Dull, dirty, and dangerous facets of human-AI collaboration for psychotherapy

1 Introduction

Driven by rapid advancements in capability, the imminent if not extant ubiquity of artificial intelligence (AI) is apparent across a range of contexts. From healthcare to education, there is significant optimism surrounding the transformative impacts of AI integration with its potential to improve safety, productivity, and creativity. One report estimates that AI will have a global market size of $4.8 trillion by 2033, representing the largest share of any frontier technology (Tou and Vezzani, 2025). This optimism for AI has extended to psychotherapy contexts that see an increasing gap between needs and available services (Mongelli et al., 2020; Dehbozorgi et al., 2025; Giotakos, 2025). Proposed use cases even include using conversational AI as self-directed therapists (Grodniewicz and Hohol, 2023; Tekin and Delehanty, 2026), though many applications of AI are meant to supplement existing practices and training efforts.

In general, as excitement for and use of AI grow, however, so too does the likelihood of harm. This rapid growth has thus necessitated urgent ethical considerations such as the risk-benefit tradeoff for various potential AI applications or the ultimate mitigation of unavoidable or already-taken risks. These ethical considerations are particularly important for contexts like psychotherapy that are integrally tied to people’s safety and wellbeing. Tekin and Delehanty, for instance, explore a number of realistic AI-related harms that may occur in the wake of the FDA’s approval of the Wysa agent for psychotherapy (Tekin and Delehanty, 2026). Many researchers have likewise responded to the growth of AI with a broader call for Responsible AI (RAI). Despite an abundance of research that aims to develop RAI principles, frameworks, models, or tools (Liao and Sundar, 2022; Shneiderman, 2020; Bhatnagar et al., 2025) and a rough consensus of what they should be based on (Jobin et al., 2019), it remains difficult to ensure that they are consistently interpreted and effectively applied in practice (Jobin et al., 2019; Whittlestone et al., 2019; Berman et al., 2024; Wong et al., 2023). It is also often the case that they are only applied retroactively, underscoring the importance of better identifying appropriate solution spaces for AI and understanding the risks in early stages of development (Saxena et al., 2025; Kawakami et al., 2024; Yildirim et al., 2023). This includes accounting for the agency of technologies by understanding the inherent risks their integrations pose (Frauenberger, 2020).

One particular challenge of identifying appropriate solution spaces for AI and understanding its risks comes as the relationship between humans and AI changes (Riva, 2025). A model of interaction that moves toward positioning humans and AI as teammates is increasing in popularity to better capitalize on the relative strengths (and minimize the weaknesses) of each (Zhang et al., 2021; O’Neill et al., 2022; National Academies of Sciences, Engineering, and Medicine, 2022; Flathmann et al., 2021; Caldwell et al., 2022; Schelble et al., 2020). This new collaborative rather than interactive paradigm reaffirms concerns about the amorality (Tolmeijer et al., 2022), competence (Khadpe et al., 2020), and transparency (Rebensky et al., 2022; Hauptman et al., 2024) of AI systems. It also raises questions about where accountability lies (Yousefi et al., 2025; Tolmeijer et al., 2022) and how trust will be distributed (Schelble et al., 2022; Duan et al., 2025, 2024) across human-AI teams.

The dynamic nature of these human-AI teams in psychotherapy—with shifting roles, responsibilities, and accountability between providers, patients, and AI systems—makes it all the more difficult to identify appropriate solution spaces, apply RAI principles, and limit practical risks ahead of time. This is especially true as these teams are built on traditional psychotherapy practices that are still actively grappling with issues of variability in treatment quality, inconsistent evidence across modalities, difficulties evaluating outcomes, and unequal access, many of which motivate the integration of AI in the first place (Barth et al., 2013; Berg, 2021; Dehbozorgi et al., 2025; Giotakos, 2025). While accounting for shifting human-AI dynamics, it must be better understood how AI addresses, extends, worsens, hides, or even creates new issues like those that have historically challenged psychotherapy practices. At a minimum, this might mean designing roles for AI in psychotherapy with lower degrees of autonomy, preserving the decisional capacity of providers (Swinger et al., 2025). This must occur alongside better AI-education tools for providers as their domain-specific expertise may be insufficient for well-mitigating AI’s risks (Swinger et al., 2025). It also means heeding the broader call for multi-disciplinary collaboration (Xu and Gao, 2024) and participatory approaches (Simonsen and Robertson, 2012) that bring in primary stakeholders to the design, development, and deployment of human-AI collaborative systems for psychotherapy. A more targeted framework that helps a broad range of stakeholders identify both best uses of AI as well as possible harms is therefore warranted (Tekin and Delehanty, 2026).

This perspective piece ultimately aims to facilitate a conversation between multi-disciplinary experts about responsibly navigating human-AI collaboration in psychotherapy and the likely risks that might occur with AI use in practice. We propose the use of the Dull, Dirty, and Dangerous (D3) framework common to robotics and human-robot interaction. With a few updates to accommodate the changing relationship between humans and AI—a relationship that can be seen as increasingly entangled (Frauenberger, 2020)—we believe the D3 framework can help guide the responsible integration of AI into psychotherapy.

2 The dull, dirty, dangerous (D3) framework for AI in psychotherapy

2.1 A brief history of the D3 framework

The D3 framework’s origins can be traced back to a Japanese expression, “kitsui, kitanai, kiken” (roughly translating to demanding, dirty, dangerous), representing different ways that migrant laborers found work undesirable (Connell, 1993; Yamamura, 2022). With the framework’s origins as a way of understanding and shaping work practices, it was naturally later applied to manufacturing, with much work on the factory floor falling into the categories of dull, dirty, or dangerous. As technologies like robotics developed to the point that they could replace some human labor, it became pertinent to consider what kinds of tasks should be taken over. Thus, the dull, dirty, and dangerous became seen as the types of undesirable, risky tasks that were better suited to automation. The D3 framework has evolved over time, for example to include a fourth D for “dear” (i.e., tasks that are expensive) (McAfee and Brynjolfsson, 2018), and has also been generally criticized for its incompleteness or insufficiency in determining the best roles for automation in the workforce (Takayama et al., 2008). As such, while the D3 framework lends itself to considerations of the tasks AI should take on, it has largely fallen out of popularity. In the following section, we therefore propose an update to the D3 framework using principles of Entanglement Theory (Frauenberger, 2020) to facilitate its application to AI for psychotherapy.

2.2 Reshaping the D3 framework for AI

While it is no stretch to imagine the D3 framework applying to novel AI systems, the original conceptualizations of dull, dirty, and dangerous differ from how we intend to use them in a few ways. First, given its most common use within robotics contexts, each of the three D’s have often been defined to represent physical tasks. That is, dull tasks are often thought of as physically repetitive and tedious, dirty tasks might imply unsanitary working conditions, and dangerous tasks refer to those that risk physical harm. In some cases, AI has been designed to work in conjunction with and augment automated systems aimed at taking over these physically dull, dirty, or dangerous tasks. For instance, AI has been used to process robotically captured images of sewer pipes and inspect them for damage (He et al., 2025) or make decisions in high-stakes transportation contexts (Almutairi et al., 2025). However, in many cases, a more metaphorical framing of the dull, dirty, and dangerous might better include the range of AI systems being devised. Dull tasks might therefore include those that are monotonous and mentally demanding, dirty tasks can include those that are complex or difficult to navigate and interpret, and dangerous tasks might refer to those that can cause mental as well as physical harm.

Beyond orientation to the physical world, another key aspect of the original three D’s is the undesirability of the tasks. In this sense, having tasks completely taken off the hands of humans can be seen in a broadly positive light. Even in robotics contexts, however, prior work has demonstrated a public preference for a human-in-the-loop approach in which robots perform jobs with workers rather than for workers (Takayama et al., 2008). This need to keep humans in the loop limits the extent to which AI might take over undesirable aspects of work. Further, given the manner in which humans and technologies like AI are increasingly entangled (Frauenberger, 2020), AI might be seen as sharing rather than taking over the burden of dull, dirty, or dangerous tasks. When considering how human-AI teams re-configure and obscure divisions of responsibility and accountability, it becomes important to consider whether outcomes are ultimately less dull, dirty, and dangerous. To this end, we argue that the D3 framework can be used to categorize escalating types of harmful outcomes of human-AI collaboration. Instead of simply asking how AI might contribute to dull, dirty, or dangerous tasks, we can also ask how collaboration with AI might risk exacerbating or creating dull, dirty, or dangerous outcomes. This shift follows ideas of Entanglement Theory, where the complex, entangled relationships between humans and technologies motivate a return to focusing on the technologies themselves—and by extension their relationship with humans—as opposed to a more user-centered paradigm (Frauenberger, 2020). It thus motivates us to think about the design of AI technologies for psychotherapy based on the world we desire, rather than just a strict set of provider or patient needs for better mental health support. It also echoes calls for participatory and speculative approaches that better center the agency of providers, patients, and the technologies themselves.

2.3 Applying the D3 framework to AI in psychotherapy contexts

Having reshaped the D3 framework from positioning AI as taking over dull, dirty, and dangerous tasks to positioning AI as extending or creating new dull, dirty, and dangerous outcomes, we now aim to better illustrate the utility of the D3 framework. Beyond simply addressing the concerns of possible stakeholders, we might start by proactively analyzing AI itself and its potential harms in relation to providers and patients. Below, we provide examples of how the updated D3 framework might be applied in psychotherapy contexts. We draw these examples from our collective experience as HCI researchers investigating and designing (often AI-integrated) systems for psychotherapy contexts (e.g., Swinger et al., 2025, 2026; Baseman et al., 2025, 2026).

2.3.1 Dull psychotherapy

In many cases, psychotherapy can be dull, for both patients and providers. For patients, finding a provider, for instance, can be a mindless and repetitive task of checking insurance coverage and scheduling initial meetings. Once patients have found a provider, some therapies also require the completion of between session homework and notes that many patients might find dull. For providers, psychotherapy involves many dull tasks like writing, retrieving, and reviewing notes and patient homework, or dealing with a host of administrative forms and tasks. AI systems are well-positioned to aid both parties with these types of tasks, but even in these seemingly low risk cases, they are not without the risk of harm. An AI system taking over dull tasks like note-taking for the provider could result in dull outcomes like providers having to manually review accuracy and representativeness of the notes and homework. This could lead to reduced attentiveness or lack of appreciation for a patient’s actual experiences. These roles for AI may also lead to dirty outcomes like patients having to grapple with extra issues of privacy and security while trying to buy into treatment. It might also lead to dirty outcomes like the missing of nuance or context in AI-generated notes, requiring providers to piece things back together while trying to decipher what AI had represented accurately or not. Finally, there may be dangerous outcomes of using AI for these low risk roles, like hallucinations entering the medical record and influencing future diagnoses or a misunderstanding of the danger a patient might be facing. Patients may even stop reflecting on their symptoms and not report them if they themselves lean too heavily on AI-assisted tools.

2.3.2 Dirty psychotherapy

Psychotherapy can also be used for tasks that are metaphorically dirty. AI may be used to screen patient surveys or intake forms, helping make sense of patient responses and screen for mental illness and evaluate severity. AI might also be used to help train new providers and help them understand where and if they have deviated from existing protocols or manuals. On the patient end, AI might be provided as an assistive coach to help challenge negative thoughts or behaviors or reinforce positive ones. In these roles too, while AI might be valuable and help improve mental health care, there is a risk of escalating harmful outcomes. Providers may have dull outcomes like having to confirm that AI has accurately represented numerical responses to surveys or spend time flagging unhelpful training notes or patient coaching prompts. Patients may have to engage in the dull task of spending time learning how to use AI tools instead of engaging with therapy exercises. There may also be dirty outcomes like it becoming unclear where AI is demonstrating fidelity to therapy manuals and protocols related to diagnosis and treatment. It could also be unclear where AI is demonstrating fidelity to cultural or demographic differences as opposed to following stereotypes. Most importantly, there may also be dangerous outcomes. Again, AI might miss nuance or context and subsequently under-report cases of patients facing a high risk of harm. AI may also enable or sycophantically reinforce thoughts and behaviors that are not healthy for a patient. In cases where AI is propagating harmful stereotypes or biases in training contexts, these biases could be reinforced in provider practices and lead to inequitable care for patients.

2.3.3 Dangerous psychotherapy

Finally, AI may be used for psychotherapy tasks that currently involve danger for providers or patients. AI may be used to conduct and monitor suicide risk during treatment or provide patients with exposure exercises in some kinds of therapy. It could also be used to monitor sensor data to determine if there is a physical health risk like an elevated heart rate during treatment. Or perhaps AI might be used by providers to deal with their own negative thoughts, emotions, or questions that come up in response to dealing with patients. While these dangerous use cases present a more obvious need to think about possible harms, it is still worth understanding what the harms might be and how they escalate in severity. Dull outcomes include verifying AI outputs, leading to cognitive surrender on the part of providers or patients themselves to monitoring the progression of symptoms and identifying risks of suicide or self-harm. Dirty outcomes might include the presence of AI in these cases leading to patients changing their behaviors so as to manipulate AI responses, making it unclear which emotions and behaviors are most representative of reality. It may also become unclear who is accountable for ensuring patient safety and being attentive to patient needs and emotions during treatment. Dangerous outcomes are not hard to imagine either, with risks like AI performing unnecessary emergency escalations or giving false reassurances related to suicide or self-harm.

3 Discussion

AI is presently taking over various roles within psychotherapy contexts and is shifting responsibilities as it becomes part of relationships with providers and patients alike. In many cases, this is happening without due caution. At best, it is happening with a focus on the roles and features that are suggested by likely users and experts toward maximizing the interaction experience and on the harms AI might cause in those cases. Even that approach doesn’t adequately account for how technologies have agency and themselves contribute to the shaping of psychotherapy practices. As an example, successful psychotherapies commonly rely on a few key factors, like the therapeutic alliance between a provider and patient (Giotakos, 2025). This alliance—which involves significant rapport building through exhibiting empathy, understanding patient expectations, and building shared goals—may be fundamentally and inherently altered by the inclusion of AI within psychotherapy practice (Baseman et al., 2025; Malouin-Lachance et al., 2025). While AI is often integrated to address weaknesses of psychotherapy like a lack of accessibility, it should certainly preserve and even enhance such core elements as the therapeutic alliance. Design approaches that simply aim to maximize user experience may neglect the agency of AI in psychotherapy contexts in such a way that the therapeutic alliance could be threatened despite the opposite intentions. By updating the Dull, Dirty, Dangerous (D3) framework to accommodate ideas of Entanglement Theory (Frauenberger, 2020), we offer a way for experts and stakeholders across multiple domains to better navigate the integration of AI into psychotherapy. The D3 framework can help these experts and stakeholders understand the needs of users, optimal roles for AI in psychotherapy considering risk-benefit tradeoffs, and better anticipate the escalating dull, dirty, and dangerous harms that AI might exacerbate or newly (and inherently) cause, such as a weakened therapeutic alliance. Future work might seek to make policy or design recommendations using this framework to improve the safety of systems that are integrated into psychotherapy contexts. We also hope that future work might leverage this framework to analyze and adjust existing AI systems, as many risks have been left to hindsight. Certainly, this future work should leverage participatory and speculative approaches that better center the agency of providers, patients, and technologies like AI. While AI for psychotherapy necessitates urgent reflection, well-directed, multi-disciplinary ethical conversations can help guide its responsible integration.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Niko Männikkö, Oulu University of Applied Sciences, Finland

Reviewed by: Orestis Giotakos, Independent Researcher, Athens, Greece