To rectify the aforementioned circumstances, we propose two courses of action. First, there must be a clear linguistic differentiation and reorganization of the concept of DTA relative to TA. Second, we propose the introduction and differentiation of interaction levels with AI in therapeutic contexts. The following presentation outlines our novel concept of the “digital therapeutic nexus” (DTN).

Our proposal aims to elucidate that the relationship between humans and AI systems constitutes a discrete type whose characteristics cannot categorically be compared with those of human-human relationships, especially in a therapeutic context. Within this framework, the term “nexus” signifies a connection between entities of different natures [45]. In this instance, it is used to denote the connection between human and artificial entities. This conceptual framework enables a novel and autonomous examination of the relationship between humans and AI systems. This enhanced understanding facilitates the identification of potential risks and the development of effective solutions.

The conceptual framework of DTA with chatbots entails a distinct understanding of TA as a harmonious entity, wherein patients and users assess the machine’s empathy, while the AI manipulates the user along the “reward function” [46]. Therefore, the DTN differs from DTA in its recognition that a genuine therapeutic relationship also encompasses conflicts, contradictions, and ruptures.

To operationalize this distinction, we propose the DTN as a conceptual framework for characterizing the user-AI relationship in digital therapeutic contexts. The DTN builds upon the stages of integration described by Stade et al [47], inspired by the autonomous-vehicle industry, which describes the AI-human therapist relationship (assistive, collaborative, and fully autonomous). However, although their work focuses on the human therapist as an expert mediator of AI integration in clinical contexts, the DTN shifts the focus to the direct relationship between AI and the user, which also evolves through 3 levels.

A transactional, knowledge-based relationship focusing on symptom understanding, psychoeducation, or the provision of instructions or guidance. Examples of interactions at this level of therapeutic nexus include step-by-step instructions for a brief breathing exercise, communication of the best strategies to adhere to a sleep hygiene plan, or a user asking questions such as “What are the early signals of a nervous breakdown?” with the system, similar to a search engine, simply providing a list of symptoms.

The system acts more empathetically, is responsive, starts to ask follow-up questions, makes suggestions, and begins to cocreate with the user. An example of a user question at this level would be, “I can’t sleep well since my divorce. I feel everything is falling apart, and now I can’t even focus at work. I’m afraid I’m making everyone around me sick of me; they want to leave me, too.” The system’s response would offer general advice, show emotional validation and constant empathy, provide support, and list possible courses of action that could alleviate psychological distress. The system would consistently avoid confrontation and opt for pleasing answers. This could also lead to stagnation if the user needs therapeutic challenges. This is also the level of highest risk, as, at this point, the constant flattery and sycophancy of the system may have helped create attachment in the user, and it is at this stage that self-destructive behaviors can be validated and promoted by the system.

A dynamic, autonomous, and symmetrical interaction (symmetrical interaction is understood here as a reciprocal, nonhierarchical exchange) in which the system challenges, reflects, and promotes psychological growth. It is worth noting that, to reach this level, the AI system must already have attained the third stage of integration described by Stade et al [47], meaning that it is fully autonomous and capable of full-range clinical interventions without human supervision. Only at this stage of the therapeutic nexus can we begin to draw parallels with the TA characteristic of human-human interactions. An exemplary response to the situation presented in level 2, instead of offering general advice, would explore the client’s specific experience of divorce, history of relationships, and current support structure to co-design a plan for sleep and work performance that fits the client’s specific life circumstances. Furthermore, there would be continuous monitoring for potential misunderstandings to handle latent ruptures by repairing and renegotiating tasks and goals. A comprehensive collection of examples of each DTN level is presented in Figure 1.

We argue that, given the ontological differences and the need for functional autonomy (functional autonomy is understood here as the capacity to perform interventions without direct oversight by a human clinician) to challenge and repair ruptures in the relationship, parallels with human TA are only sensible at level 3. It follows that, prior to level 3, the core constructs of the TA operate in qualitatively distinct ways: bond is better understood as perceived alignment, goal agreement as goal mirroring, and tasks as compliance-driven adaptation rather than as jointly endorsed practices. It is worth noting that DTN levels are intended as descriptive categories, a map of the current technological landscape rather than a normative road map. Importantly, level 3 is not an arbitrary threshold; its defining features are derived from what human therapists already do in practice. By this account, current systems can only be classified as level 1 or 2, and level 3, while not yet achieved, represents a foreseeable trajectory given that some developers are already working to address key limitations such as sycophantic behavior [34,48]. How far current systems fall short of level 3 is illustrated by empirical comparisons; for instance, in a recent comparison of text-based responses from human psychotherapists and chatbots, the chatbot responses were rated as significantly more empathetic and used supportive interventions more often. In contrast, human responses were more likely to engage in exploration and discussion of maladaptive patterns and to ask more clarification questions [49]. Similarly, psychotherapists have also stressed chatbots’ failures to ask for further information, including biographical information, the presence of suicidal thoughts, or other symptoms [50].

This finding highlights a critical limitation; currently, chatbots lack the necessary capacities for a truly transformational interaction (DTN level 3), which involves a therapeutic challenge, rupture repair, and agreement on goals. Thus, as depicted in Figure 2, the DTN levels differ not only in the interaction type they involve but also in their actual contribution to the therapeutic process. Specifically, the informational and, especially, the supportive levels may provide a degree of comfort but essentially leave the client in the same position, whereas the transformational level actively contributes to the client’s progress toward goal attainment.

Regarding transitions between DTN levels, it is important to recognize that, similar to human-human relationships, such transitions are often not sharply delineated. Higher DTN levels may temporarily incorporate behaviors characteristic of lower levels without this constituting a regression. For example, a level-3 interaction that shifts toward validation during a rupture remains level 3, given that the dominant pattern remains transformational. Similarly, a level-2 interaction that includes a list of instructions remains level 2 if those instructions are embedded in a seemingly empathic and validating context.

It is also important to note that the categorization of interactions into a DTN level should be based on the dominant pattern observed across multiple interactions. As with scales describing human-human relationships, classification accuracy increases with the number of observed sessions. To clarify these transitions as much as possible, the principal operationalization criteria are summarized in Table 1.

Because meaningful parallels with human-human therapeutic relationships can be achieved only at DTN level 3, we argue that measuring the DTA and its impact on therapeutic outcomes is appropriate only once the interaction has reached this level. Below level 3, the foundational conditions for a TA, including genuine agreement on goals and tasks, cannot be established, since such agreement requires prior challenge or disagreement rather than mere compliance or mirroring. Therefore, we recommend using the DTN as a classificatory step before any alliance-based measurement: only interactions successfully characterized as level 3 would be candidates for measurement using adapted TA scales. For this purpose, the mobile Agnew Relationship Measure appears particularly promising, as it does not simply replace the term “therapist” with “app” but undertakes a detailed adaptation of the traditional alliance scale by removing items, adding new relevant ones, and demonstrating promising psychometric properties [32,33]. A systematic comparison of the key conceptual differences between the DTA framework and the DTN is presented in Table 2.

We conclude that the full potential of “digital therapy” will not be fulfilled with systems that only agree. In psychotherapy, change often involves challenging the client and confronting maladaptive behaviors or beliefs, and is consolidated through the process of rupture repair. However, to reach level 3, it is indispensable for AI systems to have undergone an extensive expert clinical supervision and to have clear risk criteria that allow the determination of the circumstances under which AI systems can confront users or intervene for safety reasons.

The premature translation of concepts specific to human therapy into digital practice and the precipitate labeling of AI systems as therapeutic agents pose significant clinical risks, the intensity of which varies with the DTN level. One of the most serious concerns is the capacity of sycophantic systems to validate, perpetuate, or even promote risky and destructive behaviors such as suicidal ideation, homicidal ideation, or abuse [51]. This risk could be particularly high in clients with narcissistic tendencies, whose altered cognitions are especially susceptible to uncritical reinforcement. At level 1, the lack of emotional involvement does not eliminate this risk; a system functioning as a search engine can still provide content that facilitates harmful behaviors, but to a lesser extent than at level 2, where attachment and validation take place without therapeutic challenge. At level 3, similarly to what occurs in competent human-driven therapy, this risk is substantially decreased as long as the system has been properly developed, but it does not entirely disappear.

An additional concern is algorithmic bias; given that large language models are trained on human-generated text, it is well known that they tend to reproduce and perpetuate social biases and different forms of discrimination [52-54]. This risk is particularly pronounced at levels 1 and 2, where system responses are generated without a deep knowledge of the client’s personal history and are therefore less sensitive to the client’s subjective life experience. At level 3, the integration of personal context would likely attenuate this risk, although it might not eliminate it.

A third important risk is developing a parasocial relationship or even dependency. This risk would be lowest at level 1, given the lack of personal or affective exchange. At level 2, however, the risk is highest due to the system’s constant availability and unconditional validation, which can give users a false sense of mutual connection. As noted in the Introduction section, human-AI friendships focus more on availability and self-disclosure than on reciprocity, similarity, or intimacy, which we argue could be extrapolated to therapeutic relationships. Because of this imbalance, users might perceive real relational value in what is actually a 1-sided interaction. This could lead them to prefer the system over human support networks, not because it is more effective but because of emotional convenience. We predict that this risk would be lower at level 3, although it would not disappear entirely, as users may still project relational meaning onto the interaction in ways that parallel transference in human therapy.

A hard question follows: If robots begin to challenge, will users accept it, or dropout? Without a robust bond (and clear consent frameworks), challenge may be perceived as noncompliance rather than as care. This acceptance problem is central to any claim that robots can form alliances analogous to those formed by human therapists.

In this paper, we have mainly focused on unembodied systems, such as chatbots, as they constitute most of the research on digital therapy. However, the rapid development of embodied social robots presents new challenges. These systems, by incorporating increasingly sophisticated anthropomorphic characteristics, may soon allow humanlike physical interaction. This not only opens promising opportunities to strengthen the therapeutic bond but also introduces significantly higher risks, especially in sensitive clinical contexts. Therefore, progress toward more autonomous and embodied systems calls for the development of clear ethical protocols that effectively regulate the use of mechanisms such as protective override. This type of intervention, in which the system can ignore or contradict user preferences to protect users, must be clearly delimited, agreed upon, and continuously supervised.

Finally, fellow researchers are encouraged to use the DTN framework and validate our classification checklist across diverse therapeutic domains, clinical populations, and care settings to evaluate its robustness, transferability, and sensitivity to contextual variation. Clinicians are advised to implement the proposed 2-step classification approach when assessing digital tools for integration into clinical workflows, thereby supporting informed decisions regarding therapeutic roles, autonomy levels, and associated risks. Concurrently, developers are welcome to use the DTN levels as design benchmarks to guide the development of human-AI therapeutic interactions that align with clinical objectives, ethical standards, and appropriate system agency.