal conditions of change and with computer-mediated social response. In Rogers (1957), congruence, empathy, and unconditional positive regard belonged to a theory of therapeutic personality change within a human relationship. Bordin (1979) described the working alliance through agreement on goals, agreement on tasks, and the bond connecting participants. These formulations shaped the language through which later investigators would interpret relational experiences with software. Their historical contribution is an explanatory vocabulary; whether an artificial interaction reproduces clinically relevant causal mechanisms is a separate testable proposition.

Experiments on social responses to computers introduced a second research lineage. Nass (2000) documented circumstances in which users applied learned social conventions to computers, including politeness and reciprocity. Epley (2007) organized anthropomorphism around elicited human knowledge, motivation to understand an agent, and sociality motivation. These accounts explain how users may experience a system as understanding or attentive based on conversational cues. The experience may have real psychological consequences regardless of whether the system is licensed, embodied, conscious, or equipped with any particular internal mechanism.

The applied lineage passed through rule-based and retrieval-oriented agents, internet interventions, and the later emergence of large language models. Laranjo (2018) documented diverse healthcare conversational agents and substantial variation in evaluation designs. Fitzpatrick (2017) provided a landmark short randomized trial of an automated cognitive behavioral conversational agent for young adults. Generative systems subsequently expanded the range of conversational topics and interpersonal styles. This shift altered the failure space: unbounded replies can invite disclosure and respond flexibly, while making the boundaries of the intended intervention harder for users and evaluators to identify.

Defining Therapeutic Role, Clinical Care, and Support

A therapeutic role is an attributed or institutionally assigned function through which a conversational partner appears to help a person interpret distress, regulate emotion, deliberate about behavior, or pursue psychological change. Attribution can arise from a product's explicit claims, a clinician's referral, conversational content, the interface's persona, or the user's own interpretation. Roles can be described empirically by the frequency of mental-health disclosures, the kinds of guidance given, whether treatment instructions occur, and whether users report reliance in decisions. A therapeutic role is therefore observable without presupposing a credential, clinical effectiveness, or a reciprocal human-style bond.

Clinical care is distinguished by a service relationship in which assessment, treatment planning, risk management, documentation, and continuity are governed through identifiable professional and institutional processes. A clinician may use an artificial system to prepare materials, explain psychoeducation, or deliver supervised exercises, but clinical responsibility remains governed by the actual division of tasks and jurisdictional rules. Conversely, a consumer chatbot can influence consequential choices outside formal care when it encourages medication changes, minimizes a warning sign, or persuades a user to avoid human support. Its lack of formal clinical designation does not erase the possibility of consequential harm; responsibility must be examined through intended use, foreseeable use, representation, design choices, and the applicable regulatory and legal framework.

At least five domains must be separately measured: informational accuracy; emotional experience and perceived empathy; evidence-based therapeutic operation; clinically important outcome; and safety performance under acute or ambiguous risk. A response can succeed in one and fail in another. In Ayers (2023), licensed professionals favored chatbot replies to public patient questions on perceived quality and empathy. Those data concern ratings of written responses to a forum, not a longitudinal therapeutic relationship, independently verified medical correctness, or patient-level mental-health improvement. Treating the distinction as central allows favorable communication findings to coexist with rigorous clinical evaluation.

The Psychological Organization of a Therapeutic Exchange

A therapeutic exchange can be analyzed as a temporal sequence: a human enters with a need or uncertainty; the artificial system generates a response under a set of model and interface constraints; the person interprets the response as informational, empathic, authoritative, or relational; the response changes affect, appraisal, or intended action; and subsequent contact modifies expectations. The unit of analysis expands from one message to the encounter and then to a series of encounters. This expansion matters because a convincing response may confer immediate comfort while delayed outcomes include help seeking, changed treatment adherence, increased dependency, or recognition of a previously overlooked risk.

A useful research design measures each link independently. Observers can code whether an answer mirrors emotional language, corrects a cognitive distortion, offers behaviorally specified support, or shifts from general support to individualized clinical instruction. Participants can rate perceived understanding, trust, autonomy, and disclosure willingness. Behavioral logs can capture verification, clinician contact, adherence to advice, and termination of use. Validated symptom scales and prespecified adverse-event measures can assess outcomes. The analysis can then examine which intermediate processes precede improvement and whether they are attributable to the intervention rather than expectancy, baseline differences, or regression to the mean.

Human–artificial interaction also has a system-level organization. The model generates candidate text; an interface can apply routing, memory, safety filters, and contextual instructions; a monitoring service may flag risk; a professional team may be called; and the provider may revise policies after observing an incident. Cantor (2026) found that suicide-related responses varied between direct model APIs and consumer-facing interfaces. The result establishes why an apparently identical model name is an insufficient description of the clinically relevant intervention. The deployed interaction is a configuration of technologies and governance practices whose components can meaningfully change behavior.

The Research Boundary Between Human and Artificial Agency

The Artificial Era expands the number and types of systems participating in psychologically consequential interaction. Artificial responsiveness, continuity, and goal-directed selection may be studied through records, controlled prompts, model updates, and behavioral reliability tests. Human experience remains observable through ratings, disclosure, behavior, symptom reports, and care trajectories. Functional coordination between both sides is a legitimate psychological object even when competing philosophical accounts of artificial subjectivity remain unsettled. The present monograph uses the term artificial participant to specify the operational role of an artificial system in the exchange.

Agency in an accountable clinical workflow has a second meaning: the authority to set goals, authorize interventions, access protected information, override a recommendation, and respond to an emergency. A generative model may produce sophisticated recommendations while the institutional authority to act belongs to a licensed practitioner, deploying organization, patient, or other legally specified actor. Distinguishing these dimensions reveals precisely where responsibility could diffuse. A user may experience the software as the therapeutic actor while the organization assumes its outputs are merely informational and the clinician assumes the product has already validated them.

The scientific consequence is a requirement for separate descriptions of observed behavior, user interpretation, and delegated authority. Such descriptions support experimental tests of therapeutic role attribution and precise accountability audits. They also avoid importing claims from general AI companionship, where friendship and attachment are the primary outcomes, or from digital psychiatry, where predictive screening and service architecture are the primary objects. Human–artificial therapeutic interaction stands at their intersection as a distinct field organized around mental-health-relevant conversational influence and responsibility for its consequences.

Therapeutic Alliance and Its Limits of Transfer

Therapeutic alliance research provides an indispensable framework for describing cooperation, but its constructs require explicit adaptation before they can measure artificial interaction. Agreement on goals can be tested when a user selects a target, the system restates it accurately, and later responses remain consistent with that target. Agreement on tasks can be measured when a person understands an exercise, endorses its rationale, and completes it. The bond dimension poses a more complex measurement problem because a questionnaire developed for human clinicians may confound warmth of text, perceived personhood, trust, and actual interpersonal obligation. Researchers must specify whether they measure felt bond, response accuracy, conversational contingencies, or an institutionally supported therapeutic relationship.

The reference literature describes a strong association in human psychotherapy. Flückiger (2018) synthesized 295 independent studies of more than 30,000 patients, identifying an overall alliance–outcome correlation around .278 in face-to-face psychotherapy. Flückiger (2020) examined adjustment for patient characteristics and treatment processes, reporting associations in the range of r=.23–.31. Those estimates do not transfer as effect sizes to chatbot interaction. They provide a rationale for measuring alliance-like processes and a warning that even repeatedly observed associations may have bidirectional and residual-confounding explanations. Early symptom improvement can itself improve an alliance rating.

Teletherapy offers an intermediate comparison. Aafjes-van Doorn (2024) found a small association between measured alliance and outcomes across 34 independent teletherapy samples. This concerns psychotherapy mediated by technology with a human therapist, not an artificial clinician. It is useful for separating the medium of interaction from the identity and responsibilities of the clinician. Future research on artificial interlocutors should use factorial designs in which medium, human supervision, model persona, disclosure, and treatment program are independently varied.

Apparent Empathy, Validation, and Clinical Accuracy

Empathy has at least three distinguishable operational forms in this domain. Behavioral empathy denotes linguistic actions such as identifying emotional content, acknowledging distress, and responding contingently to the user's concern. Perceived empathy denotes a human judgment that a response feels understanding or caring. Clinical empathy concerns a trained practitioner's responsive understanding situated within assessment, professional boundaries, and a continuing duty of care. These constructs can correlate, but their measures and normative implications differ. Artificial systems can be evaluated for behavioral and perceived empathy through blinded experiments; clinical obligations arise from the actual service arrangement and consequences, not from a rating of verbal warmth.

Franke Föyen (2025) asked licensed mental-health clinicians to compare AI-generated and expert psychological advice under a blinded comparative design. AI responses were rated more favorably on some forms of empathy, while participants' beliefs about who wrote the advice strongly influenced ratings. In that research, actual source and perceived source were analytically distinct. This is important because an apparently empathic answer can be persuasive for reasons unrelated to clinical accuracy, and labeling can affect evaluation in ways that would disappear in an ordinary unblinded product study.

An adequate test crosses empathic style with factual and clinical correctness. It includes scenarios in which a response warmly affirms an inaccurate belief, offers overly specific advice without assessment, or agrees with a user's dangerous interpretation. The desired measure is not warmth alone but correctly calibrated understanding that preserves agency, uncertainty, and paths to qualified care. The study by Ayers (2023), built on clinician ratings of answers to public forum questions, demonstrates what high-quality prose can accomplish perceptually. It cannot establish that the same system manages therapeutic rupture, recognizes risk across sessions, or maintains patient safety over time.

Disclosure, Authorship Attribution, and Informed Choice

Disclosure is a psychological intervention in its own right because source identity influences how a recipient interprets advice. Relevant information includes that the interlocutor is artificial, whether a human clinician is participating or available, the purposes for which the product has been validated, what happens to records, whether conversation content is used to train or improve systems, the limits of crisis recognition, and how to leave the system or reach a human service. A general statement about artificial intelligence is insufficient when a product presents itself in the role of a therapist or responds to emergencies. Informed choice depends on the user's ability to connect disclosure to decisions, not simply on the appearance of a notice.

Scientific experiments should distinguish the effect of accurate source disclosure from the effects of tone and product reputation. In a disclosure experiment, participants may be randomly assigned to truthful descriptions of the same validated intervention while the content of advice is held constant. Investigators can measure changes in perceived competence, willingness to share, judgment about confidentiality, and intention to seek human care. Because some products differ in whether clinicians actually intervene, study arms must not use fictional provider claims: the disclosed information must be true for the intervention being tested. The comparison can be ethically conducted with explicit debriefing and appropriate safeguards.

The findings of Franke Föyen (2025) support the importance of perceived authorship, but a blinded rating task is not a longitudinal consent trial. A person deciding whether to disclose trauma, suicidal thinking, or medication use faces a qualitatively different decision from an evaluator comparing short advice excerpts. An accountability audit should therefore test recall and comprehension of terms, whether a user can identify a human escalation channel, how frequently a notice is encountered, and whether the actual user journey accords with what was disclosed. Provenance labeling also needs to survive interface redesign and changes in model version.

Trust Calibration and the Authority of Fluent Systems

Trust calibration describes whether reliance is proportionate to performance under the conditions in which the system is used. Appropriate reliance is neither maximal use nor categorical rejection. It consists of choosing to act on reliable outputs, seeking verification for uncertain outputs, detecting known limits, and retaining access to competent human evaluation when stakes warrant it. Trust can be measured by behavior rather than only self-report: acceptance of recommendations, cross-checking, delay before acting, use of override controls, and correction following demonstrably false answers.

Research on automation bias provides an instructive precedent. In a simulated aviation decision task, Skitka (1999) found omission and commission errors associated with an imperfect automated decision aid. This result concerns an entirely different operational domain; it supplies a mechanism worth testing in clinical interaction, not an estimate of psychiatric harm. Generative systems may compound reliance when they provide warm explanations, remember earlier disclosures, and use confident clinical language. A user's sense of rapport can make advice feel especially credible even when its factual basis is weak.

A robust experimental program varies reliability, confidence signals, source identification, and opportunities for human verification. It measures which users detect misinformation, when they update trust after a known error, and whether the system's admissions of uncertainty improve actual decisions. High levels of subjective trust after successful low-stakes conversations cannot substitute for demonstrated accuracy in uncommon high-stakes scenarios. Trust calibration should be assessed separately for supportive text, symptom interpretation, emergency guidance, and treatment decisions, because reliability is task-specific.

Psychological Disclosure, Privacy, and Sensitive Data

Artificial conversation can offer immediate availability, a perceived absence of judgment, and a consistent format that facilitates self-disclosure. These features may support reflection and engagement. They also create a concentrated record of intimate information: symptoms, interpersonal conflict, identity, trauma, family relationships, substance use, self-harm concerns, and details about third parties. The psychologically experienced intimacy of a conversation is not equivalent to legal confidentiality, professional privilege, or technical secrecy. Users can reasonably mistake an interface resembling therapy for an environment subject to the same duties as a licensed clinician.

Privacy research in this domain must therefore study both behavior and understanding. A disclosure experiment can compare actual information shared with the participant's estimate of storage, retention, vendor access, and data reuse. A consent-comprehension test can measure whether participants understand that a safety escalation may cause human review. Data minimization requires that only information necessary for the declared function is collected and retained; retention schedules, access controls, deletion procedures, and incident response need independent verification. Sensitive disclosure from a third party introduces obligations that cannot be resolved solely through the original user's consent.

World Health Organization. (2024) emphasizes governance of large multimodal models in health contexts, including transparency and risk. These principles must be translated into concrete measurement rather than adopted as reassuring language. A system promising human escalation must specify whether the contact details collected are accurate, whether a professional is staffed at the relevant times, how users learn about emergency information sharing, and how incidents are audited. Clinically meaningful privacy is the combination of truthful communication, proportionate data handling, access safeguards, and accountable institutional practice.

Dependence, Engagement, and the Changing Therapeutic Relationship

Repeated supportive exchanges can shape habits and expectations. A person may return because the system offers useful exercises, remembers prior goals, provides temporary relief, or appears to provide unconditional acceptance. Those pathways need distinct constructs. Engagement is the observed frequency and pattern of use; adherence is completion of an indicated intervention; dependence concerns difficulty regulating use or withdrawing from a relationship despite impairment. Frequent use alone is insufficient evidence of pathological dependence, especially for users managing chronic symptoms or limited access to services.

The study of dependence requires time-sensitive measures. Daily assessments can connect mood, loneliness, symptoms, social contact, chatbot use, and help seeking; repeated interviews can document meanings attached to the system; and carefully designed interruptions can test whether reliance persists after tools are unavailable. Stronger designs should examine whether deterioration precedes increased use or follows it, whether human care is displaced, and whether the artificial interaction bridges the user toward effective assistance. A single survey linking loneliness to chatbot use cannot determine direction. Randomized designs comparing comparable human-supported and AI-supported approaches are more informative than comparisons against no contact.

Clinical accountability becomes important when products maximize conversational retention. Optimization for session duration or recurring engagement may differ from the goals of therapeutic progress, autonomy, and discharge. Evaluation should therefore include appropriate termination, informed referrals, constructive disagreement, and skill transfer to offline life. A system that makes a user feel continuously supported while delaying a needed assessment may score well on momentary satisfaction and poorly on clinical outcome. Providers should measure both and report how they resolve incentives when the metrics diverge.

Experimental Evidence From Structured Therapeutic Chatbots

The first generation of therapeutic conversational agents often combined scripted choices, psychoeducation, reminders, and exercises associated with established psychological treatments. Their scientific importance lies in isolating a reproducible intervention package. In the Fitzpatrick (2017), Fitzpatrick, Darcy, and Vierhile (2017) assigned 70 young adults reporting depression and anxiety symptoms to an automated cognitive behavioral agent or an informational control. The two-week study provided initial evidence of symptom improvement in a narrowly defined population, with an intervention designed around cognitive behavioral principles. Its sample, duration, comparator, and self-reported outcomes limit the conclusions that can be drawn about severe illness, crisis management, sustained remission, or equivalence to clinician-delivered psychotherapy.

The proper estimand in such a trial is the effect of assignment to a named intervention under observed conditions. Its result includes content, interface, prompts, reminders, participant expectations, and engagement. A positive score change does not isolate the conversational agent's apparent empathy as the mechanism. Where treatment content is drawn from independently studied techniques, evidence about those techniques does not automatically validate a new delivery mechanism. The system must still demonstrate fidelity, acceptable use, adverse-event response, and outcomes in the relevant population. Such trials contributed a crucial methodological template: clinical claims require a defined intervention and an adequate comparator rather than anecdotal reports that people felt helped.

Laranjo (2018) surveyed conversational agents across healthcare and found substantial diversity in purposes and evaluation measures. This heterogeneity complicates synthesis even within the pre-generative generation. Retrieval, scripted decision trees, embodied assistants, and unconstrained language models differ in what they can say, how reproducibly they say it, and which conditions can be controlled. For scientific comparison, a chatbot is not a single intervention class analogous to one drug compound. Investigators must describe its functional components, constraints, target population, clinical indication, deployment setting, and safeguards.

Randomized Evidence for Generative Conversational Treatment

Generative systems make it possible to respond flexibly to a wider range of statements while preserving elements of structured intervention. Heinz (2025) tested a specially tuned chatbot, Therabot, in a national randomized trial of 210 adults who had clinically significant depressive or generalized anxiety symptoms or were at high risk for feeding and eating disorders. Participants were randomized to four weeks of chatbot access or a waitlist control. The trial provides evidence that a particular developed and monitored system can improve measured short-term outcomes relative to waiting for the service in the studied setting. Such a result is important precisely because it applies to an identified intervention rather than to every possible general-purpose model.

Interpretation requires comparison-aware reasoning. A waitlist control may differ from active treatment in attention, expectancy, self-monitoring, structure, and perceived support. Improvements beyond a waitlist establish a useful initial contrast but do not establish equivalence to a licensed therapist, superiority to established internet-based therapy, or long-term benefit. Risk behavior, hospitalization, sustained functional improvement, and rare adverse outcomes require longer and larger studies. The delivery environment and its protocol also belong to the intervention: supervision, trained evaluation, and safeguards may make a trial product substantially different from a consumer chatbot made available without those features.

The question of which components caused improvement remains open. A generative program may combine cognitive reframing, activity scheduling, exposure principles, journaling, psychoeducation, and conversational affirmation. Factorial or dismantling trials are needed to determine whether one component, their interaction, or simple structured engagement mediates outcomes. It would be methodologically premature to conclude that simulated empathy itself treats depression because an integrated product reduced symptoms. A credible causal mechanism must connect independently measured intermediate changes to subsequent outcomes under suitable experimental or longitudinal controls.

The Meta-Analytic Evidence and Its Heterogeneity

Aggregated evidence now supports cautious claims about mean symptom outcomes while revealing profound heterogeneity. Zhang (2025) synthesized randomized trials of generative AI mental-health chatbots. Their meta-analysis included 14 RCTs and 6,314 participants and estimated an overall effect size of 0.30 with a 95% confidence interval from 0.004 to 0.59. The reported 95% prediction interval extended from -0.85 to 1.67. This very wide prediction interval matters clinically: the average pooled result provides limited information about the likely effect of a specific future implementation. The authors also identified risk-of-bias concerns and differences between social-oriented and task-oriented interventions.

A synthesis combining different populations, endpoints, exposure doses, and controls answers a broader question than any single efficacy trial. Pooled effects are sensitive to small-study bias, variation in outcome instruments, selective analysis, and the inclusion of quasi-independent intervention variants. Researchers must distinguish immediate post-intervention outcomes from sustained improvement, and compare clinical severity as well as statistical significance. Even a statistically positive mean effect can coexist with harmful responses for a small subgroup facing unusual risks. Randomized trials not designed to detect rare adverse events cannot prove their absence.

The present monograph therefore assigns evidence to three distinct questions. Whether a specific chatbot can yield benefit under a trial protocol has positive experimental support for certain products and symptoms. Whether generative chatbots as a class provide reliable benefit for every population and indication remains unresolved by the available synthesis. Whether a model can safely undertake the responsibilities of an autonomous mental-health clinician is a separate question that symptom endpoints and satisfaction scores do not answer. All three questions should be maintained separately in public communication and clinical authorization.

Comparative Evidence About Empathy and Advice Quality

Comparative rating studies can reveal why users are drawn to artificial advice. In Ayers (2023), three licensed healthcare evaluators judged 195 questions paired with physician and chatbot responses drawn from a public forum. Chatbot replies were often preferred and rated higher on empathy and quality. The published correction notes that accuracy was not independently assessed in the manner some summaries implied. The findings concern a particular answer-generation task and rating arrangement; they do not establish the safety of making individualized psychiatric treatment decisions from such responses.

Franke Föyen (2025) provide a more directly psychological comparison. Licensed clinicians evaluated brief AI-generated and expert-authored advice without knowing the actual source. AI-written advice received favorable ratings on emotional and motivational empathy, while perceived expert authorship predicted much stronger preference than actual origin alone. This distinction demonstrates that human judgments of competence and warmth are shaped by source expectations as well as content. The authors used a blinded comparative design, making the study valuable for separating message properties from attribution.

Neither comparison measures a complete course of therapy. Therapeutic work involves eliciting information over time, negotiating goals, responding when change stalls, recognizing adverse effects, managing crises, revising hypotheses, and making decisions under uncertainty. A good reply can facilitate a single moment of reflection while the broader system performs poorly on continuity or risk. These experiments therefore justify testing user experience and clinician perceptions as separate endpoints alongside independent measures of correctness, treatment integrity, safety, and outcome. Their strongest contribution is methodological: they reveal why perceived therapeutic value can diverge from clinical validation.

Evidence for Remote Human Therapy as a Comparison

Remote psychotherapy by licensed human clinicians is an important comparator because it separates the technological medium from the identity of the provider. Aafjes-van Doorn (2024) analyzed the association of alliance with outcomes in teletherapy and found a small significant association across 34 independent samples. Its interpretation belongs to human-delivered treatment through communications technology, with access to ordinary professional obligations and the therapeutic alliance established through human interaction. It should not be described as evidence that a standalone chatbot possesses the same therapeutic alliance.

Comparison groups for future AI trials could include clinician-led teletherapy, evidence-based digital CBT with a human coach, structured self-help without conversational personalization, and general-purpose chat-based support. These alternatives answer different policy questions. Superiority to a waitlist may justify further investigation; noninferiority to an active established treatment requires a prespecified margin, sufficient power, equivalent outcome collection, and serious adverse-event monitoring. If a product is proposed as an adjunct to professional care, its incremental effect should be tested against that care alone. If proposed as a substitute, the burden of evidence is higher.

Access and cost considerations should enter analysis without becoming shortcuts around validation. A scalable system may reach people who otherwise have no treatment; this is a meaningful public-health motivation. But the relevant comparison depends on local service availability, literacy, disability access, language, crisis resources, and the probability of delayed diagnosis. Pragmatic trials measuring equitable access and downstream care use can test the real-world value of a chatbot better than universal claims about affordability derived from the marginal price of one conversation.

Clinical Meaning of Short-Term Symptom Change

Symptom scales quantify perceived distress or impairment over a reference period and are indispensable outcomes when suitably validated. However, clinical meaning depends on baseline severity, instrument reliability, meaningful change thresholds, comorbidity, functional outcomes, and duration. An average reduction in depression score may reflect genuine improvement yet obscure deterioration among high-risk participants. Missing outcome data can be especially consequential when disengagement results from worsening symptoms or adverse experiences. Reports should present attrition, treatment adherence, prespecified analysis populations, uncertainty, and heterogeneity across relevant demographic and clinical groups.

The research corpus presently favors short-term studies, selected participant groups, and outcomes accessible by questionnaire. Longer trajectories could involve relapse, shifts in help seeking, increased or reduced clinical service use, changes in interpersonal functioning, and dependence on the system. Trials should include follow-up after access ends, because sustained benefit is distinct from benefit obtained during use. Clinically important negative outcomes should be solicited systematically, rather than inferred from the absence of spontaneously reported complaints. Investigators also need a route for participants to access appropriate human care if adverse symptoms emerge.

A public claim such as evidence-based AI therapy should be linked to the precise population, product version, intervention protocol, comparator, endpoint, and interval supporting it. It should also disclose whether the evaluation involved live human oversight, especially when the same model family appears in products without that oversight. This form of attribution protects the value of experimental achievements by preventing unsupported extrapolation. It offers a stable foundation for clinical accountability: the organization responsible for a claim can identify what evidence it possesses and what remains to be tested.

Crisis Signals and the Limits of Conversational Recognition

Acute mental-health crises may be expressed directly through statements of immediate danger, indirectly through hopelessness or coded disclosure, or ambiguously through humor, quotation, dissociation, or changes in context. Recognition requires more than matching keywords. It involves understanding what an utterance might imply, what additional assessment is required, and which action is justified by the available evidence. In practice, a crisis workflow must establish the distinction between a conversational response, a structured screening process, and a professional clinical risk assessment. The latter may require information outside the conversation, including current intent, prior history, protective factors, means access, functional context, and local emergency resources.

Arnaiz-Rodriguez (2026) developed a clinically informed taxonomy and a dataset of 2,252 selected mental-health crisis examples, with 2,046 used in model testing. Their analysis examined five large language models and observed variations in safety across systems and scenario types, including weaknesses with indirect or ambiguous crisis cues. The test design itself deserves scrutiny: selected examples were assembled from public datasets, and much evaluation relied on a model-based judge whose agreement with human assessment was examined. Such a benchmark provides consequential evidence of potential failure modes but cannot measure real-world incidence, the prevalence of crises among product users, or the true effectiveness of emergency intervention.

A separate clinical problem is the response to a recognized signal. A tool can identify a phrase suggestive of suicidality yet produce an inappropriate message, fail to ask needed clarifying questions, or have no way to transfer responsibility. Conversely, a cautious refusal or generic emergency message may avoid one type of dangerous output while failing to connect a person to help. Evaluation should therefore independently measure risk recognition, severity discrimination, context-sensitive response, successful routing, and subsequent contact. A single aggregate safety score hides potentially material weaknesses in any one step of this chain.

Consumer Interfaces, APIs, and Deployment-Specific Safety

Safety belongs to the deployed system, including the model, instruction hierarchy, moderation layers, retrieval sources, user interface, escalation channels, and version. In Cantor (2026), five chatbot systems were examined through public interfaces and direct programming interfaces using standardized suicide-related prompts. The analysis found variation by both system and mode of access, particularly for higher-risk prompts. The study establishes that outcomes observed with one access pathway cannot be assumed for another even when a common model family appears underneath.

The design also limits inference. A response to a standardized hypothetical prompt is not a patient outcome, and a categorical determination of whether an answer is direct may not fully describe whether it was supportive, clinically appropriate, or harmful in a particular lived context. Nonetheless, the mode effect is a direct challenge to model-only certification. Procurement committees and clinical researchers need exact records of the deployed interface, safety configuration, update policy, crisis routing, and access channel. The unit of validation should be the user-facing service with all relevant safety layers, followed by monitoring under real operating conditions.

Version changes create an additional accountability challenge. A vendor may alter a model, its refusal policy, system prompt, external retrieval tools, or memory function without any visible change in the product name. Even a previously validated risk protocol may then behave differently. Change control should identify which modifications require renewed scenario testing, clinical review, or suspension. A system that passes crisis scenarios in a laboratory should not receive indefinite approval when its operative components remain mutable. Observability must include records sufficient to reconstruct the relevant configuration for a disputed incident, subject to strict privacy controls.

Human-Validated Safety Benchmarks

Clinical participation in safety evaluation improves construct validity but does not transform a simulated benchmark into a clinical outcome study. Bentley (2026) examined human validation of the open-source VERA-MH framework for chatbot suicide-risk detection and response. Its contribution is to define and test evaluation procedures against clinician judgments, enabling structured comparison of candidate systems. The work was supported by a commercial mental-health provider, and authors reported employment and equity interests. Such disclosures matter because safety benchmarks can influence procurement and competitive claims; independent replication remains an important safeguard.

Benchmarks must represent the phenomena they purport to measure. An overrepresentation of explicit, short, single-turn crisis statements may inflate apparent safety relative to multi-turn cases in which meaning emerges gradually. Rare but consequential events are difficult to sample from routine use. Synthetic conversations can improve scenario diversity and permit privacy-preserving tests, yet synthetic realism itself requires validation. Clinicians may disagree about the most appropriate response to an ambiguous cue, and this uncertainty should be documented rather than erased through a single majority label.

Good benchmark design includes clear taxonomies, severity strata, demographic and linguistic diversity, adversarial and indirect disclosures, and predefined scoring for omissions, harmful assertions, and actionable referral. It also measures false positives, because excessive escalation can burden clinicians, undermine trust, and deter honest disclosure. Beyond average scores, the reporting should reveal performance for the hardest and least common cases; uncertainty, model version, evaluator conflicts, and access mode; and the specific remediation undertaken after a serious failure. Arnaiz-Rodriguez (2026) and Bentley (2026) together point toward this more differentiated safety science.

Prospective Human-in-the-Loop Escalation

Human oversight is clinically relevant only when a defined pathway can be executed. It includes an agreed trigger, reliable detection, notification, a trained person with decision authority, timely contact, appropriate assessment, documented disposition, and follow-up. Glasner (2026) prospectively reported implementation of a human-in-the-loop suicide risk screening and escalation protocol in a generative chatbot study. The investigators used standardized screening and human clinical contact in a single-arm context. Their design is especially informative for operations because it links a digitally detected signal to an actual human response rather than ending at a textual recommendation.

At the same time, implementation feasibility and clinical effectiveness are different claims. A single-arm study can describe screening yield, referral attempts, contact success, and operational bottlenecks. It cannot by itself show that the protocol prevents suicide, improves longer-term symptoms compared with another service, or generalizes across healthcare systems. The use of a fixed screening questionnaire is another important distinction: when a system presents a validated instrument unchanged, it does not follow that the generative model itself validly detects all forms of suicidal thinking in open dialogue. The components should be evaluated independently and as an integrated pathway.

An accountability system must also account for failed contact. If a flag is raised but the user cannot be reached, the event remains clinically relevant. Records should distinguish attempted from completed contact, acknowledged referral from attended care, and advice to seek help from a documented handoff. Response-time benchmarks need to match the urgency implied by the protocol and the practical availability of local services. A product should not promise continuous emergency monitoring when human teams are staffed only at certain times or when technical alerts cannot be reliably delivered.

False Negatives, False Positives, and Base-Rate Problems

Crisis screening involves a tradeoff between missed risk and unnecessary escalation, influenced by the event's prevalence in the evaluated population. When an event is uncommon, even a tool with seemingly favorable sensitivity and specificity can have a low positive predictive value. This is a mathematical property of conditional probability, not a defect unique to AI. Reporting only the proportion of known-risk scenarios correctly identified gives incomplete information about the burden on clinical staff, patients receiving alarming false alerts, and the number of events left undetected in routine use.

Different thresholds imply different duties. A broad screening tool may correctly favor sensitivity and refer uncertain cases for human review, but the organization must have capacity for the resulting referrals. A narrower trigger might be operationally easier while missing subtle warnings. The choice should be justified by use case, population, available services, and consequences of both error types. Structured screening and clinical assessment must retain their distinct purposes: an algorithmic score does not independently settle whether a patient requires emergency intervention.

Performance evaluation should report the denominators for each stage: people eligible, people screened, signals detected, alerts transmitted, users contacted, assessments completed, and referrals attended. Each stage has its own error process and failure mode. A high detection rate with low contact completion is insufficient for a claim of effective escalation. Likewise, a high proportion of completed calls among those reached omits users whose contact data were wrong or who withdrew. The entire pathway and its missing data must be visible to clinical reviewers.

Adverse Events and Iatrogenic Possibilities

Potential adverse effects arise from both the content and the relational organization of interaction. Harmful content can include inaccurate clinical interpretations, encouragement of unsafe action, unjustified diagnostic certainty, or advice that delays appropriate intervention. Relational and service-level effects may include escalating emotional dependence, displacement of valued support, inappropriate disclosure of sensitive information, and a sense of abandonment following abrupt product change. Some of these pathways have observational or simulated support; others remain plausible hypotheses requiring prospective measurement. Their scientific treatment should clearly identify the evidentiary status.

Adverse-event monitoring cannot rely exclusively on voluntary complaints submitted by engaged users. Individuals who disengage, deteriorate, or are hospitalized may be systematically absent from feedback dashboards. Trials should define events prospectively, use independent monitoring where appropriate, specify attribution procedures, and collect follow-up beyond the final conversation. Qualitative investigations can detect harms missed by standardized scales, while controlled investigations can test whether particular conversational features increase risk. The reporting standard should distinguish verified adverse events from simulated failures and from theoretical hazards.

For particularly vulnerable patients, safety controls must be examined against the intended use and actual service contract. A system offered as a general wellness companion may still receive clinically urgent disclosures, but it should communicate boundaries and direct users to appropriate resources. An intervention marketed to patients with a diagnosed disorder carries a different validation and oversight burden. Across settings, American Psychiatric Association. (2026) identifies professional and patient-safety concerns surrounding unsupervised generative systems. Scientific risk assessment should translate those concerns into measurable standards, avoiding claims of proven danger or proven safety that exceed the empirical evidence.

Clinical Accountability as an Operational System

Accountability in human–artificial therapeutic interaction is an operational system connecting justified claims, assigned duties, ongoing evidence, and accessible remedies. Its necessary elements are identification of the party responsible for each function; a record of what the system is intended and authorized to do; methods for detecting failure; human capacity to respond; and a mechanism for learning from incidents. An organization cannot establish accountability solely by announcing that its software is experimental or that users should seek professional advice. Such disclosures may help define an appropriate relationship, but they do not substitute for performance evidence, truthful representation, and effective safety practices.

The accountability unit is the complete service configuration. A developer may train or fine-tune a model; an application provider chooses prompts, memory, retrieval, and interface design; a healthcare organization determines eligibility and escalation policy; a licensed clinician may review certain outputs; and a patient decides whether to accept a suggestion. These actors contribute different capabilities and hold different duties under contracts, professional norms, product regulations, and jurisdiction-specific law. The monograph proposes an accountability trace in which every consequential decision can be linked to its input information, technical version, decision-maker, applicable policy, and recorded outcome. This is a research and governance model rather than a claim that responsibility is universally allocated in the same way under all legal systems.

A clinically responsible system must identify foreseeable interactions that exceed its tested competence, define responses to them, and verify those responses through independent appraisal. If the user asks for a diagnosis, reports acute risk, requests medication changes, or expresses uncertainty about consenting to care, the interface should respect its validated scope and provide proportionate routes to qualified services. The quality of that routing is an observable outcome. Merely generating a disclaimer is a weak substitute for demonstrating that a user can access assistance and that a clinician receives sufficient information for safe continuation.

Responsibility Across Developers, Deployers, and Clinicians

Developers are best positioned to document the model's training and evaluation provenance, system limitations, known error modes, and response to software changes. Application deployers control which features reach users, which prompts and policies shape output, how data are retained, and whether escalation is genuinely available. Healthcare institutions have responsibilities for procurement, training, patient inclusion, incident reporting, and integration into clinical practice. Clinicians remain accountable for their own professional decisions and the reasonable use of tools within their competence and setting. Patients exercise autonomy over participation and disclosures, while the design and accuracy of representations materially shape what choices are possible.

American Psychological Association. (2025) emphasizes assessment of AI quality, professional judgment, and appropriate oversight. This professional position identifies responsibilities when clinicians integrate AI into practice. It does not transform every consumer chatbot into a medical device or every use of an AI interface into a formal therapist–patient relationship. Applicability depends on what services are being provided, by whom, for whom, and under which rules. These particulars should appear in the deployment dossier so that reviewers can determine who actually bears each obligation.

Failure investigations should avoid treating the phrase human in the loop as conclusive. A human who receives too many alerts, lacks the relevant training, cannot access the dialogue, or has no authority to intervene may add little safety. Conversely, a human supervisor equipped with concise context, valid assessment tools, clear response deadlines, and authority to halt a system may materially change outcomes. A strong research design compares different oversight configurations using prospective metrics rather than assuming that a human presence guarantees responsible care. Operational performance, staffing arrangements, and shared accountability must be evaluated together.

Clinical Governance, Professional Standards, and Product Claims

A claim that a chatbot offers therapy raises questions about the conditions under which a psychological intervention should be represented as treatment. These questions include the intended indication, age range, evidence base, contraindications, clinician role, outcome monitoring, adverse events, informed consent, and whether the claimed benefit applies to the currently deployed version. Professional guidance reflects these stakes. American Psychiatric Association. (2026) addresses therapeutic use of generative AI and emphasizes safeguards, disclosure, and the present limitations of autonomous treatment claims.

In the United States, the regulatory characterization of a software function turns on its intended use and applicable law. Food and Drug Administration. (2026) explains criteria affecting whether certain decision-support functions fall outside the statutory medical-device definition and describes examples of regulated and non-device functions. A mental-health chatbot cannot be categorized accurately from the generic label AI or wellness alone. Function-specific claims, users, decision independence, underlying information, and intended clinical use matter. The guidance is jurisdiction-specific, and legal classification of a particular product requires fact-specific assessment.

Clinical governance should establish a written intended-use statement, evidence register, named clinical lead where appropriate, versioned risk assessment, monitoring calendar, adverse-event procedure, incident escalation, accessible complaints route, and decommissioning plan. These controls become meaningful when audited against records and patient outcomes. World Health Organization. (2024) and the National Institute of Standards and Technology. (2024) provide complementary frameworks for organizational governance. They are governance sources rather than direct trial evidence that any individual chatbot improves mental health.

Documentation, Incident Investigation, and Learning

An auditable service requires records calibrated to the sensitivity of the data. Technical logs may document model version, interface configuration, safety-layer state, high-risk alert triggers, and escalation decisions. Clinical records may document the assessment and disposition taken by a professional. Privacy law and professional confidentiality requirements limit what should be collected, who may read it, and how long it may be retained. The aim is sufficient accountability with proportionate data minimization. Unrestricted collection of every intimate conversation can itself introduce major harm.

A serious-incident review should reconstruct the sequence of events from initial user disclosure to system output and human follow-up. Questions include whether the system recognized a relevant warning, whether its response agreed with validated protocol, whether an alert was issued, whether the alert arrived, who had authority to act, whether contact succeeded, and which practical obstacles affected the outcome. Reviewers must also examine whether the product's public claims or interface design contributed to reliance on the system beyond its evaluated scope. Each identified failure should link to a corrective action and a method for testing that action.

Incident rates need denominators: number of users, relevant sessions, identified crises, escalation attempts, interventions completed, and subsequent outcomes. A service that experiences no reported incidents may lack the ability to detect them. Conversely, increased incident reports after implementation of better monitoring can reflect improved detection rather than worsening product performance. Transparent reporting should separate confirmed harm, suspected harm, process failures, near misses, and simulation findings. This distinction protects the scientific record and enables comparisons across deployments with different monitoring intensities.

Institutional Liability and the Limits of Generalization

Clinical accountability has legal, ethical, scientific, and operational aspects that can overlap without becoming identical. A professional's ethical duty may be broader than a narrow contractual obligation; a regulatory authorization concerns a specified product function rather than every foreseeable clinical use; and the presence of a technical safeguard does not settle a legal claim. Jurisdictions differ in negligence standards, medical-device definitions, privacy frameworks, professional licensing, consumer protection, and the status of AI-generated advice. The evidence synthesized here supports analysis of risk and governance, while particular legal liability remains a fact-specific matter.

The most useful interdisciplinary question is how responsibility can be made visible before an incident occurs. A care organization should know which AI tasks are included in its clinical pathways, which protocols a human clinician has reviewed, what the patient was told, which outcomes are monitored, and which person or department can halt deployment. A product provider should know what evidence supports its marketing claims and whether previously documented safety behavior persists after updates. Researchers should know whether the treatment package in a clinical trial matches the service being offered to users.

The result is a unified accountability question: can every consequential transition in the human–artificial therapeutic process be explained, tested, and corrected by an identifiable responsible actor? The question is empirically tractable through process tracing, interface testing, longitudinal outcomes, incident audits, and pragmatic clinical evaluation. It is also institutionally significant because responsibility without effective control becomes nominal. The evaluation of accountable artificial therapy should therefore assess allocation of authority together with real capacity to intervene.

Construct Validity and the Measurement of Therapeutic Interaction

Psychological measurement in this domain begins by specifying the construct and its referent. Perceived empathy is a judgment made by a person about a response. Response appropriateness is an assessment of content against a defined clinical or ethical criterion. Therapeutic alliance is a multidimensional relationship construct with a history of use in human psychotherapy. Clinical utility is improvement in a patient-relevant outcome. These measurements answer different questions and require distinct validity evidence. A model may be reliable at producing comforting language while users misunderstand the scope of the service; it may also deliver accurate information that users experience as cold or alienating. Neither pattern can be described adequately by a single satisfaction score.

Measures developed for human psychotherapy need revalidation when used for artificial systems. Items invoking mutual commitment, therapeutic care, therapist feelings, or a clinician's professional role may change meaning when the interlocutor is software. Content validity should be examined by experts and diverse users, followed by cognitive interviews that reveal how respondents interpret items. Factor structure, reliability, measurement invariance, convergent validity, and discriminant validity should be tested across systems and use cases. An alliance-like scale should not simply reproduce perceived anthropomorphism or liking. Its ability to predict outcomes should be compared with alternative measures such as usability, accuracy, goal concordance, and engagement.

A scientifically informative instrument distinguishes the person, system, and interaction. User characteristics include symptom severity, expectations, prior digital literacy, access to care, and relational needs. System characteristics include response fidelity, version, latency, disclosure practices, personalization, and safety constraints. Interaction characteristics include cooperation on goals, disclosure, response relevance, repair after error, and help-seeking transitions. Model comparisons should report how these attributes were measured instead of collapsing them into the label AI therapeutic quality.

Randomized Designs, Causal Mediation, and Active Controls

Random assignment is essential for identifying the effect of access to a particular intervention when conducted and analyzed appropriately. Yet causal interpretation also depends on what the treatment actually changes and whether a comparison isolates the intended mechanism. A waitlist group distinguishes an intervention package from waiting; a credible active control better addresses attention, expectancy, structured engagement, and nonspecific support. To evaluate apparent empathy as a mechanism, investigators might independently randomize empathetic wording, factual accuracy controls, and therapeutic content within an ethically approved support program. Outcomes must then be measured after the proposed mediator and before or at the planned endpoint.

Randomization of the full product does not automatically identify mediation through perceived alliance or warmth. These variables may be affected by early symptom changes, engagement, user preference, and unmeasured characteristics, creating feedback and confounding. A proposed mediation analysis must address temporal order, measurement error, and the causal assumptions needed to interpret indirect effects. Repeated assessments and experimental manipulation of a treatment component improve inference, though even they may leave unresolved the relationship between observed proxy measures and the underlying mechanism.

The field needs several types of trial. Initial efficacy trials can compare a clearly defined treatment package with a suitable comparator. Dismantling trials can identify which conversational or cognitive behavioral components contribute to outcomes. Pragmatic service trials can evaluate access, retention, actual clinical referrals, functional outcomes, and safety in realistic settings. Noninferiority trials against established care require defensible margins and adequate detection of clinically important adverse outcomes. Liu (2020) and Cruz Rivera (2020) provide reporting considerations that support reproducibility and specification of the system's interaction with humans.

Longitudinal Trajectories, Feedback, and Selection

Longitudinal evaluation must separate short-term affective relief from durable change. A chatbot encounter can alter mood within minutes; repeated use may influence coping, interpersonal expectations, and help seeking over months. The relation between use and symptom change can be bidirectional: distress may prompt more use, while beneficial or adverse exchanges may alter later distress. Simple correlations between conversation frequency and well-being can therefore be misleading. Within-person designs, repeated symptom measurement, event-based diaries, and appropriately modeled time-varying confounding help distinguish patterns that cross-sectional surveys cannot.

Attrition is not neutral. A person who discontinues because the product is repetitive differs from someone who stops because symptoms have improved or because a conversation was distressing. Investigators should seek reasons for discontinuation, analyze missing-data mechanisms, and disclose how conclusions depend on imputation assumptions. Product telemetry supplies detailed records of use but cannot independently establish why behavior changed. Ethical collection must permit meaningful consent and data minimization; not every person seeking confidential help will agree to extensive behavioral tracking.

Version drift requires special treatment in long studies. A participant who begins with one set of safety constraints and continues after a model update has experienced more than one intervention state. Trials should retain version histories, document exposure windows, and evaluate whether changes alter the causal contrast. When updates are frequent, the research may need sequential or platform-level designs rather than an analysis that treats a nominal product as permanently stable. The resulting evidence should apply to precisely dated implementations and disclose when generalization to current software remains untested.

Human Oversight as an Experimental Variable

Oversight should be studied as an intervention component with its own fidelity, burden, and effect on care. Relevant design contrasts include AI-only informational support, AI supported by scheduled clinician review, triggered human escalation, and clinician-led therapy assisted by an AI tool. Ethical constraints require appropriate safeguards for all arms, particularly where participants could experience acute risk. Comparisons can examine whether clinician review improves accuracy, crisis management, engagement, service linkage, and patient-reported autonomy; whether alerts cause fatigue; and whether the additional oversight is cost-effective in specific populations.

Glasner (2026) illustrates the value of recording an actual risk-routing workflow. The next scientific step is to compare oversight configurations while monitoring consequences over longer intervals. Researchers should specify staffing, hours of coverage, response times, escalation thresholds, escalation completion, clinician training, and competing responsibilities. Human oversight without these operational details is too broad a treatment label for reproducible clinical science.

Human review can also change therapeutic meaning. Knowledge that a clinician may read selected conversations could increase trust for some users and reduce intimate disclosure for others. Randomized disclosure of truthful oversight arrangements and qualitative interviews could identify how these effects vary. Such research should include users who decline surveillance, because evaluation limited to those willing to share detailed transcripts may overstate acceptability and underrepresent privacy concerns. Safeguards must accommodate the distinction between clinical necessity and platform preferences for data collection.

External Validity, Culture, Language, and Access

Findings from English-language, technologically literate adult samples cannot automatically be generalized to adolescents, older adults, people with disabilities, communities with low health-service access, or users speaking other languages. Clinical expression of distress is culturally and linguistically patterned; metaphors, indirect disclosure, and pragmatic conventions influence whether a system recognizes concern and responds respectfully. A translation that retains surface meaning may still change the apparent authority or emotional quality of a therapeutic message. Validation should include participants from the intended communities rather than assume that a high average benchmark score implies equitable performance.

Subgroup analysis must be designed in advance when a product is expected to serve different populations. Measurement invariance tests whether symptom scales and alliance-like instruments retain comparable meaning. Stratified analyses can investigate variation in benefit, false negatives, false positives, response appropriateness, and referral access. Samples must be sufficiently powered for planned contrasts, and apparent differences from small exploratory subgroups should be reported with uncertainty. Clinician adjudicators and service partners should have appropriate language and cultural competence.

The relevant outcome extends beyond the screen. A referral is useful only if the user can reach a service. Cost, transportation, disability accommodations, insurance, legal status, cultural stigma, and local availability all shape the meaning of human escalation. A system may perform similarly on standard prompts across regions yet lead to unequal patient outcomes because follow-up pathways are inaccessible. Evaluating human–artificial care in real settings requires measuring the service ecology and the continuity between digital advice and actual assistance.

Reproducibility, Model Updates, and Transparent Reporting

Research reproducibility requires an adequate description of the deployed system without disclosing information that would compromise security or patient privacy. Reports should identify the product, release period, model family or fixed version where available, system configuration, safety policy, retrieval and memory functions, human intervention components, and the interfaces through which participants interacted. Trial protocols should specify eligibility, crisis handling, version changes, outcome schedules, missing-data rules, and adverse-event monitoring. Deviations and corrective actions should be dated and documented.

Safety benchmarks benefit from independent clinical ratings and open methodological artifacts when feasible. Bentley (2026) published resources supporting examination of a suicide-risk benchmark; Arnaiz-Rodriguez (2026) described a taxonomy and evaluation procedure for mental-health crisis exchanges. A benchmark should expose its sampling method, scenario categories, scoring rubric, evaluator agreement, and limitations. When tests use private real-user information, privacy protections and restricted independent auditing may be preferable to unrestricted release.

Replication must target both efficacy and safety. A successful intervention trial should be reproduced by investigators outside the development team with a credible comparator, independent outcome assessment, and appropriate subgroup coverage. A safety claim should survive different prompt sets, interface modes, multi-turn contexts, and foreseeable deployment changes. Evidence of one successful outcome at one time remains valuable; documenting its boundary conditions converts that finding into cumulative knowledge rather than a product-wide guarantee.

Clinical Populations and Diagnostic Complexity

Mental-health-related conversations arise in people with diverse clinical histories and goals. Some seek general reflection during transient stress; others live with major depressive episodes, anxiety disorders, trauma-related symptoms, psychosis-spectrum experiences, substance-related problems, neurodevelopmental differences, or multiple co-occurring conditions. A system evaluated for general stress management in a community sample has established a different evidence base from a clinician-supervised intervention for a diagnosed disorder. The clinical context changes what questions must be asked, how uncertainty should be expressed, and when care must be transferred. Comorbidity and fluctuating impairment also make standard prompt-level testing an imperfect model of actual treatment needs.

A disorder-specific trial should define inclusion and exclusion criteria, relevant diagnoses or symptom thresholds, baseline care, concurrent treatments, and the outcome instruments appropriate to that population. Its interpretation must account for co-occurrence and external care received during the study. Clinical accuracy also requires an adequate response when the user's presentation does not fit the product's target population. For example, an intervention designed for mild depressive symptoms must not infer from one successful trial that it can manage acute bipolar mood states or psychotic symptoms. The question is an evidentiary one: do the validation population and protocol match the contemplated use?

Differential diagnosis and medication management involve risks that exceed those of general educational coping suggestions. In a clinical setting, medication advice, individualized assessment, and treatment planning require appropriately qualified professional judgment and an assessment of relevant information. An artificial assistant may help organize patient questions or present validated information, but its suggestions should be governed by their intended function and independent evaluation. Diagnostic uncertainty and competing explanations should be acknowledged when clinical decision support is used. A reassuring conversational tone cannot eliminate the need for assessment of symptoms, context, medical causes, and safety.

Children, Adolescents, and Developmental Vulnerability

Young people may interpret conversational cues differently from adults, and their decisions unfold within families, schools, health systems, and peer relationships. Developmental differences in autonomy, privacy understanding, emotion regulation, and help-seeking pathways affect the meaning of consent and therapeutic reliance. The emergence of naturalistic conversations with generative systems creates new opportunities for social support and psychoeducation, alongside questions about age-appropriate explanations and intervention boundaries. The evidence from adult or young-adult trials must be distinguished from trials designed for minors.

Developmentally responsive research should measure comprehension of the system's artificial nature, understanding of confidentiality and data-sharing rules, perceived authority, social substitution, and the ability to seek suitable adult or professional help. Recruitment and monitoring require age-appropriate ethical review and legally compliant consent or permission processes. Because young people may use indirect language about distress, crisis evaluation should include linguistic contexts relevant to the intended age group. Safety design should not assume that adult-centered emergency instructions will be interpreted or followed in the same way by adolescents.

Subgroup protections should preserve useful access without requiring children to disclose intimate information to systems that have not been validated for their needs. A responsible deployment must clarify whether it is intended for minors, what oversight and escalation are available, how privacy and mandatory reporting obligations apply in the setting, and whether appropriate human services exist. The research agenda includes longitudinal developmental effects, comparison with established youth services, and the circumstances under which artificial support complements rather than displaces protective relationships.

The Artificial Era and Interdisciplinary Psychological Knowledge

Human–artificial therapeutic interaction brings into contact clinical psychology, social cognition, human–computer interaction, medical informatics, safety engineering, health law, and the study of artificial systems. Each discipline contributes a different method: clinical trials estimate treatment effects; experiments on attribution test perceived social meaning; psychometrics evaluates constructs; software audits examine reproducibility and failure; ethics and law specify professional and institutional obligations. A coherent scientific account preserves their connections without treating one mode of evidence as a substitute for another.

Within the Artificial Era, the central research question concerns the psychological organization of a relationship involving distinct forms of participant and capacity. The human side contributes lived distress, goals, judgment, vulnerability, interpretation, and changing social context. The artificial side contributes generated responses, memory configurations, interaction constraints, policy-governed behavior, and system updates. The encounter is formed through reciprocal causal influence at the behavioral level: user input changes the system's next output, and that output may change user belief, feeling, or action. Claims about inner experience or personal status of an artificial system are philosophical and scientific questions that require their own operationalization.

The clinical relevance of the arrangement is determined by demonstrated effects, foreseeable uses, and enforceable responsibility. The capacity of a system to behave as a convincing helper is an important object of inquiry; the legitimacy of a therapeutic role depends additionally on validation and accountability. This principle helps define useful research links with Psychology of Artificial, Psychology of Human–Artificial Relations, Counseling Psychology and Psychotherapy, and Postsubjective Psychology, while preserving the distinctive research object of the present dossier.

An Integrated Model of Benefit, Risk, and Responsibility

The synthesis can be expressed through four interacting domains. Therapeutic affordance comprises the opportunity for supportive exchange, psychoeducation, structured psychological activity, and access at a particular time. Interactional attribution comprises perceived empathy, credibility, social presence, and therapeutic role. Clinical consequences comprise symptom outcomes, functional change, care-seeking behavior, adverse effects, and risk escalation. Governance comprises evidence standards, consent, safety design, human oversight, monitoring, and repair. The model is an organizing synthesis of existing literature, designed to generate testable comparisons rather than to assert an independently proven causal theory.

The links between domains are contingent. A high-empathy message may increase engagement, but may also increase trust in an inaccurate recommendation. A therapeutic role may support self-directed skill practice or create misplaced expectations about human oversight. A well-designed escalation protocol may reduce missed referrals but increase disclosures to clinical teams that users did not anticipate. The net effect depends on the product, clinical setting, population, and implementation. This model therefore treats outcome and responsibility as functions of the full interactional system, not properties that can be read off from conversational fluency alone.

The evidence supports an asymmetric conclusion. It is reasonable to investigate and, under appropriate validation, implement particular artificial therapeutic tools for defined purposes; randomized evidence for selected programs makes their potential empirically meaningful. It is not scientifically justified to infer autonomous clinical competence from conversational persuasiveness or average symptom improvements in limited trials. Safety benchmarks, professional guidance, and deployment studies show why the conditions of responsible use require independent testing. The clinically actionable question is what a specified system demonstrably does, for whom, with what safeguards, and under whose control.

Unresolved Research Problems and a Future Program

The first major open problem is mechanism identification. Current trials establish outcome contrasts for particular products more directly than they establish whether perceived empathy, goal agreement, cognitive restructuring, behavioral practice, expectancy, or therapist-like alliance explains change. The next generation of research should isolate components experimentally and measure intermediate processes with appropriately ordered observations. Claims about relational mechanism should be tested against competing explanations that include engagement, patient preference, pre-existing distress, and human service contact.

A second problem is the scarcity of prospective safety outcomes, especially for rare severe events. Simulation studies provide essential failure discovery and allow systematic stress testing, but they must be complemented by real-world monitoring, representative cohorts, independent clinical adjudication, and ethical follow-up. Researchers should develop scenario tests for indirect crisis expression and multi-turn contexts while disclosing their uncertainty. Incident reporting systems must be capable of finding harms among users who disappear from engagement metrics.

A third problem concerns governance experiments. It is necessary to determine which forms of human review improve outcomes, which generate alert fatigue, and which are acceptable to users with different privacy expectations. Prospective comparisons could vary screening protocols, oversight staffing, patient choice over monitoring, and clinician response pathways. Cost-effectiveness should consider actual downstream care and the resources required to maintain the safeguards, rather than assuming that a low marginal inference cost implies low-cost clinical delivery.

The fourth problem is transportability through time and social difference. Studies should compare languages, age groups, diagnostic presentations, socioeconomic contexts, disabilities, and national health systems. They should retain model and interface histories and measure whether validated performance persists after software updates. Open, reproducible methodological artifacts and independent audit access can support cumulative findings even where sensitive clinical data require controlled handling. Public research reports should establish not merely that a product works, but the boundaries within which the evidence warrants that conclusion.

Critical Scientific Discussion

The literature reveals three recurring inferential errors. The first substitutes perceived warmth for professional empathy and clinical competence. The second substitutes short-term symptom response to a named intervention for broad claims of psychotherapy equivalence. The third substitutes successful responses to selected crisis prompts for demonstrated patient safety in a real service. Each error can be addressed through a different research design: controlled source-attribution and response-quality experiments; adequately powered randomized comparisons with meaningful clinical endpoints; and longitudinal safety monitoring linked to verifiable human escalation. Their joint correction creates an evidence architecture appropriate to high-stakes human–artificial interaction.

There are also unresolved normative tensions that have empirical consequences. Extensive safety monitoring may improve detection while changing users' willingness to disclose; close conversational responsiveness may increase access while strengthening uncalibrated reliance; automatic risk detection may support clinicians while producing alarm burden and privacy concerns. These tensions should be evaluated with real stakeholders and measured as consequences, rather than resolved through promotional promises or purely abstract principles. A robust service can specify how it values autonomy, access, accuracy, and safety and can show whether its practice accords with those commitments.

The central scientific advance is a more precise unit of analysis: the entire therapeutic exchange plus the clinically consequential systems around it. This unit incorporates human psychological experience, artificial response generation, professional and institutional control, and the time course of care. It integrates the experimental evidence for specific interventions with emerging empirical safety research without making either evidence stream carry claims it was not designed to support. Under that perspective, clinical accountability becomes a testable characteristic of a real service pathway.

Conclusion

Human–artificial therapeutic interaction is a scientifically identifiable phenomenon formed by mental-health-related disclosure, artificial response, perceived therapeutic meaning, subsequent behavior, and the systems that authorize and govern consequential decisions. Its study connects historic theories of empathy and therapeutic alliance with experiments on anthropomorphism, randomized digital intervention trials, comparative judgments of AI-generated advice, crisis-response benchmarks, and the practical organization of clinical care. Each literature contributes a distinct kind of evidence and a distinct boundary of inference.

The strongest affirmative findings are that specific automated and generative conversational interventions can improve measured symptoms under particular study conditions; humans can experience artificial responses as empathic and clinically relevant; and safety behavior changes materially with the deployed interface, evaluation scenario, and oversight configuration. The central unresolved questions concern durable outcomes, comparative effectiveness against established treatments, mechanisms of benefit, heterogeneous patient effects, recognition and handling of rare crises, the consequences of dependency, and reliable clinical oversight at scale.

Clinical accountability is the capacity to identify and substantiate what an artificial therapeutic service does, assign duties to actors who can control its behavior, detect and respond to failures, and document the consequences for people using it. Responsible scientific and clinical development requires specific product claims, reproducible evaluation, transparent disclosure, proportionate protection of sensitive information, reachable human pathways, and ongoing verification after system updates. The Artificial Era makes this form of inquiry increasingly consequential: effective therapeutic support and accountable clinical judgment must be studied together in the actual interactions through which artificial systems enter human psychological life.

References

  • Aafjes-van Doorn, K., Spina, D. S., Horne, S. J., & Békés, V. (2024). The association between quality of therapeutic alliance and treatment outcomes in teletherapy: A systematic review and meta-analysis. Clinical Psychology Review, 110, 102430. https://doi.org/10.1016/j.cpr.2024.102430
  • American Psychiatric Association. (2026). Position statement on the use of artificial intelligence for mental health services and treatment. Approved July 2026. https://www.psychiatry.org/getattachment/4f7b33ca-40f2-4f23-b29a-f2204cc17d4e/Position-Use-of-AI-Mental-Health.pdf
  • American Psychological Association. (2025). Ethical guidance for AI in the professional practice of health service psychology. Updated December 2025. https://www.apa.org/topics/artificial-intelligence-machine-learning/ethical-guidance-ai-professional-practice
  • Arnaiz-Rodriguez, A., Baidal, M., Derner, E., Annable, J. L., Ball, M., Ince, M., Perez Vallejos, E., & Oliver, N. (2026). Between help and harm: An evaluation study of mental health crisis handling by large language models. JMIR Mental Health, 13, e88435. https://doi.org/10.2196/88435
  • Ayers, J. W., Poliak, A., Dredze, M., Leas, E. C., Zhu, Z., Kelley, J. B., Faix, D. J., Goodman, A. M., Longhurst, C. A., Hogarth, M., & Smith, D. M. (2023). Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. JAMA Internal Medicine, 183(6), 589–596. https://doi.org/10.1001/jamainternmed.2023.1838
  • Bentley, K. H., Belli, L., Chekroud, A. M., Ward, E. J., Dworkin, E. R., Van Ark, E., Johnston, K. M., Alexander, W., Brown, M., & Hawrilenko, M. (2026). AI chatbot suicide risk detection and response: Human validation study of the open-source VERA-MH safety evaluation. JMIR AI, 5, e92817. https://doi.org/10.2196/92817
  • Bordin, E. S. (1979). The generalizability of the psychoanalytic concept of the working alliance. Psychotherapy: Theory, Research & Practice, 16(3), 252–260. https://doi.org/10.1037/h0085885
  • Cantor, J., Huang, E., Zhang, L. A., Salas, A., Zhang, F., Burnett, A., Lohn, A., Pataranutaporn, P., Rader, B., Yu, H., & McBain, R. K. (2026). Suicide-related responses from AI chatbots through application programming interfaces and consumer-facing apps: Observational cross-sectional comparative study. Journal of Medical Internet Research, 28, e97968. https://doi.org/10.2196/97968
  • Cruz Rivera, S., Liu, X., Chan, A.-W., Denniston, A. K., & Calvert, M. J. (2020). Guidelines for clinical trial protocols for interventions involving artificial intelligence: The SPIRIT-AI extension. Nature Medicine, 26(9), 1351–1363. https://doi.org/10.1038/s41591-020-1037-7
  • Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
  • Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017). Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Mental Health, 4(2), e19. https://doi.org/10.2196/mental.7785
  • Flückiger, C., Del Re, A. C., Wampold, B. E., & Horvath, A. O. (2018). The alliance in adult psychotherapy: A meta-analytic synthesis. Psychotherapy, 55(4), 316–340. https://doi.org/10.1037/pst0000172
  • Flückiger, C., Del Re, A. C., Wlodasch, D., Horvath, A. O., Solomonov, N., & Wampold, B. E. (2020). Assessing the alliance–outcome association adjusted for patient characteristics and treatment processes: A meta-analytic summary of direct comparisons. Journal of Counseling Psychology, 67(6), 706–711. https://doi.org/10.1037/cou0000424
  • Food and Drug Administration. (2026). Clinical decision support software: Guidance for industry and Food and Drug Administration staff. January 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software
  • Franke Föyen, L., Zapel, E., Lekander, M., Hedman-Lagerlöf, E., & Lindsäter, E. (2025). Artificial intelligence vs. human expert: Licensed mental health clinicians' blinded evaluation of AI-generated and expert psychological advice on quality, empathy, and perceived authorship. Internet Interventions, 41, 100841. https://doi.org/10.1016/j.invent.2025.100841
  • Glasner, S., Michero, D., & Stewart, M. (2026). Human-in-the-loop suicide risk detection and clinical escalation in a generative AI mental health chatbot: Prospective implementation study. JMIR Formative Research, 10, e106118. https://doi.org/10.2196/106118
  • Heinz, M. V., Mackin, D. M., Trudeau, B. M., Bhattacharya, S., Wang, Y., Banta, H. A., Jewett, A. D., Salzhauer, A. J., Griffin, T. Z., & Jacobson, N. C. (2025). Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI, 2(4). https://doi.org/10.1056/AIoa2400802
  • Laranjo, L., Dunn, A. G., Tong, H. L., Kocaballi, A. B., Chen, J., Bashir, R., Surian, D., Gallego, B., Magrabi, F., Lau, A. Y. S., & Coiera, E. (2018). Conversational agents in healthcare: A systematic review. Journal of the American Medical Informatics Association, 25(9), 1248–1258. https://doi.org/10.1093/jamia/ocy072
  • 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. The Lancet Digital Health, 2(10), e537–e548. https://doi.org/10.1016/S2589-7500(20)30218-1
  • Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
  • National Institute of Standards and Technology. (2024). Artificial intelligence risk management framework: Generative artificial intelligence profile (NIST AI 600-1). U.S. Department of Commerce. https://doi.org/10.6028/NIST.AI.600-1
  • Rogers, C. R. (1957). The necessary and sufficient conditions of therapeutic personality change. Journal of Consulting Psychology, 21(2), 95–103. https://doi.org/10.1037/h0045357
  • 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
  • World Health Organization. (2024). Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. World Health Organization. https://www.who.int/publications/i/item/9789240084759
  • Zhang, Q., Zhang, R., Xiong, Y., Sui, Y., Tong, C., & Lin, F.-H. (2025). Generative AI mental health chatbots as therapeutic tools: Systematic review and meta-analysis of their role in reducing mental health issues. Journal of Medical Internet Research, 27, e78238. https://doi.org/10.2196/78238