Conversational agents and the making of mental health recovery

Introduction

“Recovery” has been iterated as the new “paradigm in mental health policy and practice.” ¹ In “No health without mental health: A cross-government mental health outcomes strategy for people of all ages’ ² the coalition government set six mental health objectives, one of which was that more people with mental health problems will recover. This recovery-guided approach is continued in “Closing the gap: Priorities for essential change in mental health”, ³ where a commitment is made to commission services with an emphasis on recovery. More recently the “Five year forward view for Mental Health’ ⁴ document was published that stresses that commissioners should prioritise early intervention, choice/personalisation and recovery.

Yet, despite its increased visibility and importance within the mental health arena, recovery remains a vague, “polyvalent concept’ ⁵ that defies definitional consensus. Within the clinical literature, recovery is often situated as the amelioration of symptoms so that a person can resume activities within what is considered a normal range. A second form of recovery has its origins in the Independent Living and Civil Rights Movement of the 1960s and 1970s. This sense of recovery “does not require remission of symptoms or other deficits, nor does it constitute a return to normal functioning. Rather, it views mental illness as only one aspect of an otherwise whole person”.^{6, 7} Whilst the multiple meanings of recovery are often situated as dualisms or either/ors, Pilgrim and McCranie ⁸ suggests that there are four “different shades” of the meanings of recovery: (i) recovery as a personal journey; (ii) recovery as a critique of services that emphasises choice, empowerment, reform; (iii) recovery as therapeutic optimism; and (iv) recovery and the social model of disability. Whilst there is a conceptual separation in their emergence from different social groups there is a practical co-presence in their everyday application that can create “working misunderstandings”. ⁸

Recovery is also (bio)political. Firstly, dominant discourses of recovery are said to individualise what are social problems. ⁹ There has been a neoliberal intrusion on the word and it has been taken over by marketisation, language, techniques and outcomes. ¹⁰ Within this neoliberal, individualising arena recovery has become the “struggle for recognition” and “it is only when the collective, structural experiences of inequality and injustice are explicitly linked to processes of emotional distress that recovery will be possible.” ⁹

Secondly, “definitional debates about recovery reflect wider ideological debates about the nature of mental health.” ⁹ Policy documents talk about recovery within the context of the “global burden” of mental illness^7,11 whilst often sidestepping controversy surrounding diagnosis in psychiatry. The Campaign for the Abolition of Schizophrenia Label, for example, suggests that “the idea that schizophrenia can be viewed as a specific, genetically determined, biologically driven, brain disease has been based on bad science and social control since its inception”. ¹² As Lasalvia and Ruggeri (2018) note, “even within the psychiatric field there have been calls to change the term schizophrenia as not scientifically valid and unhelpfully focusing on a biological explanation of what is a heterogeneous and context-influenced disorder”. ¹³

In short, recovery remains a vague and political concept that sits at the core of current mental health policy despite significant critical voices. In this context, it is timely to explore whether, and in what forms, recovery is situated within conversational agents (chatbots) - an emerging technology viewed as having potential to transform mental health treatment. The Topol Review on the digital future of mental healthcare and its workforce concluded that there is already a strong evidence base for these kinds of interventions and that by 2021 chatbot systems could be offering advanced automated or semi-automated diagnostic and therapeutic tools. ¹⁴ However, whilst considerable research has been carried out on the efficacy, feasibility and ease of use of mental health chatbots, ¹⁵ very little attention is being paid to how chatbots relate to ideas of mental health recovery. We do not know, for example, the extent to which chatbots are designed to work with – and encourage – particular versions of recovery. The paper therefore asks: to what extent are chatbots shaped by, and shaping, the concept of mental health recovery?

We begin with a critical review of the existing literature on chatbots and mental health. Exploring how chatbots shape, and are shaped by, the concept of mental health recovery requires us to look both at overt reference to recovery as well as possible sites of socio-political struggles that may come to conceptualize recovery. ¹ To operationalise this, we derive analytic directives by bringing contradiction, contention and debates within the literature on chatbots into dialogue with the existing literature on the meanings of mental health recovery. We then move to outline the methodological and analytical focus. We interrogate a range of online sources to explore how chatbots are being positioned and the claims being made of them.^16–18 Analysis starts with the foreshadowed ideas from the critical literature on recovery and asks: how do chatbots relate to recovery? How are chatbots positioned in terms of expertise? How do chatbots relate to time? How do chatbots relate to everyday experience and clinical subjectivities? We conclude with a discussion of how recovery may be being troubled by chatbots and the implications for the future of mental health recovery.

Chatbots and mental health recovery: A view from the literature

Chatbots - “systems that are able to converse and interact with human users using spoken, written, and visual languages” ¹⁹ – are increasingly being used in the field of mental health. A scoping review conducted by Abd-alrazaq et al. (2020) identified 53 studies that involved 41 different chatbots. In 17 studies, chatbots were used for therapeutic purposes and of these 10 were based on cognitive behavioural therapy and just under half were focused on people with depression and anxiety. ¹⁹ Woebot, for example, treats “depression and anxiety using digital version of time-tested cognitive behavior therapy”, ²⁰ whilst #BeALifeline Direct Message (DM) Chatbot aims to help start conversations about suicide and self-harm. ²¹ They are also being tailored to populations, with HARR-E being launched to tackle the male mental health problem and Wysa being originally tested and co-designed with an adolescent user group. As well as therapy, chatbots also aim to provide training, improve social skills and act as a screening tool for such things as depression, dementia and posttraumatic stress disorder. ¹⁹

Recovery has an ambiguous presence within this literature. Whilst there is minimal overt mention of recovery, there is a clear focus on efficacy and symptom reduction with Fulmer et al., ²² for example, examining scores on the Patient Health Questionnaire (PHQ-9), Generalized Anxiety Disorder Scale (GAD-7), and Positive and Negative Affect Scale (PANAS) at baseline and 2 to 4 weeks later. Other studies investigate differences in effect index across groups and time, ²³ or report on the development of a chatbot to deliver support for caregiving professionals; also, as measured by the PHQ-9 and GAD-7. ²⁴ Recovery, then, appears to be tacitly situated as a (clinical) outcome and the amelioration of symptoms.

With that said, there are areas of contradiction, contention and debate within the literature that point towards emerging social, political and technological tensions around the conceptualization of recovery. Several of these are made apparent in the following clinical abstract:

“A World Health Organization 2017 report stated that major depression affects almost 5% of the human population. Major depression is associated with impaired psychosocial functioning and reduced quality of life. Challenges such as shortage of mental health personnel, long waiting times, perceived stigma, and lower government spends pose barriers to the alleviation of mental health problems. Face-to-face psychotherapy alone provides only point-in-time support and cannot scale quickly enough to address this growing global public health challenge. Artificial intelligence (AI)-enabled, empathetic, and evidence-driven conversational mobile app technologies could play an active role in filling this gap by increasing adoption and enabling reach. Although such a technology can help manage these barriers, they should never replace time with a health care professional for more severe mental health problems. However, app technologies could act as a supplementary or intermediate support system. Mobile mental well-being apps need to uphold privacy and foster both short- and long-term positive outcomes.” ²⁵

Struggles over expertise manifest in the literature where conversational agents are situated as offering low-cost, on-demand, highly adaptable, self-help which will help address a growing global challenge. However, at one and the same time, it is made clear that conversational agents should supplement rather than replace traditional therapeutic options. In many respects this is a variant on an old theme. As telepsychiatry developed it was suggested that this could increase access. ¹⁷ However, clinicians were also presented as being resistant to its advancement because of the deleterious effect on the doctor-patient relationship deemed central to therapy. For example, and as Pickersgill notes, ¹⁷ in his 1952 address as President of the American Psychiatric Association, Leo Bartemeier expressed the need to be cautious about the rise of technological approaches to clinical practice and alert to the possible impact on clinical and therapeutic interpersonal relationships. The dual emphasis on reach and tradition creates complex biomedical virtues – which “refers to the (profession‐defined) praxis of goodness within the laboratory and the clinic” ¹⁷ – where telepsychiatry and direct clinical care need to be carefully situated as extending clinical care rather than replacing it. Pickersgill ¹⁷ uses the example of Chill Panda, an app that captures biometric data – such as heart rate and blood flow – and suggests playful tasks that suit your current state of mind. Users are encouraged to learn how to manage their stress and feel better. However, this is offered with provisos and “clinical expertise thus remains salient, even as (potential) patients are encouraged to contribute to their care”. ¹⁷

Within the literature on chatbots, these tensions between tradition and reach intersect with ambiguities surrounding artificial intelligence. Vaidyam et al. ¹⁵ define chatbots as “digital tools existing either as hardware (such as an Amazon Echo running the Alexa digital assistant software) or software (such as Google Assistant running on Android devices or Siri running on Apple devices) that use machine learning and artificial intelligence methods to mimic humanlike behaviours and provide a task-oriented framework with evolving dialogue able to participate in conversation”. Seen in this way, AI becomes an essential part of chatbots and conversational agents are situated as part of the general AI revolution said to be “transforming mental health”. ²⁸ The Alan Turing Institute is also attempting to drive forward research in to AI-based precision mental health. There is also a move towards a language of “prediction”. A 2018 paper ³⁰ for example, suggested that AI could be used to identify how well people at risk of psychosis will function in the future. The news forum “Scimex” reported this under the banner – “AI could be used to predict mental health recovery”. Yet, the majority of the chatbots identified in a review (92.5%) depended on decision trees or predefined rules to generate their responses. Only 7.5% used machine learning approaches. ¹⁹ The nature of artificial intelligence within the chatbots remains ambiguous and it is therefore unclear whether struggles over expertise involve users gaining more control and autonomy, or the chatbot itself becoming the dominant agent: further individualising “treatment” and ultimately alienating both service users and clinicians from it.

There are also temporal ambiguities surrounding the ways in which AI is discussed; including a mix up of tenses between what AI could do and what AI actually does presently. This temporal ambiguity complicates the role being claimed for AI in recovery whilst also complicating what recovery is. Temporality has long been central to discussion of recovery. For example, literature framed by the biomedical/disease model suggests that disturbances in the experience of time are a commonly reported feature of mental disorders. Also working within a biomedical/disease model, Eugene Minkowski, noted that with those diagnosed with schizophrenia “experienced time is altered in its flow, being experienced as frozen, immobilized, without “elon vital”. Experience of time is also affected by specialization – that is time is felt as divided in juxtaposed elements – and temporal fragmentation: “With the fracturing of time flow, we observe an itemization of now-movements in consciousness so that each now-movement in a person’s stream of consciousness will be experienced as detached from the previous one and from the following, hence as extraneous to one’s stream of consciousness and sense of selfhood.” ³¹ Time continued to be a methodological tool and pathological focus of clinical psychiatry and psychology throughout the first half of the 20th century – and Mnkowski played a key role in shaping subsequent understandings of abnormal time and depression and schizophrenia. ³²

Amongst other things, this literature also notes a “subjective decrease in the experienced velocity of the flow of time in depressive disorders” ²⁹ and that “depressed patients” showed “general changes in their experience of the passage of time. The most common was that of time having slowed down.” ²⁹ Perhaps in contradiction, Vogel et al. ²⁹ also found some indication of an increase in the experienced velocity of the passage of time. As noted further below, it therefore becomes important to consider how chatbots frame and portray the experience of time. Words like relapse, recurrence, deterioration, periods of improvement remain common within biomedical discussions of depression. As Fava and Visani ³³ suggest, for example: “clinicians working with depressed patients are often confronted with the unsatisfactory degree of remission that current therapeutic strategies yield, and with the vexing problems of relapse and recurrence”. Yet, as illustrated in the abstract above, the literature on chatbots makes redress to “short and long term” positive outcomes. In doing so, it appears to move us away from expressing and experiencing recovery through a language of relapse, recurrence and deterioration. Chronicity remains in uncertain ways and the language of “outcomes” sidesteps, rather than resolves, tensions that inhabit notions of recovery.

Methods

This paper focuses on the ways official online materials explain and make sense of chatbots, their imagined functionality and value for (potential) users. ¹⁶ Our interest is in how chatbots are positioned and the claims being made for them, rather than the experience of using the chatbot in itself. As apps are interpreted by surrounding materials that give them meaning, ³⁶ empirical data reported on in this paper is comprised of publicly available online materials. Materials are sourced from official company pages and include blog posts, promotional videos, marketing materials and user guides.

As noted, chatbots have been developed for a range of tasks including therapy, training, screening and improving social skills. It has also been proposed that voice signal extraction can be used to detect mental illness. ³⁷ Within the present study, we adopted a purposive, homogeneous, sampling strategy ³⁸ and focused on those offering forms of therapy. Sampling proceeded as follows: First, a decision was made to explore those apps which are situated as offering assistance for depression/anxiety. This is because reviews suggest that this is the most common focus of chatbots. ¹⁹ Second, three chatbots for mental health were selected for analysis: Woebot, Wysa and Tess. These are three of the most popular chatbots; for example, appearing at the top of Clarke’s ³⁹ list of “AI Bots and Apps for Depression”. Woebot launched in the Summer of 2017 and is “designed to offer convenient care to those struggling with depression by mimicking human conversation, offering self-help related guidance and companionship to its users”. Woebot uses chat-based functions to mimic online human interactions. It is built around principles from Cognitive Behavioural Therapy (CBT) and uses natural language processing to “get to know” users over time and “more accurately detect and meet your emotional needs at a given time, offering personalized resources, self-help guidance, information and support related to your concerns”. It won the google play award for “standout wellbeing app” in 2019. Wysa is reported to have around 1,500,000 users. Like Woebot, both Wysa and Tess incorporate principles of CBT. However, Wysa is also built around dialectical behavioural therapy, meditation practices and motivational interviewing. Tess moves away from app based approaches and uses text-based messaging to track user goals and provide guidance and interventions. ³⁹

These chatbots are developed outside of mainstream mental health services and can be expected to be driven, to some extent, by commercial interests and the desires of potential customers, including mass healthcare purchasers such as in the UK the National Health Service. Yet, at one and the same, whilst it might be assumed that chatbots will draw from a biomedical framework because of this, developers also stress the important role that users play in their evolution. For example, the Wysa website states that “Over 60 psychologists and 15,000 users have provided specific inputs to shape how Wysa helps them. 65 users have volunteered to help us translate Wysa into their language.” Materials analysed are thus formulated within a complex space that includes developers, psychologists, “users”, consumer concerns and healthcare commissioning rules. Materials were systematically collected. For each chatbot, we began with the official company webpage. These pages were downloaded, along with anything that was directly linked. For example, the Woebot website lists two scientific references on its “how it works page”. These were also downloaded and analysed. The Tess website also includes a series of references as well as embedded videos from both clinicians and users. As Berg ¹⁶ notes in his analysis of self-tracking devices, “Since all of these materials are part of the companies” marketing strategies and social media presence, they are regarded as not only official presentations of the products, but also as naturally occurring empirical materials.” Online Appendix 1 lists the sources analysed and accompanying urls. All sources were analysed between August 2019 and February 2020. Links are accurate as of then.

Analysis proceeded in three steps. First, materials were examined for explicit reference to “recovery”. Where the word recovery was present, we reflected on the context and framing and the discursive resources being deployed. Second, a wider reading was conducted foreshadowed by the ideas and questions from the critical literature review and materials were read sensitized to possible struggles over expertise, technology and temporality that implicitly invoked recovery. Finally, we drew these two sets of observations into the set of cross-cutting themes presented below. This discussion is not an exhaustive set of findings but focuses on sets of tensions we consider particularly pertinent to the development of future notions of recovery.

Findings

Across Woebot, Wysa and Tess there is largely an absence of overt reference to “recovery”. Exceptions include the statement on the Tess website that “92% of people moved towards recovery” (Tess_2) and Woebot’s claim that “struggle is actually a necessary part of recovery” (Woebot_18). There is no further elaboration on the statement and it is instructive that it appears on a page titled “outcomes”. Across the three chatbots, outcomes dominate as an idea. The “science” page of the Woebot website, for example, has very little information beyond a graph showing “reduction in depressive symptoms” following use of Woebot (Wobebot_2) Readers can follow a link to a paper by Fitzpatrick et al. ⁴⁰ which reports that “Intent-to-treat univariate analysis of covariance revealed a significant group difference on depression such that those in the Woebot group significantly reduced their symptoms of depression over the study period as measured by the PHQ-9 (F=6.47; P=.01) while those in the information control group did not”. Somewhat similarly, Wysa’s webpage points the user to a paper by Inkster et al. ²⁵ that evaluates the Wysa app using ideas of impact and engagement. Impact was measured by capturing the average improvement (pre-PHQ-9 minus post-PHQ-9) between a group of high and low users (Wysa_2 & Wysa_6)

There is a prima facie case, then, for suggesting the conversational agents are invoking an idea of recovery that involves smooth and linear reduction in symptoms over time and locates the agency for bringing about that recovery in the app. Debates surrounding diagnosis are made redundant by claims that the chatbots are “agnostic when it comes to diagnosis” (Woebot_18) and very little critical attention is paid to who is deciding which outcomes are core or to the standard questionnaires and measurement devices employed in the clinic to capture these outcomes. Screening tools arguably constitute the phenomena they purport to track rather than simply reflect them. ⁴¹ – especially through techniques such as “reduction”, “expression”, “quantification” and “normalisation”.

Further analysis does suggest complex and competing ideas which coexist with this narrow view of recovery. In making these apparent, we firstly discuss three themes that map on to the struggles over practice, expertise and evidence ¹ foreshadowed by the literature review: “troubled clinical responsibility”, “extended virtue of (technological) self-care” and “altered ontologies and psychopathologies of time”. Within the concluding discussion we explicitly connect these back to recovery.

Concluding thoughts – What then of recovery?

Following Williams et al. ¹⁸ firm conclusions may be unwise, given the embryonic nature of these developments. However, a number of tentative conclusions may be drawn. For one, multiple versions of recovery appear to be approximated within the materials. On the one hand, there appears to be a discursive move towards a recovery from model. As Davison and Roe note,

“Recovery from serious mental illnesses involves the amelioration of symptoms and the person’s returning to a healthy state following onset of the illness. This definition is based on explicit criteria of levels of signs, symptoms, and deficits associated with the illness and identifies a point at which remission may be said to have occurred. This definition thus has many advantages from clinical and research perspectives, as it is clear, reliable, and relatively easy to define, measure, and link to dysfunctions or wellbeing in other areas of life. People who enjoy this sense of full recovery could be considered to have recovered from psychosis in the same way that other people may recover from an infection, a broken leg, or, in the case of recovery over a longer period of time, asthma”. ²⁶

Notwithstanding this, a third form of recovery is also made apparent across the contradictions and areas of debate within the materials. Both the recovery from and the recovery in model are challenged by the vague ontological status of mental illness across the materials analysed. Where the disease model is questioned – and mental illness is situated as a “problem of living” - recovery can no longer be about recovering from, or learning to live with limitations caused by, illness. In the absence of overt discussion, users appear to be implicitly invited to bring together knowledge from recovery and positive psychology and, in the process, see recovery as “personal”: “Personal recovery involves working towards better mental health, regardless of the presence of mental illness.” ⁵⁵ Across the materials, “personal recovery” appears accompanied by an ever-increasing imperative built on the normalisation of apps as a form of self-care. Conversational agents should be used by all; not least because shifting temporalities mean we do not always find others awake when we may need them. AI becomes somewhat diminished in status as simply a way to deliver more accessible care.

“Personal recovery” is increasingly individualised and the experiences of the user may become further alienated from their own recovery. When we talk of recovery from versus recovery in, scientific versus consumer models of recovery, clinical versus personal recovery or versus social recovery ⁵⁶ we are often talking about whether it is clinical or personal expertise which is brought to the fore. The conversational agents explored here potentially trouble this dichotomy by situating biomedical virtue within Cognitive Behavioural Therapy itself rather than the individual user or therapist. It may be more appropriate then to suggest that chatbots are being shaped by, and shaping, (im)personal recovery imperatives.

Further research is needed here. The limited scope of the current study means there remains a need to delve deeper into the positionings embedded within use of the apps. “Walkthroughs” offer the possibilities for further exploration of digital applications” “sociocultural representations as much as its technological features or data outputs, which also have social and cultural influences.” ⁵⁷ The “walkthrough” method requires researchers to go through individual chatbots paying attention to such things as (I) the app’s vision; (ii) its operating model; (iii) its governance; (iv) mediator characteristics, such as how the app guides users through activities via menus and buttons; (v) registration and entry; (vi) and app suspension and closure. ⁵⁷ Such approaches can provide data on intended purpose, embedded cultural meanings, and implied ideal users. ⁵⁷ Applying this method to a range of chatbots can shed light on how different design contexts impact on the ways that recovery is embedded within these technologies. Capturing stakeholder meanings and motivations is also important moving forward.

As well as further exploring how chatbots are shaped by, and shaping, recovery and asking what models are excluded and what are the processes of truth-making, future research needs to consider how social scientists might work with key stakeholders to challenge enduring hegemonic forms. It is also important that research engage with the question of whether we should focus on recovery. As Pilgrim and McCranie ⁸ note, whatever sociological frame we employ: “it is clear that recovery as rhetoric in the policy and services field is neither self-evidently desirable or worthy”.

Supplemental Material