Borderline personality disorder (BPD) is associated with significant functional impairment and profound distress [1]. It is characterized by intense, unstable interpersonal relationships marked by alternating idealization and devaluation, an unstable sense of self, affective instability, and a heightened risk of self-harm and suicidal behavior [1]. Yet access to specialist treatments such as dialectical behavior therapy (DBT) is limited [2], and there are implementation challenges [3,4]. Earlier digital interventions for BPD, predating AI-CA and commonly based on DBT or CBT, have shown mixed efficacy, with meta-analytic benefits for suicidal ideation and paranoia but not for overall symptom severity [5]. These DBT- and CBT-based tools nonetheless outperformed alternative approaches for suicidal ideation [5], suggesting that grounding digital support in an evidence-based framework matters. Interest in AI-enabled support for mental health is now growing and could help to extend evidence-based treatment to more people, though practical and ethical concerns complicate implementation [6,7]. For people with BPD, who experience interpersonal challenges, self-harm, and high risk for suicidal behavior [1], it is essential that we understand how to integrate AI into BPD treatment responsibly. This study provides an in-depth account of how DBT psychologists anticipate, imagine, and evaluate possible AI-CA integration within a DBT framework for BPD, which may also inform care for other populations with similar clinical needs.

Standard DBT is an evidence-based treatment for BPD [8], delivered through individual therapy, group skills training (including homework assignment and review), phone coaching, and therapist consultation team [9,10]. These four modes serve DBT’s core functions: building patients’ skills, strengthening motivation to change, promoting generalization of skills to daily life, structuring the environment, and sustaining therapists’ capability and motivation [9,10]. DBT assumes that many difficulties reflect skills deficits [9,10]. Across treatment, therapists balance acceptance and change while helping patients act from Wise Mind, the integration of emotion mind and reasonable mind [9,10]. DBT combines validation of patients’ responses as understandable in context with contingency management to reinforce adaptive behavior [9,11]. Phone coaching supports real-time skills generalization before self-destructive behaviors occur [12,13].

For example, a patient who has just received a poor exam grade calls with high urges to self-harm. The therapist first validates the distress (“you worked so hard on that exam, and now you are anxious about failing the course”), then coaches a specific plan for the rest of the evening, and, when risk warrants it, also helps the patient reduce access to means. This real-time intervention reduces the risk of self-harm while helping the patient plan to cope with pain in more adaptive ways. Yet phone coaching is among the DBT components most challenging to implement and the least frequently delivered [3,4]. Perceived provider burden is a frequently cited reason for omission [14]. These implementation difficulties provide a clinical rationale for considering AI-CA in DBT, particularly for hard-to-sustain functions such as between-session skills generalization.

AI-based conversational agents are increasingly being investigated as potentially scalable tools for mental health support, although evidence for their effectiveness and implementation in routine care remains limited [7,15,16]. Conversational agents relevant to mental health vary both in technical architecture and primary design intent. Architecturally, they range from rule-based systems using predefined scripts or decision trees, through hybrid systems combining natural language processing or machine learning with rule-based constraints, to generative AI systems, often large language model (LLM)-based, that produce more open-ended and contextually responsive dialogue [16-18]. In terms of design intent, agents may be developed specifically for therapeutic or wellness support [16], function primarily as social or relational companions [19], or be general-purpose generative AI systems that are not designed for mental health but are nevertheless used for emotional or psychological support [20]. We use the term AI-CA to refer only to hybrid and generative (LLM-based) AI agents with therapeutic intent; where evidence concerns the generative AI (eg, ChatGPT [Open AI]) specifically, we use the term generative AI-CA. While evidence on AI-CA in mental health care remains limited overall [17,21], research specifically addressing its use in BPD or within DBT frameworks is, to our knowledge, absent. This may reflect iatrogenic-risk concerns and the ethical, legal, and practical challenges of managing self-harm and suicidality in this population [22,23].

How patients and professionals perceive AI-CA is central to responsible implementation, as multiple factors at both the patient and provider level shape the acceptability and adoption of conversational agents in health care [24]. Several studies have examined how users perceive generative AI-CA and AI-CA for mental health support [25-28]. Common findings indicate that AI-CA are perceived as accessible, helpful, and nonjudgmental, but also limited by a lack of emotional depth, difficulties with complex situations, and an inability to actively guide the therapeutic process. Clinicians, in turn, tend to favor administrative uses of AI, such as documentation, while showing uncertainty toward client-facing applications, though openness may increase with greater exposure [29].

The relevance of AI-CA to DBT cannot be evaluated only by asking whether AI can produce empathic, coherent, or clinically plausible responses. DBT therapists respond strategically, reinforcing skillful behavior; that is, responding so that the behavior becomes more likely to recur. Whether a behavior is skillful is not given by its topography (the observable form of the behavior, described independently of its consequences) but by its function (the effect it produces in a given context) [9,11]. Determining this requires attention to the context, frequency, and duration of the behavior, to the patient’s current capabilities, and to the patient’s short- and long-term goals. Calling for support during an intense urge to self-harm is skillful for a patient who cannot yet tolerate the emotion alone. The same call becomes a safety behavior when the patient is no longer seeking coaching but repeated confirmation that nothing bad will happen. Safety behaviors are actions taken to prevent a feared outcome that is in fact unlikely; they relieve anxiety in the moment, which maintains them, while preventing the patient from learning that the feared outcome does not follow [30].

Two features of AI-CA are relevant against this background. First, generative AI-CA tends toward sycophancy, affirming users’ statements and self-image rather than challenging them, a tendency linked to reduced prosocial intentions and increased dependence [31]. An agent that responds affirmingly to whatever the patient brings may therefore reinforce what treatment targets for change, for example, by supporting avoidance where exposure is indicated, or by confirming a patient’s belief that they cannot cope. The concern is not merely hypothetical: studies of suicide-related AI responses show inadequate crisis responses, limited contextual understanding, and a need for safety evaluation [22,32]. Second, unlike human therapists, AI-CA respond immediately and on demand. This lowers the threshold for seeking support, which may aid skills generalization but may equally sustain safety behaviors and reduce opportunities to tolerate distress or learn skills more independently. AI-CA have also been found to facilitate self-disclosure by reducing perceived fear of judgment, although in nonclinical samples [33]. How a constantly affirming and always-available agent affects patients who already may struggle with relationships, boundaries, and expressing their own wants and feelings remains an open question.

There is a need for understanding how AI-CA might be responsibly integrated into DBT, particularly given limited access to specialist BPD treatments such as DBT [2] and the potential for AI-CA to support between-session skills practice and generalization. However, methodological guidance on how to integrate AI into psychotherapy remains underdeveloped [7], and the perspectives of method-specific clinicians who would be responsible for such integration have received limited attention. Because the adoption of AI-enabled health care tools depends on stakeholders’ perceptions and acceptability [24], qualitative methods are well suited to capture how clinicians make sense of risks, responsibilities, and fit within clinical practice. Since AI integration in DBT is not yet established, participants’ accounts are treated as anticipatory and interpretive rather than experiential reports of implemented practice. Such expectations are analytically relevant because emerging technologies are shaped not only by technical capability, but also by professional imaginaries, anticipated risks and benefits, and role negotiations before implementation occurs [34]. Therefore, this study aims to explore how DBT clinicians in Sweden understand the anticipated integration of AI-CA into DBT for individuals with BPD, specifically regarding the relational positioning of AI-CA and perceived suitability of DBT functions or tasks for AI-CA involvement.

This qualitative study used semistructured interviews to explore DBT clinicians’ expectations about the imagined integration of AI-CA into DBT. Data were generated and analyzed using reflexive thematic analysis (RTA) within a constructivist epistemological framework [35], in which knowledge and meaning are understood as actively constructed through the researcher’s engagement with the data, rather than as preexisting patterns awaiting discovery. RTA within a constructivist framework was chosen because the knowledge sought is interpretive rather than observational. The first author’s dual positioning, as a DBT clinician and as a developer of an AI-CA prototype, was treated as an analytic resource rather than a bias to be neutralized, enabling recognition of treatment-specific meaning in participants’ accounts. Consistent with a constructivist RTA, the findings are understood as constructions shaped through dialogue between participants’ accounts, the interview framing, and the research team’s clinical and theoretical positioning [35]. The study is reported in accordance with the Standards for Reporting Qualitative Research (SRQR) checklist (Checklist 1)[36].

This study was approved by and followed guidelines from the Swedish Ethical Review Authority (Diary Number: 2024-03541-01). All participants gave written informed consent, participated voluntarily, and did not receive compensation. To protect privacy, data were pseudonymized, removed identifiers before publication, and stored securely by the research team in compliance with the General Data Protection Regulation (GDPR).

Seventeen psychologists were purposively recruited by emailing DBT teams in publicly funded psychiatric outpatient services in Västra Götaland and Skåne, 2 of Sweden’s largest health care regions, serving approximately 3.2 million inhabitants combined. Inclusion criteria were at least 1 year of experience working with DBT and being a licensed psychologist. In Sweden, psychologist licensure requires a 5-year MSc in Psychology and 1 year of supervised clinical practice. DBT training typically comprises about 10 days over 1 year, considered the minimum team-based standard for delivering DBT in Sweden. Fifteen of the seventeen participants worked in standard 4-mode DBT programs; the remaining 2 in programs without telephone coaching. Participant characteristics are summarized in Table 1.

The interview guide provided a common structure oriented toward DBT-specific implementation questions, including relational positioning, responsibility, treatment functions, and BPD-specific concerns (Multimedia Appendix 1); it was used flexibly, with follow-up probes pursuing clinically and analytically salient issues. Attention to anticipated AI-CA capacity to understand reinforcement contingencies developed early, after participant 3, and informed subsequent interviews. Participants were asked to consider a not-yet-implemented, DBT-integrated AI-CA whose autonomy and role were left open for them to position themselves.

The mean interview length was 51 (SD 9.9; range 36-80) minutes. The participant group comprised 13 women and 4 men, with a mean age of 43 (SD 10.3, range 24-60) years and a mean of 6.4 (SD 5.6; range 1-18) years of DBT experience. No participant dropped out after agreeing to take part. One pilot interview was conducted to refine the interview guide and was included in the study. Three participants from the first author’s DBT team were included but interviewed by coauthor LA, who has no connection to the participants, and the rest by the first author. Interviews were conducted via secure video (12 of 17) or, when it was possible, in person from January 28 to June 3, 2025, at the workplaces of participants. Information power was assessed reflexively throughout the process, and the 17 interviews provided rich and diverse data that enabled deep engagement with the research questions [37]. This decision regarding the number of participants relied on the study’s narrow scope (AI integration in DBT), the highly specific participant group (DBT psychologists), and the rich quality of the dialogue, although no established theory was applied and the analysis was cross-case, which may have lessened the information power of the participant group. Recruitment was guided by information power rather than data saturation and ended when the focused aim, specific participant group, and rich interview material were judged sufficient to address the research questions. Data saturation was not applied, in accordance with RTA methodology [38].

The analysis followed Braun and Clarke [35] 6-phase framework. The process was recursive, moving back and forth between phases as patterns of meaning were developed and refined. The first author conducted the primary analysis, with 2 team discussions to explore and challenge interpretations of codes and themes, and more frequent analytic consultations with supervisors (SS, LA, and SIL). Analytic discussions functioned as critical reflexive dialogue to challenge and deepen interpretations rather than to establish coding consensus. The initial focus on power, responsibility, and potential iatrogenic effects was gradually refined into a more nuanced concern with relational dynamics and the difficulty of determining the function of AI-generated responses within a therapeutic context. Later stages involved refining, defining, and naming themes, and then producing the final report.

The first author became familiar with the data by underlining transcripts and listening to each interview at least 3 times. Each iteration clarified the rationale for asking certain questions while omitting others, for example, questions concerning the management of suicide risk. Early analytic coding used Post-it notes and mind maps, which were iteratively clustered into candidate patterns of meaning (see Multimedia Appendix 2 for process documentation). Visualization made it easier to see overarching patterns of meaning and form themes. Later, NVivo 15 (Lumivero) was used to manage the dataset; however, final coding and theme development were conducted in a spreadsheet format, as this afforded greater flexibility for the recursive, interpretative work of collapsing, splitting, and renaming codes in line with RTA. We did not aim for coder consensus or interrater reliability. A reflexive diary was also used, primarily before and after interviews, to document affective responses such as discomfort or enthusiasm and to help distinguish researcher reactions from participants’ meanings.

Generative AI tools (ChatGPT, Claude) were used only for language editing and manuscript organization, not for data analysis, coding, or theme development; no interview excerpts or participant material were entered into them. Two authors (SS and LA) have previously coauthored work on a separate mental health chatbot prototype (BETSY) for non-DBT contexts [39,40]. This shared research trajectory likely sensitized the analytic dialogue to questions of codesign, feasibility, and acceptability of AI-CA, and was kept in view during reflexive discussions. All interviews were transcribed verbatim in Swedish by the first author. Analysis was informed by the first author’s clinical training in DBT and undergraduate grounding in both psychodynamic and cognitive behavior theory. Behavioral-theoretical concepts such as reinforcement contingencies, the dialectical tension between acceptance and change [9], and relational or projective sensitivities functioned as interpretive lenses rather than deductive frameworks; no predefined codebooks were used.

In line with RTA, the first author’s positioning was considered throughout. A licensed psychologist with over a decade of DBT experience, including work with suicidal patients, the first author is also developing a noncommercial AI-CA prototype within the doctoral project. This dual role, as clinician and as developer of the technology under study, sensitized the analysis to questions of risk, clinical responsibility, and implementation constraints; a fuller reflexive account appears in Multimedia Appendix 2.

The results are organized around three main themes (Table 2): (1) “Who Are We in Therapy?,” on the relational positioning of AI; (2) “The Stoic Helper,” on functions seen as suitable for AI support within DBT; and (3) “The Well-Intended Accommodator,” on the difficulty of reliably determining whether an AI response supports change or reinforces maladaptive behavior (AI functional ambiguity) and the dependency and responsibility concerns this raises. Across themes, participants often held conflicting views, moving between confidence and uncertainty about whether and how AI should be implemented. Throughout, “AI” refers to AI-CA as defined above.

Taken together, clinicians’ views were structured by relational positioning, functional hopes, and ethical concern rather than by simple approval or rejection. Figure 1 presents a conceptual synthesis: theme 1 frames themes 2 and 3, which form a dialectical pair in which the same feature that makes AI a “stoic helper” can make it a “well-intended accommodator,” the juncture at which functional ambiguity arises. These positions sometimes coexisted within a single account, for example: P4 hoped AI could relieve after-hours telephone coaching yet feared it might reinforce a patient’s harmful “ways of thinking,” while P8 saw AI as a possible “bridge” for relationally avoidant patients yet doubted it could grasp the function behind their words. The three themes were developed from recurring patterns across interviews, not from single cases. However, some positions were more widely shared than others. Across interviews, participants commonly discussed availability, boundaries, skills support, roles, dependency, and responsibility. More specific ideas, such as AI-CA as a supervisor or competitor, were less frequent and are presented as illustrative examples rather than prevalence claims.

This theme centers on relational dynamics rather than technical capabilities, exploring how AI is defined and positioned in relation to both therapist and person in treatment, and which relational capacities it may or may not possess compared with human therapists.

The availability and relational aspects of AI are problematized in terms of what they do to the therapeutic relationship, and that unlimited access is not always an asset or necessarily desirable.

Expanding on this, a participant reflected on the universal human need and capacity for relationships, attachment, and forming relational bonds; “we are made to form relationships. I mean, I just ask ChatGPT about rabbit poop and I want to write thank you. Even though I get that there is nobody to thank” [P5]. Participant [P17] takes this somewhat further by considering what this relational capacity may mean for the individual who forms an attachment: “I think that the risk that it [AI] is experienced as supportive without being there” [P17]. These accounts are read as suggesting that AI can evoke relational attachment patterns, for both patients and therapists alike, where boundaries between therapeutic and personal relationships are less clear.

AI’s position in relation to the therapeutic frame shaped what it was seen as allowed or expected to do, with participants describing AI variously as a tool, a team member, or as having the characteristics of a supervisor.

Given that DBT is a team-based treatment, several participants suggested that AI could be understood as part of that team, also suggesting that in some respects AI may help therapists make better decisions, much like a supervisor would.

How AI is positioned here also shapes which qualities it is expected to possess, those a good colleague or supervisor would need in order to perform such tasks.

Many participants described availability as important for patients. However, this participant points to how it could also matter for therapists. This account is interpreted as suggesting that, from a therapist’s standpoint, the risk of fostering dependency is less obvious and is instead framed as a question of whether AI’s competence can be “trusted.”

One way of understanding this is that there is an ambiguity around what AI could or should do. AI was also attributed considerable competence here, alongside a skepticism about even entertaining that possibility. On the other side of the relational spectrum, some participants felt more comfortable with a more passive or dependent AI that carried out tasks delegated by the therapist, like an assistant.

But somewhat later in the interview the participant expanded further:

These accounts can be read as AI being easier to construct as less threatening when positioned as a tool or more dependent on the therapist, which may function as a response to perceived uncertainty and a sense of losing control and ambivalence regarding the role or function of AI in therapy.

Several participants constructed AI as something that should have boundaries in its contact with the patient, like those of a human therapist or team member.

In this example, the participant accepts that AI is ’setting a boundary’ with the patient, which can be read as positioning AI as more than a tool, as something closer to having a subject. This could be interpreted as making two claims at once: that AI is something to be respected, or that patients with relational difficulties need to treat AI ’as if’ it were a person in order to learn new behaviors.

Below, a question is raised about what happens when, or if, the trust between AI and patient is high.

The therapist and AI are situated as competing for the patient’s trust on the basis of perceived competence and adaptability, with the therapist being diminished within the therapeutic relationship.

A participant reflects on the timing of AI introduction, similar to participant [P2], suggesting that the therapist and patient may first need space to establish a relational connection before AI is incorporated into therapy: “In the orientation phase when you are supposed to form some kind of alliance and trust. All of those parts feel very strange that you would bring in AI there” [P11]. The participants frame the solution as establishing a therapeutic alliance before introducing AI, without in this instance, considering that the therapeutic relationship often fluctuates over the course of therapy.

This theme addresses how AI is relationally defined and where its boundaries are drawn, which shapes what participants saw as possible in team-based DBT.

AI was understood as “a stoic helper” that was able to provide support at times when human therapists either could not or preferred not to be present, such as after hours or on weekends. More concretely, the stoic helper was exemplified by participants in performing difficult-to-sustain therapeutic DBT functions. These could be skills generalization, psychoeducation, help with homework, and between-session availability and telephone coaching. When AI is described as performing these functions, it is interpreted as implying that AI would possess characteristics of being available, competent, adaptable, and tireless.

Many participants thought about telephone coaching:

Here it is understood that AI could become an “AI therapist,” with less focus on why the patient should not call the human therapist, or what consequences calling AI instead might have. The next excerpt highlights telephone coaching as a burden, while also conveying hopes that AI could share some of this load. “In DBT treatment I probably perceive it as a good complement. Maybe even some form of relief regarding the phone coaching specifically” [P16]. Expanding on the capacity to stay with the patient in a wider range of ways: “that it [AI] doesn’t get tired and stuff, that is a possible advantage. Some patients maybe need to hear something very many times” [P14]. AI is interpreted as possessing therapeutic qualities in areas where human therapists may sometimes fall short. Neither participant acknowledged that they themselves might become dysregulated during telephone coaching. This participant further reflects on telephone coaching: “To use as phone coaching for example when we are not available. Because it’s often at night that they can need phone coaching and then we are not available” [P4], interpreted as expressing a therapeutic wish to be present in moments when human availability is inevitably limited and needed.

Below, expanding on contacting AI instead of the human therapist, the contact between therapist and patient is not itself problematized. Instead, AI is positioned as a possible solution to what could be constructed as a shared problem within the therapeutic relationship.

Analytically, the participant can be interpreted as hoping that a frustration in therapy, patients not reaching out, could be addressed by introducing an AI-CA that would always be available and could not be burdened. Expanding on this further, the next participant mentioned: “Sometimes we give really super complicated homework assignments, I think. And then the patients just stand there and just, “I have no idea [how to do the homework]” [P12]. AI was interpreted as competent and benefiting not only patients but also therapists, as the examples participants referred to often concerned situations they themselves experienced as difficult to manage. These accounts framed AI as a response to patient needs and pointed to strain points in DBT delivery that clinicians may experience as difficult to sustain, without directly relating this to their own limited resources.

Another participant reflected on AI’s relational capacities, interpreted here as a concern that AI might compete with the therapist for the therapeutic relationship, which needs to be protected.

However, this is understood as suggesting that, if introduced the “right” way, AI could be seen as aligned with the therapist, an extension of the therapy rather than a separate, competing entity.

A substantial component of DBT involves skills learning and generalization. The nonhuman aspects of AI were in this account read as being beneficial since it removes a layer of relational complexity for some patients.

AI was interpreted as more adaptable and competent than human therapists and could provide a higher degree of individualization when teaching skills in treatment, and points to that AI is experienced differently than a human therapist.

This participant’s account expands on skills training further and could be read as expressing frustration with limits within the therapy setting and the potential of AI in this regard.

AI is interpreted as having capacities distinct from those of human therapists and as compensating for therapist difficulties, or as being better positioned to teach skills because it is not constrained by time or group-based settings.

How a therapeutic situation can be understood or analyzed varies by its nature, and more than one answer can be “right.”

The participant is interpreted as suggesting that AI could broaden the perspective in therapy and give the patient more options rather than just listening to the therapist’s voice.

This theme addresses AI’s anticipated functions in DBT and the opportunities its accessibility offers, often at points clinicians find difficult to manage but essential to treatment.

Participants raised concerns that on the surface helpful AI behavior may reinforce maladaptive behavior, foster dependency, and operate as a safety behavior (well-intended accommodation).

The boundary between helpful support (such as skill generalization) and maladaptive accommodation is difficult to determine in practice as understood in this account:

The same participant further reflected on technology in a broader societal context and what skills we might lose or not learn or internalize:

While also situating themselves within the same context as the patients, but more on a personal level, not professional. Instead of auxiliary ego, another participant used external knowledge, which here is understood as constructed as the same concept:

Here AI is positioned more as something “outside” therapy, hindering learning or growth instead of helping or facilitating it as seen in theme 1. Not mentioned is that the therapist is also external to the patient and someone you could become unhealthily dependent on.

Several participants raised the concept of Wise mind [9,10], which is constructed as the opposite of an external auxiliary ego and connects the effects of AI to the ability to trust one’s sense of inner wisdom or certainty.

The participant reflects on how AI might influence our inner wisdom, interpreted here as a concern about how AI may shape identity, particularly for individuals whose sense of self is more vulnerable to external influence, which AI may have considerable capacity to exert.

Here, the participant highlighted both the potential benefits and drawbacks of increased accessibility.

This participant is understood as both pointing to tension between risks and benefits and situating themselves within the same sphere of vulnerability as the patient, similar to [P5]. Also read in this account is the notion of responsibility and using AI to learn or using AI instead of learning.

The next participant reflected on what therapeutic work involves and how AI currently seems to operate, thereby indirectly raising questions about its potential consequences.

The participant is understood as pointing to therapeutic qualities seen as problematic in AI, while also suggesting that patients should, to some extent, be directed within the therapeutic process, not just led. The notion of mirroring could here be viewed as being sycophantic, and therapeutic solutions are given instead of being sycophantic within treatment, which AI currently might struggle with doing.

Here, P4’s reflections are understood as similar:

In these accounts, both AI and patients were interpreted as limited in their capacity to know what is needed therapeutically, whereas the therapist was constructed as the one better able to determine this.

Here, the participant articulates that a therapist does not merely listen to what the patient says but attends to the function it serves within a specific context.

The participant questioned how effectively AI could accomplish the task of determining the function of behaviors, given its potential inability to challenge the patient or move beyond the surface of the written text, which expands on P3’s mirroring or P4’s concerns about AI learning to promote harmful behaviors.

Furthermore, another participant is understood as expressing concern about affirming responses or beliefs that, although understandable, do not fit the situation or work against the person’s therapeutic goals.

Taken together, these accounts suggest an anticipated concern that the same AI response may be experienced as validating in one context yet invalidating in another, and participants did not expect AI to reliably distinguish between the two.

Furthermore, participants expressed concerns about how responsibility for treatment would be managed.

The participant is understood as contemplating how ruptures might be handled in therapy when an AI-CA is involved, perhaps more so when working with suicidal clients. The next participant is also read as implying that contact with patients is not predictable and you need to be able to adjust and that therapists, in the end, need to maintain oversight, since AI cannot be trusted in this context, “I think that’s hard to hand over this [suicide assessment] to an AI because it’s so much... tone of voice and like knowledge of patients and experience as a therapist of many similar situations and stuff. I think that’s hard to teach to an AI” [P14]. This could also be exemplified with “You still need a person who can check what it says” [P15]. All these accounts are read as concerns around how AI-CA responds in high-risk situations.

This theme concerns challenges to DBT adherence and whether AI-CA can address BPD-specific considerations, particularly their ability to understand patients’ behavior in clinically complex contexts.

This qualitative interview study of DBT psychologists’ perspectives on AI-CA developed three themes: Who Are We in Therapy? (theme 1), The Stoic Helper (theme 2), and The Well-Intended Accommodator (theme 3). Theme 1 examined how participants relationally positioned AI-CA, which in turn shaped how they imagined its practical integration into DBT. Theme 2 reflected how clinicians imagined AI-CA as compensating for strain points or helping in DBT delivery. Theme 3 highlighted an anticipated central implementation challenge, AI functional ambiguity, that is, the difficulty of determining whether an AI-CA response is helpful or harmful in a specific context. Taken together, our analysis suggests that participants’ views of AI-CA appeared closely tied to their understanding of their own role and to the boundaries of what AI-CA should or could do. Figure 1 is a conceptual overview intended to reflect how these themes relate to each other. In the figure, how AI-CA is positioned relationally (theme 1) shapes what it is expected to do (theme 2) and the risks this raises (theme 3); themes 2 and 3 are linked by AI functional ambiguity.

A qualitative study found that therapists conceptualized AI as a “bridge” or complement to therapy, supporting accessibility and structured tasks rather than the relational core of treatment [41]. In this current study, participants positioned AI-CA in more than a single role. Mapping these accounts onto a behavioral intervention technology continuum [42], they range from therapist-led use with AI-CA support to AI-CA-delivered care under therapist oversight, near but never fully automated use.

When participants imagined an AI-CA in DBT with relational capabilities, they thought this might affect the alliance, a question that predates generative AI [43]. Recent AI-CA studies suggest that users may form alliance-like experiences with AI agents [16,44], and perceived alliance has been linked to greater engagement and symptom improvement [44]. Qualitative research with psychotherapists suggests that integrating AI-CA into psychotherapy raises questions about professional role boundaries, ethical responsibility, and the preservation of the human therapeutic relationship [45]. However, these studies examined the patient’s experience of the AI-CA, rather than how an AI agent affects an ongoing therapist-patient relationship. To our knowledge, direct evidence remains limited on alliance when AI, patient, and therapist are integrated within a psychotherapy model such as DBT.

AI-CA have been proposed to support several parts of the suicide prevention pathway, such as disclosure, risk detection, crisis signposting, and clinician-assisted triage [46], which would be important if AI-CA were integrated into DBT. Responses to suicide-related queries appear to have improved over time [32]. Autonomous risk assessment and crisis management remain insufficiently validated and may be unsafe, especially for high-risk or relationally vulnerable groups [19,22], which several participants in this study raised concerns around. Given the AI-CA’s tendency toward sycophancy and the fact that patients with BPD often present with self-destructive and suicidal behaviors, where it is essential not to reinforce maladaptive behavior, the management of anticipated AI functional ambiguity becomes a central clinical concern.

AI functional ambiguity is related to but structurally distinct from safety behavior, a patient-enacted behavior such as seeking reassurance from an AI-CA. AI functional ambiguity, by contrast, is not a patient behavior but a condition pertaining to AI responses in therapeutic interaction. It is the condition under which it cannot be reliably determined whether a given AI response supports treatment-consistent change processes or inadvertently undermines them. The concerns expressed by participants under “The Well-Intended Accommodator” align with an emerging empirical literature on LLM sycophancy, in which interaction with sycophantic models has been shown to reduce prosocial intentions while paradoxically increasing user trust and willingness to continue using such models [31]. Within DBT, validation is defined as communicating to the patient that their responses make sense and are understandable given their history and current life context, while explicitly not endorsing what is maladaptive [9]. DBT operationalizes this through the therapeutic frame, in which the therapist tries to understand each behavior in the patient’s whole context. A generative AI-CA may instead default to sycophancy, for example, by delivering the reassurance a patient seeks, and so reinforce the patient’s safety behavior. Whereas sycophancy is an AI response whose therapeutic function has already been classified as unhelpful, AI functional ambiguity is the condition in which such classification is not reliably available. What distinguishes the concept is that its concern lies one level higher, at the point of functional classification itself.

Clinicians did not test an AI-CA, a deliberate choice: any demonstrated model would quickly become outdated and could steer participants’ narratives, and a brief demonstration would capture only a narrow aspect of AI-CA in DBT. In addition, the interview guide framed AI- and therapist-generated responses as potentially hard to distinguish, an assumption supported by prior research [47,48]. Because the interview framing positioned AI-CA as a potential adjunctive tool or support, and at times as resembling an additional therapist, relational issues such as alliance, role boundaries, and responsibility may have been made more salient during interviews. We therefore treat the themes as reflexively developed interpretations shaped by the dialogue between participants’ accounts, the interview framing, and the research team’s DBT-informed clinical perspective. We collected no standard measures of computer or AI-CA use to characterize the participants. DBT program fidelity and individual therapist adherence were not formally assessed. Some participants’ limited AI-CA experience may also have shaped the range and depth of their reflections.

Transferability should also be considered in relation to the EU AI Act, given that participants were situated within this legal context. To our knowledge, no AI-CA is currently implemented in publicly funded DBT. Under current EU and Swedish legal systems, this appears to be complicated [49], which may have shaped participants’ AI-CA expectations. Another more permissive legal setting might have generated different themes. Swedish survey data describe a public that adopts technology early yet holds cautious, mixed AI attitudes [50], possibly shaping participants’ baseline orientation toward AI-CA. Because the sample was predominantly female and drawn from a caring profession, transferability to more gender-diverse clinician groups should be considered with caution, particularly for findings concerning care, emotional availability, dependency, and AI-CA’s relational role. The themes should not be read as gendered, as no gender-comparative analysis was conducted. The participants’ shared DBT frame of reference is, by contrast, possibly a less context-sensitive frame of reference. Participants worked in large public regional services rather than private settings, likely treating patients with substantial psychiatric burden. Because DBT is team-delivered, a participant group including nurses, social workers, or counselors might have surfaced concerns about liability, professional identity, and scope of practice absent from this psychologist-only group. Purposive recruitment may have shaped which views became most visible, particularly among clinicians with more positive or critical views of AI-CA. While telephone coaching practices can vary across DBT programs and health care systems [3], and most of our participants worked in programs that included telephone coaching (15 of 17), which may have shaped their perspectives.

Participants stressed establishing a patient-therapist alliance before AI becomes involved. They also emphasized making AI less directive later in treatment, so that responsibility returns to the patient. In our view, this implies that the role of AI-CA may need to vary across treatment phases and with patient need. Greater therapist oversight seems warranted when patients present high-risk behaviors, positioning AI-CA as more adjunctive and less autonomous. Future research should develop methods to examine AI functional ambiguity across roles and therapeutic modalities. A key design and research problem is whether AI-CA responses can distinguish skills coaching from reassurance-seeking, including whether explicit therapeutic context reduces maladaptive AI responses and supports responses that balance validation with change, such as prompting skills use. Research should also examine how patient and therapist can converge on the AI’s role, function, and limitations, and how that agreement may need to be adapted across therapy to what both require.

In this study, 17 DBT psychologists were interviewed to understand the anticipated integration of AI-CA within a specific treatment context, DBT, and for a clinical population, individuals with BPD. In the accounts we developed, AI-CA was constructed not merely as a tool but as introducing a relational position within therapy that likely needs to be addressed explicitly. The hopes participants attached to AI-CA may be understood as a response to demands within therapy itself; imagined uses appeared shaped not only by technological possibility but also by clinicians’ frustrations, wishes, and unmet needs. In the end, the same difficulty therapists face may also apply to AI-CA, most notably what we conceptualize as AI functional ambiguity. This is the risk of not knowing whether responses such as validation foster dependency, function as safety behaviors, or help patients move forward in treatment. Such a development raises familiar therapeutic questions in a new form, including how to maintain shared goals, adherence to specific treatment methods, and clarity around responsibility.