Liability analysis in health care AI often begins with a general question: if an AI system contributes to harm, should responsibility attach to the manufacturer, the deployer, the clinician, or some combination of these actors? This framing is useful, but it overlooks what is distinctive about psychiatry. In this article, “psychiatric AI” is shorthand for AI systems whose outputs or interactions materially influence psychiatric assessment, treatment, monitoring, documentation, or patient support; it does not imply that the underlying technology is unique to psychiatry. Many risks are shared with health care AI, but they may be intensified in psychiatry because decisions depend unusually heavily on testimony, contextual interpretation, therapeutic trust, and liberty-sensitive risk judgments. Psychiatric practice depends on first-person reports, family testimony, behavioral signs, temporality, and shifting contextual meaning. The clinician’s task is rarely binary detection; it involves integrating narrative, risk, diagnosis, and relationship in a setting where the patient may be frightened, ambivalent, cognitively impaired, socially vulnerable, or legally exposed. AI therefore alters not only efficiency, but also the epistemic structure of the encounter [1,6-8]. For this reason, the standard of care in psychiatry cannot be reformulated merely in terms of improved predictive accuracy. It must remain grounded in justified clinical reasoning. Muyskens et al [21] argue that epistemic humility in clinical practice requires recognition of uncertainty, awareness of the limits of knowledge, and respectful engagement with other epistemic sources, including the patient. In computational psychiatry, this is crucial. If algorithmic output is treated as inherently superior because it appears quantitative or objective, the clinical encounter risks becoming a form of machine-credentialed disbelief. A patient’s account may then be displaced because it is inconsistent with a model’s score or category. In psychiatry, where patient testimony is often the primary source of knowledge, this is a structural threat to good practice [1,21].

This problem can be described as testimonial injustice in a digital context. AI does not create testimonial injustice, which can arise from ordinary clinical power asymmetries, but it may amplify and institutionalize it by attaching quantitative authority to an opaque output and reproducing that judgment across clinical workflows. A patient reporting escalating suicidal thoughts, newly commanding hallucinations, or intolerable medication effects may be implicitly disregarded when an AI system labels risk as low or adherence as satisfactory. Once institutions present such systems as validated, clinicians may feel legally, professionally, and institutionally exposed if they depart from the tool without documenting why, but equally exposed if they follow it and harm results. Liability pressure therefore does not arise only after error; it shapes the distribution of authority before decisions are made, including in triage, detention, discharge, and therapeutic alliance [2,14,22,23].

This prelitigation effect is supported by empirical work on psychiatric defensive practice. In an Italian national survey of 254 psychiatrists, most respondents reported defensive medical behaviors in the context of perceived professional liability and medicolegal pressure, a concern sharpened in Italy by the medicolegal position of guarantee, under which psychiatrists may be held responsible not only for diagnostic and therapeutic appropriateness but also for some consequences of patient behavior [24]. Additional analysis from the same survey found that defensive practice was associated with a group that is younger in age, had fewer years of professional experience, and female, while prior involvement in complaints was not the main differentiating factor [25]. These findings do not establish a Europe-wide pattern, but they support the psychiatric claim that liability pressure may alter ordinary clinical reasoning before litigation occurs, particularly in decisions about admission, discharge, risk assessment, compulsory care, and documentation.

The legal literature often assumes that human oversight resolves accountability. The AI Act similarly requires meaningful human oversight for high-risk systems [15]. In practice, formal oversight does not guarantee substantive independence. Work on automation bias and selective adherence shows that human decision-makers tend to over-rely on automated outputs when systems appear technically advanced, carry institutional endorsement, or operate in time-pressured settings [26]. Direct psychiatry-specific evidence remains limited, but general automation-bias research and the defensive-practice literature discussed above suggest that uncertainty and concerns about defensibility may intensify deference to automated outputs. A clinician faced with a probabilistic risk score or large language model (LLM) summary may outsource doubt to the machine. While AI promises consistency, opacity surrounding model development, validation, and domain transfer can make challenging its conclusions difficult. Psychiatric professionals may remain formally responsible while becoming practically deferential, creating a misalignment in which liability remains human while practical epistemic authority shifts toward the system [14,27-29]. Here, authority means the weight given to the output in reasoning, not a legal transfer of decision-making power. Reliance may be more difficult to evaluate than reliance on a validated scale or a colleague’s advice when the model is opaque, dynamically updated, institutionally mandated, or difficult to challenge. The appropriate response is not a symbolic human-in-the-loop requirement, but a standard of human-in-reasoning: the substantive work of understanding the output’s intended use and limits, testing it against the patient and context, and recording reasons for reliance or departure. Clinicians must be able to interpret outputs sufficiently to assess their fit with the patient, context, and legal stakes. Institutions should require active articulation of reasons when clinicians follow or override high-stakes AI outputs, especially in admission, discharge, compulsory care, and suicide-risk assessment. AI explainability need not entail full technical transparency, but psychiatric systems should provide intelligible reasoning, confidence limits, performance boundaries, and warnings when extrapolating from weakly matched data; otherwise, clinicians may not have sufficient or adequate outputs to support interpretation [4,15,28,30,31].

Responsibility refers to the substantive duties and causal contributions of actors; accountability refers to the mechanisms through which they must explain, document, audit, and answer for those duties; and liability refers to the legal consequences and remedies that may follow after harm. The terms overlap but are not interchangeable. Theoretical accounts of AI liability in medicine often distinguish between 3 models. Under a reliance model, clinicians are encouraged to rely on certified systems and may be partly shielded when doing so reasonably. Under a deviance model, clinicians remain fully liable because AI is treated as a tool. Under a neutrality or no-fault model, compensation is socialized through insurance or collective funds, reducing the plaintiff’s burden to prove individual fault [14]. Each model captures an important dimension, but none is sufficient for psychiatry. A pure reliance model risks amplifying automation bias and diminishing the clinician’s duty to interpret the patient. A pure deviance model places disproportionate responsibility on psychiatrists while allowing developers and institutions to externalize risks associated with opaque systems they control. A pure neutrality model may improve access to compensation but risks underdeterrence unless combined with robust regulatory enforcement and recourse mechanisms. At present, claimants may face substantial costs in overcoming opacity and producing evidence of wrongdoing, breach of duty, causation, and harm. Psychiatric practice therefore calls for a layered model of accountability and a specific liability framework for AI.

Responsibilities are allocated according to the degree of control over the relevant source of risk and the capacity to prevent or mitigate foreseeable harm. Actors should primarily be accountable for aspects of AI-assisted care that they design, introduce, govern, or directly influence. This functional allocation reflects well-established principles in product liability, organizational governance, and professional negligence, while recognizing that psychiatric AI operates through sociotechnical systems rather than isolated individual decisions. Responsibilities therefore follow control and preventability rather than professional status alone. Developers, as they control system architecture, validation, documentation, intended use, and foreseeable misuse, should be responsible for design defects, inadequate validation, misleading performance claims, cybersecurity failures, and insufficient safeguards against foreseeable misuse. Deployers, typically health care organizations or institutions responsible for procuring and implementing AI systems, should be accountable for procurement, local validation, workflow integration, staff training, monitoring, and escalation procedures because they determine how the technology enters and functions within clinical services. Clinicians should remain responsible for informed, context-sensitive judgment, avoiding uncritical reliance, and documenting reasons in high-stakes cases because they retain immediate responsibility for applying AI outputs to an individual patient within a particular clinical context [27,30]. Table 1 proposes the main liability models in health care AI and their psychiatric implications.

Because fault-based liability is largely governed at the national level, the European legal and practical landscape remains uneven. The following examples are an illustrative mapping of different governance trajectories among countries represented by the authors or commonly discussed in European digital health governance: strategy-led implementation, centralized supervision, human-primacy legislation, evidence-based reimbursement, guidance-led trust governance, and sandbox preparation. Portugal illustrates a strategy-led approach. The country approved a National AI Agenda for 2026 to 2030 through Council of Ministers Resolution No. 2/2026 of January 8, but has not enacted a comprehensive standalone AI liability regime. The existing Center for Responsible AI has no regulatory or oversight capacity. Neither the Charter of Human Rights in the Digital Era nor the designation of ANACOM (Autoridade Nacional de Comunicações) as the single point of contact for the EU AI Act has addressed AI liability in psychiatric care. In health care, this means continued reliance on general health law, tort law, data protection law, medical device law, labor law, and incoming EU obligations. This offers flexibility but limited certainty in psychiatric malpractice scenarios involving AI-assisted care [37]. Spain has moved earlier toward institutional supervision through AESIA (Spanish Agency for the Supervision of AI), created under Royal Decree 729/2023 (August 22, 2023). This may support administrative oversight and coherent implementation of EU requirements, but ordinary negligence law must still determine reasonableness when clinicians follow, ignore, or depart from AI recommendations [38,39]. Italy has adopted one of the clearest human-primacy positions. Law No. 132/2025 (September 23, 2025) frames AI in health care as auxiliary rather than autonomous, prohibits discriminatory conditioning of access to health care, and affirms human responsibility for clinical decision-making. For psychiatry, this resists attributing autonomous clinical authority to machines, but may leave clinicians carrying extensive responsibility unless recourse against defective systems or negligent deployment is available.

Germany offers a different model through the Digitale Gesundheitsanwendungen framework. Digital health applications may be reimbursed under statutory insurance when they meet regulatory and evidentiary criteria, including demonstration of a positive health care effect. Although Digitale Gesundheitsanwendungen does not primarily address AI liability, it shows how evidence, reimbursement, and regulation can shape the practical standard of care [40]. France emphasizes governance and trust through the Haute Autorité de Santé Initial Key Guiding Principles for the Use of Generative AI in Health Care and the Health Data Hub, stressing appropriate use, verification, and organizational responsibility [41,42]. The Netherlands illustrates supervisory preparation of the AI Act implementation through sandbox approaches and continued reliance on civil code doctrines. These trajectories show convergence on governance but divergence on civil liability. For psychiatry, the same AI-supported act may face similar ex ante duties across Europe but different ex post litigation risks. Table 2 shows selected European jurisdictions toward AI governance.

The following 4 use cases are illustrative rather than exhaustive. They were selected to represent common or rapidly expanding psychiatric AI domains. Each tests a different legal category: simulated therapeutic interaction, risk prediction, data-driven surveillance, and clinical record production. These technologies should not be treated as a single regulatory class. Patient-facing conversational agents may operate with limited clinician mediation; suicide-prediction systems function as clinical decision support; passive monitoring produces continuous inferential data; and documentation assistants are primarily administrative tools with indirect but potentially important clinical effects. Their intended purpose, degree of autonomy, and position in the workflow affect classification and liability exposure. The first is the conversational or therapeutic chatbot. Responsible AI principles focused solely on fairness, autonomy, and robustness may be insufficient when the clinical issue is relational: dependency, vulnerability, emotional attunement, and meanings generated through interaction must also be considered [23]. Chatbots may appear empathic and continuous without clinical responsibility or capacity for care. Reviews show rapid expansion, heterogeneous purposes, variable evidence, and uneven safeguards, while recent analyses warn about role confusion, dependence, and mismatches between the perception of empathy and accountable care [11,12,32,43,44]. When systems approximate therapeutic dialogue, crisis response, or management of vulnerable users, some duty of care becomes harder to deny [10,13,19,34]. The second use case is suicide prediction. AI-based models may outperform unaided clinical judgment in some settings, but reviews emphasize heterogeneity and difficulty translating prediction into ethical intervention [45]. Studies using electronic health records and social media content illustrate improved stratification, but also show how prediction can become detached from actionability, liberty, and fairness [46-49]. False negatives may produce underintervention; false positives may increase coercion, stigma, and deprivation of liberty. Risk scores should not become warrants for restriction [45,50]. The third use case is digital phenotyping and passive monitoring. Systems that infer mood, relapse, cognition, or adherence from smartphones, wearables, language, or behavioral traces rely on data that may be noisy, socially patterned, or weakly contextualized. Patients may not understand the extent of inference, and services may lack capacity to respond to alerts. Reviews stress the gap between proof-of-concept performance and safe clinical translation [3,6,8]. The fourth use case is LLM documentation and summarization. Psychiatric records shape continuity of care, risk assessment, legal processes, insurance decisions, and patients’ self-understanding. If AI compresses nuance, overstates certainty, omits qualifiers, or translates narratives into confident diagnostic shorthand, downstream care may be distorted. These systems should be evaluated for documentation fidelity, calibration, and downstream effects, not only fluency or time saving [28,29,51]. When appropriately validated and integrated, these tools may improve access, identify clinically relevant patterns, support continuity, and reduce documentation burden. The purpose of the liability analysis is to enable those benefits through safer use rather than to argue against adoption. Table 3 shows psychiatric AI use cases, liability risks, and standard-of-care implications.

Justified integration refers to the reasoned incorporation of AI into psychiatric decision-making through critical clinical interpretation rather than routine acceptance or automatic rejection of algorithmic outputs. We adopt the Integrated Ethical Approach for Computational Psychiatry proposed by Putica et al [27] because it is specifically designed for mental health AI and integrates the principal ethical domains that are repeatedly implicated in psychiatric AI, namely beneficence, autonomy, justice, privacy, transparency, and scientific integrity. These principles provide the ethical rationale for justified integration while remaining distinct from legal duties. In this Viewpoint, it supports the ethical content of justified integration but is not treated as a source of European legal duties. AI should inform psychiatric reasoning without replacing it. The standard of care should neither require routine acceptance of AI outputs nor treat AI use as inherently suspect. Under justified integration, psychiatrists should know the intended purpose and limits of the system, assess whether the patient resembles the population on which the tool was developed, consider whether the output fits the clinical picture, discuss AI involvement when material to consent or trust, and document reasons when high-stakes decisions rely on or depart from AI outputs. This approach aligns the augmented clinician model with the AI Act’s emphasis on meaningful human oversight, while translating oversight into a clinically operational duty [4,15,29,30]. Table 4 relates the clinical meaning of requirements for the reasoned integration of AI in psychiatric practice.

A further implication concerns the nature of evidence. Evidence for AI should not be limited to aggregate predictive accuracy; it should include subgroup performance, false positives and false negatives, user comprehension, impact on clinician behavior, impact on coercive interventions, and effects on therapeutic trust. A suicide-prediction tool that modestly improves discrimination but increases unnecessary coercion in specific populations may be clinically and legally inferior to a less accurate model deployed more cautiously. Likewise, an apparently effective chatbot may perform well on symptom reduction while generating role confusion, emotional overdependence, or unsafe crisis responses. Although European law does not yet mandate all such metrics, professional guidance and institutional governance should increasingly consider them when defining reasonable psychiatric use [29,43,44]. Implementation must be proportionate to clinical risk and workflow. Low-risk administrative tools may require verification and periodic audit, whereas systems influencing diagnosis, suicide risk, admission, discharge, or compulsory care require formal training, local validation, escalation procedures, and protected time for review. Institutions should avoid converting human oversight into unresourced documentation: concise templates, audit sampling, and clear ownership of alerts can reduce burden. Smaller services may require shared procurement expertise, regional validation support, or phased adoption rather than placing technical governance on individual clinicians.

The central discussion is not whether AI should be used in psychiatry, but how responsibility should be allocated when AI becomes part of psychiatric judgment. A layered accountability model is preferable because it matches the sociotechnical structure of AI-assisted care. Developers primarily control design, validation, performance claims, technical documentation, cybersecurity, safeguards, and foreseeable misuse. Deployers primarily control procurement, local validation, workflow, training, escalation, monitoring, and audit. Clinicians control contextual interpretation, explanation to the patient, assessment of clinical coherence, and documentation of reliance or departure. These boundaries are functional rather than absolute, and mixed-fault cases may involve more than one level. This model avoids treating the psychiatrist as the default bearer of all digital risk, while preserving the clinician’s duty to reason with the patient rather than through the machine. Such a model also clarifies the relationship between law and ethics. The law must determine fault, causation, defect, disclosure, and compensation after harm. Ethics must preserve the conditions under which psychiatric care remains a relationship of interpretation and moral responsibility. In psychiatry, these 2 tasks are closely connected. A legal framework that encourages passive deference to AI would weaken care even before litigation arises. Conversely, a framework that places all risk on clinicians may discourage useful tools or incentivize defensive documentation. A defensible European approach should therefore combine preventive regulation, product liability, national malpractice law, professional guidance, and institutional governance rather than relying on any single mechanism.

Mixed-fault scenarios are particularly important. If a hospital mandates use of an opaque suicide-risk tool without local validation, the main failure is unlikely to be the psychiatrist’s ordinary clinical judgment alone. If a vendor markets a chatbot as supportive while designing it to maximize emotional engagement among vulnerable users, the central issue may be product design, misleading claims, or foreseeable misuse. If an LLM documentation system subtly changes the evidential tone of notes, responsibility may be shared between the developer that produced insufficient safeguards, the institution that integrated the tool without audit, and the clinician who failed to verify a high-stakes record. Conversely, when a transparent, validated, locally governed tool is used as intended, the clinician remains responsible for contextual interpretation rather than mechanical obedience. Layered accountability therefore does not dilute responsibility; it locates responsibility at the level where control and preventability were greatest.