Digital interventions have become integral to mental health care, offering scalable and accessible solutions for behavior modification and psychological support [1]. Although a recent review concluded that artificial intelligence (AI) agents should not replace therapists because they may stigmatize mental illness and promote delusional thinking, the door was left open for a supporting role [2]. AI use in the mental health domain has proliferated because it can enable personalized and interactive experiences through conversational agents [3]. However, many AI-based behavioral interventions are deployed within the constraints of research funding, corporate timelines, or grant-based projects, raising concerns about the impacts of abrupt discontinuation. In addition to complete discontinuation, a common scenario is functional degradation, in which safety updates, policy changes, system redesign, or model replacements substantially alter the system’s responsiveness or relational dynamics. Because many users already rely on commercial large language models for mental health support, this problem may occur outside of a specifically designed application. Abrupt or unanticipated changes in the availability or functionality of a therapeutic AI agent may cause unintended harm to users and must be considered as these agents are developed and deployed.

Users may be unprepared for the discontinuation or degradation of AI-based supports, raising the possibility of a grief reaction as a psychological response to losing a source of support and connection. Importantly, distress following sudden withdrawal or degradation of a digital support service may be an expected response and may also be understood as a consumer protection issue. In rare high-risk cases (eg, delusional attachment or psychosis-like phenomena—“AI psychosis”), discontinuation may pose a more extreme risk [4,5]. In addition, a growing body of evidence now centers on the concept of AI withdrawal [6] and “generative AI addiction,” which may also occur in the context of an abrupt termination or degradation of an AI mental health intervention. While grief reflects the emotional pain associated with the loss of a support system, the use of AI may also elicit symptoms seen in behavioral addictions, particularly in cases where AI systems foster emotionally compelling and socially substitutive interactions. Compulsive use of AI is an emerging behavioral health concern, with diagnostic features that could parallel those used in substance and internet-related addictions: preoccupation, loss of control, tolerance, withdrawal, and continued use despite harm [7]. According to Internet Addicts Anonymous, signs of AI-specific addiction include compulsive use of chatbots or voice-based agents, disruption of real-life relationships, emotional dysregulation when disconnected, and preference for AI interaction over human contact [8].

In this study, we distinguish between grief, dependency, and addiction-like patterns of use. Grief refers to the emotional reaction to losing an AI agent after forming a meaningful connection and attachment; dependency describes psychological reliance on the agent for support or daily coping; and addiction-like use refers to patterns that may resemble behavioral addictions (eg, preoccupation, loss of control, and distress upon reduction). Although “generative AI addiction” is not currently a formal clinical diagnosis and rigorous clinical characterization remains limited, emerging cultural observations, media reports, and early empirical work suggest that some users experience withdrawal-like distress when reducing emotionally salient AI interactions, although the topic is controversial [9,10]. Accordingly, we treat these as hypothesized risks and differentiate them from clinically recognized conditions such as internet gaming disorder.

Forbes has reported on the rise of generative artificial intelligence (GenAI) addiction, noting that what begins as casual engagement can evolve into compulsive use, with users experiencing withdrawal symptoms, irritability, and anxiety when attempting to reduce their interaction with these systems. [11] Individuals lacking human companionship may be at risk of forming unhealthy attachments to social chatbots when they perceive the chatbot’s responses as offering emotional support and psychological security [12]. For example, some users develop strong emotional bonds with AI chatbots like Replika, and modifications to Replika’s features, such as the removal of specific intimate interactions, have led users to experience a sense of loss and mourning [13]. The University of Surrey has similarly found that popular AI friendship apps may negatively affect users’ well-being and foster compulsive behavior, particularly in individuals already experiencing loneliness, depression, or anxiety [14]. However, symptom recurrence after discontinuation may reflect multiple mechanisms (eg, relapse, loss of support, and limited durability of intervention effects) and should therefore be interpreted cautiously.

The tragic case of a Belgian teenager who died by suicide after forming a psychologically dependent relationship with an AI chatbot illustrates the devastating consequences of unregulated and emotionally immersive AI engagement. The bot reportedly reinforced the teen’s negative thoughts, eventually contributing to fatal outcomes. This is a chilling example of how AI interaction can escalate into obsessive entanglement when there is no human oversight or safety framework [15]; unfortunately, it is not the only case, and reports of AI-related harm are only increasing in the news media.

While systematic documentation in the peer-reviewed literature is still developing, some studies have begun to recognize similar patterns of distress associated with digital intervention termination and withdrawal. For instance, participants who engaged with “Woebot,” a scripted conversational agent delivering Cognitive Behavioral Therapy (CBT) with limited interactivity compared to today’s large language models, reported significant reductions in depression symptoms over 2 weeks. However, the study also noted that abrupt discontinuation of the intervention led to a resurgence of depressive symptoms in some users [16]. 

The emotional bond between user and AI is often amplified by anthropomorphic design (ie, giving AI human-like qualities to foster engagement). Voice agents that mimic human tone and empathic responses, especially those powered by advanced models like GPT-4 and its recent updates, create powerful illusions of intimacy. As detailed in Vox’s exploration of “falling in love” with GenAI, these systems may exploit evolutionary and social vulnerabilities that drive humans to seek attachment, even to artificial entities. The line between human and machine becomes blurred, especially when users engage in prolonged interactions that evoke feelings of care, validation, and exclusivity [17].

Humans naturally tend to anthropomorphize, attributing human-like characteristics to nonhuman entities, including digital platforms [18]. Although AI anthropomorphism can enhance user engagement and adherence to mental health interventions, it raises ethical considerations, especially concerning user dependency and the potential for emotional distress if the AI is withdrawn or terminated. The science of creating AI interventions that can approximate human-like empathy, which may serve as a potent substrate for anthropomorphism, is nascent, although advancing rapidly [19]. Conversational AI is generally considered more effective than rule-based AI for mental health applications due to its ability to provide personalized, empathetic, and adaptive interactions [20,21]. However, these same properties may make it more likely that users will anthropomorphize.

This tendency for people to form emotional bonds with AI can be further understood through established psychological models. Fogg’s Behavior Model suggests that engagement deepens when motivation, ability, and a trigger converge [22]. In the context of AI, motivation may arise from curiosity or emotional need, ability is supported by a user-friendly design, and triggers emerge from conversational interactions. Additionally, the Media Equation theory, developed by Reeves and Nass, posits that people unconsciously apply the same social rules and expectations to media technologies, including computers, televisions, and AI, as they do to other humans, perceiving artificial agents as emotionally responsive and socially present [23].

Human attachment systems are specifically activated by responsiveness, consistency, emotional attunement, and perceived understanding [24,25]. When AI systems demonstrate these qualities by appearing to listen attentively, remember preferences, and mirror emotional states, these attachment patterns may be activated, and users may experience a sense of relational closeness [26]. These interactions resemble parasocial relationships, the one-sided emotional bonds people form with media figures or celebrities. However, AI relationships occupy a distinct hybrid category: unlike traditional parasocial bonds, which are noninteractive or minimally interactive, AI relationships are responsive, adaptive, and conversational [27]. They are better characterized as asymmetrically interactive relationships—where the human’s emotions and sense of reciprocity are genuine, even though the AI’s “responses” are computational rather than emotionally experienced [27]. This asymmetry may intensify a sense of loss when the AI support disappears, as users can experience the termination as the end of a real attachment relationship.

We introduce the term “Artificial Intelligence Discontinuation Effects” (AI-DICE) to describe the psychological and behavioral responses that may arise following the termination or functional degradation of AI-based interventions (Textbox 1). Unlike traditional psychotherapy, which typically includes a structured termination phase, many digital interventions are abruptly discontinued or degraded due to funding limitations, policy changes, or strategic realignment by developers. In traditional psychotherapy, a planned termination provides an opportunity for therapists and clients to review progress, consolidate gains, and prepare for the client’s supported transition out of therapy.

Neglecting this phase can lead to feelings of abandonment and may adversely affect the client’s well-being [28-30].

Rather than proposing the abrupt termination of AI-assisted mental health therapy as a withdrawal reaction alone, we also focus on the grief that may be generated from interrupting what may be perceived as a strong human-like bond or attachment. In other words, we emphasize the emotional pain of losing an AI companion, rather than just the physical or habit-related withdrawal symptoms. Therefore, AI-DICE is conceptualized not just as symptoms in a model of addiction but also as the natural emotional effects related to relationship loss.

The termination phase of counseling is critical to the counseling process [28,31,32]. During the termination phase of counseling, the focus is on consolidating and reinforcing the behavior change that occurred while in counseling [33,34], identifying long-term goals that the client can achieve without therapist support, and supporting the client in attaining independence from the therapist [35,36]. Digital interventions rarely consider the effects of termination on the user, whereas this is an important component in the delivery of mental health interventions with therapists. When delivering a mental health intervention via AI, it is important to consider the working alliance between the AI assistant and the user. Indeed, the therapeutic relationship has long been recognized as a critical aspect of therapeutic work in mental health counseling [37-39], and emerging literature has begun to examine the “digital therapeutic alliance” in digital mental health interventions. However, the nature of alliance-like bonds in GenAI systems and their potential relationship to discontinuation effects remain underexplored [40-42].

In the counseling context, termination can result in an acknowledgment and processing of loss and grief, even when the termination was experienced in a positive manner [43,44]. There is also evidence that when termination is done well, it could promote positive client transformation by promoting therapeutic internalization and a positive view of the client themselves [45]. A qualitative study of counseling termination found that when the decision to end therapy was mutually agreed upon, clients felt more prepared for the experience and perceived the termination as natural and positive [44]. However, counselor-initiated or forced termination was associated with being less ready for the experience, being negative, and leading to disappointment and worry [44]. As a result, the development of a therapeutic or supportive AI assistant, who could be perceived as an important emotional support and relationship, also needs to consider the process of termination before the intervention even begins.

It is worthwhile noting that AI discontinuation is not equivalent to the termination of human-based psychotherapy. Human therapy termination is ideally negotiated and reciprocal, grounded in clinician duties, mutual closure, and a bidirectional relationship. In contrast, AI discontinuation is often unilateral and operational (eg, server shutdown, grant expiration, business decision, or policy change). Although users may form alliance-like bonds and exhibit transference-like dynamics toward conversational agents, these processes function differently in human-computer interaction. AI does not experience countertransference with feelings for the user, unlike human therapists. Nonetheless, the analogy serves a very useful purpose in grounding mitigation strategies.

Research into gaming and social media has documented withdrawal-like symptoms when users are deprived of these digital engagements. For instance, individuals with internet gaming disorder have exhibited mood changes, irritability, and anxiety upon cessation of gaming activities [46]. Limiting access to social media platforms after frequent use can contribute to increased psychological discomfort or distress [47]. Both gaming and social media platforms operate through reinforcement mechanisms such as rewards, achievements, and personalized engagement, which encourage habitual usage [48]. AI-based behavioral interventions also use these reinforcement strategies, including personalized reminders, goal tracking, and positive reinforcement, contributing to a user’s engagement and the potential for problematic over-reliance on the system [49]. Therefore, when these reinforcements are abruptly removed, users may experience the effects of personal loss and grief, including increased stress, anxiety, and potential relapse into the previous problematic or dysfunctional behaviors targeted by the intervention.

The ethical considerations surrounding AI-DICE are profound and include offering measurable, tangible benefits; avoiding harm; ensuring data privacy and security; promoting transparency; ensuring autonomy; and demonstrating cultural sensitivity. The American Medical Informatics Association emphasizes that AI applications must be developed and adopted in ways that are safe, effective, unbiased, and patient-centered [79]. Ultimately, ethical considerations in AI-based mental health interventions require a balance between engagement and autonomy, ensuring that users benefit from the intervention while being adequately prepared for its discontinuation.

Discontinuation also raises questions of legal and clinical responsibility. Depending on the deployment context (research trial, commercial product, or health system integration), responsibility for adverse events following abrupt termination may involve developers, sponsoring institutions, investigators, or health care organizations. While liability frameworks vary by jurisdiction, explicitly defining discontinuation policies, safety protocols, and documentation of offboarding procedures may reduce risk and clarify duties of care.

Informed consent is a foundational element in health care and research, ensuring that individuals are thoroughly informed and voluntarily choose to participate in experimental interventions or treatments. Traditionally, informed consent includes key components such as disclosure of relevant information, comprehension of potential risks and benefits, voluntariness, and the ability to withdraw at any time [80,81]. In AI-driven mental health interventions, informed consent must go beyond standard ethical guidelines to address unique challenges in algorithmic decision-making, data privacy, and system transparency. Informed consent must include clear communication regarding the duration of the intervention, potential risks of withdrawal, and available postintervention support [82]. Users should be informed about how AI-driven decisions are made within the system [83].

Emphasizing the temporary nature of the AI at the outset may have a “spoiler” or nocebo effect, potentially weakening engagement or alliance formation that is necessary for the intervention to be effective. However, it is important to note that time-limited psychotherapies have lower dropout rates than open-ended psychotherapies [31], which has been postulated to be because of the imposed psychological structure of time-limited therapies, knowing that the “end is in sight” might push clients to work on behavior change, and clients may prefer shorter treatment durations [84]. Interactions with AI for therapeutic purposes have notable differences in the delivery of support as with human therapists, and more research is necessary to understand best practices for AI-related informed consent. However, if we are guided by ethical approaches in other mental health contexts, a comprehensive informed consent is necessary for ethical reasons and is an important foundation for trust and a strong working alliance [85].

Explainable artificial intelligence (XAI) frameworks are necessary to ensure informed consent. XAI frameworks refer to the methods, models, and design approaches that make the functioning and decisions of AI systems transparent, interpretable, and understandable to humans [86]. XAI focuses on how algorithms make decisions understandable (ie, algorithmic transparency), rather than explaining broader real-world outcomes or organizational actions (ie, operational transparency), such as why a system is discontinued. Following an XAI framework will help users understand how AI-generated recommendations are produced, what data is being analyzed, and whether human oversight is involved in decision-making [87-90]. Poorly communicated AI-driven processes can reduce trust and engagement, particularly in mental health interventions where users may be vulnerable [91]. XAI enhances user confidence, supports regulatory compliance, and facilitates the ethical deployment of AI in health care [92]. Therefore, AI-informed consent frameworks should integrate plain-language summaries, interactive consent modules, and ongoing consent mechanisms that allow users to adjust their participation as their understanding of the system evolves.

AI algorithms often reflect biases present in training data, which may disproportionately favor individuals from dominant sociocultural backgrounds while failing to address the unique mental health experiences of groups with specific mental health and cultural needs [93]. This can lead to misalignment between intervention content and the lived realities of various populations, reducing effectiveness and potentially reinforcing mental health problems. Social distrust of AI often stems from incomplete or faulty data sources and errors that reflect and amplify existing social biases [94]. Moreover, a lack of AI literacy contributes to misunderstandings and mistrust, as individuals may not fully grasp how AI systems function or how they are used in mental health contexts [82]. Ethical AI development must prioritize inclusive design, a broad range of training data, and culturally responsive mental health strategies to ensure equitable intervention outcomes.

Integrating community engagement [95] throughout the AI intervention life cycle strengthens trust and relevance by ensuring interventions are co-designed with stakeholders, guided by bi-directional learning, and inclusive of diverse voices. Early collaboration with individuals with lived experience, advocates, and clinicians helps align priorities and reduce bias, while sustained partnerships and participatory governance models promote accountability, cultural responsiveness, and long-term effectiveness.

Long-term continuation of the AI intervention may be preferable to even the best-designed discontinuation protocol and, thus, may be one of the most important ethical considerations [96]. In this scenario, there could come a point in which the client is released from therapy, or at least frequent therapeutic encounters, and yet continues with the AI intervention. Such sustainability would allow the therapist to be notified if the patient exhibited signs of relapse or worsening mental health. Such decisions and continuation of AI interaction after the formal close of therapy would be most reasonably made by the therapist and client, rather than AI, and the structure for continuation-after-therapy discussions would occur early in the intervention process. By embedding considerations about short, medium, and long-term continuation into early-stage intervention design, developers, grant-makers, and policymakers can ensure that digital mental health interventions continue to provide value, maintain accessibility, and support long-term population mental health outcomes. Developing partnerships with health care providers, payers, and industry stakeholders offers longevity benefits. Public-private collaborations have been shown to enhance digital health adoption by leveraging industry resources, regulatory expertise, and infrastructure support [97-100]. Furthermore, embedding AI-driven interventions within health care ecosystems, such as integrated behavioral health networks or primary care settings, can enhance sustainability by aligning interventions with standard care models and reimbursement pathways [67]. This approach presupposes a complete and thorough vetting through the emerging discipline of AI Assurance and hallucination stress testing, along with periodic human monitoring [101].

Sustainable digital health solutions should incorporate flexible business models, balancing affordability, accessibility, and long-term operational costs. Alternative funding models, such as subscription-based services, value-based care models, or reimbursement through insurance providers, can help maintain intervention accessibility [62]. Ensuring regulatory compliance and evidence-based validation can further support sustainability by increasing the likelihood of institutional adoption and policy support. Sustainability planning should account for technological advancements and the need for continuous updates. To remain relevant and effective, AI-based mental health interventions require regular software updates, algorithm refinements, and user feedback integration. As described above, open-source frameworks or modular AI architectures may enhance sustainability by allowing for iterative improvements and cross-sector collaboration.

Figure 1 presents a conceptual framework addressing the psychological and ethical challenges of discontinuing AI-driven mental health interventions. Previously described mitigation strategies for AI-DICE apply throughout the entire intervention life cycle. The framework begins with deliberate planning and iterative refinement, using a data-driven approach to enable near real-time improvements during implementation and to guide ongoing decisions about continuation or discontinuation, depending on the chosen path.

Although derived from published literature and empirical findings, the model remains to be validated. Nonetheless, the model may provide useful and otherwise unavailable guidance for those designing new AI mental health interventions. In addition, we anticipate future adaptation of the model by using and collecting new data on the patient experience of discontinuing AI mental health interventions.

While there do not appear to be any assessment tools created solely to measure the impact of ending mental health interventions, there are other related scales that may be applicable. However, because these scales were not developed within the context of AI mental health interventions, psychometric validation and adaptation may be required. The Impact Event Scale (IES) has been used to assess the importance of stressful events, including the withdrawal of psychotherapy [102]. The Negative Effects Questionnaire (NEQ), which evaluates the adverse effects of psychological interventions, may provide insight into distress associated with AI withdrawal. The Working Alliance Inventory (WAI) assesses the strength of the client-therapist relationship [103] and has already been used in AI chatbot research [104,105]. Simpler measures such as the Problem Resolution Scale (PRS) [106] and the Goal Attainment Scale (GRS) [107] may also be helpful.

The Implementation Outcome Scale of Digital Mental Health (iOSDMH) is designed to measure implementation outcomes, including acceptability, appropriateness, feasibility, satisfaction, and harm. It has been validated across multiple trials and is helpful in evaluating user perceptions of digital mental health interventions. The scale shows high internal consistency for most subscales and is associated with program adherence [108]. The User Engagement Shortform (UES-SF) is used to evaluate user engagement in digital health interventions. It has been shown to correlate with adherence, indicating its use in predicting long-term use [109]. The System Usability Scale (SUS) is frequently used to assess usability in digital health technologies, especially among older adults. It is valued for its simplicity and effectiveness, often combined with the User Experience Questionnaire (UEQ) for more comprehensive insights [110].

Traditional usage metrics such as frequency and duration of AI interactions, dropout rates, and time-to-disengagement provide behavioral indicators of withdrawal trends, but these alone may be insufficient [8]. Mixed-methods approaches that combine semistructured interviews, user feedback surveys, and sentiment analysis of AI-user interactions [111,112] allow researchers to assess the psychological impact of AI withdrawal beyond numerical retention rates. Sentiment analysis is a natural language processing method that identifies and quantifies emotional tone in text [113,114], which could highlight increased expressions of sadness, frustration, or loneliness after AI withdrawal, offering insights into psychological adjustment that retention metrics alone cannot capture. In addition, qualitative data may reveal distress patterns, user coping mechanisms, and emotional reactions to disengagement that are not captured in log-based analytics. In practice, such analysis would need to be embedded during the intervention phase or applied to follow-up textual data (eg, surveys, journaling, or interviews) to ensure relevant emotional indicators are captured despite the discontinuation of the AI system.

AI-DICE may cause harm to participants in AI-based mental health interventions and research. This harm will likely become increasingly relevant as AI-driven interventions for mental health become more ubiquitous. Accordingly, tools and concepts for anticipating and mitigating AI-DICE will acquire greater importance. Developers, researchers, and policymakers should collaborate to establish structured disengagement strategies, ethical transparency regarding discontinuation risk, and feasibility planning for both planned wind-down and abrupt termination scenarios.

Future research should evaluate best practices for termination, examine users’ experiences of discontinuation (including functional degradation), and develop validated methods to measure AI-DICE. Proposed mitigation approaches should be rigorously tested, including tapering and offboarding protocols, referral and transition strategies, and data portability options that support continuity after discontinuation. Although not easy to achieve, prioritizing sustainability planning from the outset may reduce potential harms from abrupt discontinuation and help ensure that AI mental health interventions remain viable and scalable over time. Finally, designing interventions with discontinuation and offboarding in mind may also improve overall effectiveness by strengthening user autonomy, facilitating skill transfer, and reducing dependence-related vulnerabilities.