For

We are told that artificial intelligence (AI) could replace most, if not all, human jobs.^{Reference Grace, Salvatier, Dafoe, Zhang and Evans1} Here, we argue that psychiatry is far from immune to such a coup. We will consider our looming obsolescence in two domains: technical and relational.

It is uncontentious to assert that AI will transform the technical aspects of psychiatry. For example, smartphone technology enables the capturing of rich, longitudinal, multimodal data, the analysis of which promises vastly improved characterisation of illnesses and their trajectories.^{Reference Onnela and Rauch2} Such methods have already shown predictive potential in forecasting relapse in bipolar disorder.^{Reference Vanello, Guidi, Gentili, Werner, Bertschy and Valenza3–Reference Gruenerbl, Osmani, Bahle, Carrasco, Oehler and Mayora5} Coupled with improved models of treatment efficacy^{Reference Huys, Maia and Frank6} and increasingly naturalistic, data-driven taxonomies of mental illness,^{Reference Grisanzio, Goldstein-Piekarski, Wang, Rashed Ahmed, Samara and Williams7} it seems likely that computers' technical mastery in diagnosis and treatment-planning will soon out-do that of humans.

However, psychiatry offers more than the sum of its technicalities. Clinical exchanges in psychiatry involve a dynamic interplay of facts and values, through which patients might find relief in feeling listened to and understood. Trust within the ensuing relationship further enhances treatment effects.^{Reference Krupnick, Sotsky, Simmens, Moyer, Elkin and Watkins8} Might this richness of communication prove beyond the abilities of AI? If so, could the human psychiatrist survive as skilled mediator between patient and machine – a human front-end to an AI operating system? In our view, there are good reasons to believe that AI-led care, even in psychiatry's most relational aspects, could ditch the human middleman.

AI is not yet able to converse with sufficient flexibility to hold a psychiatric interview. However, natural language-processing is rapidly advancing and conversational agents have already found application in assessing alcohol-drinking habits.^{Reference Elmasri, Maeder, Ascoli, Hawrylycz, Ali, Khazanchi and Shi9} Indeed, there is good evidence to suggest that people can build therapeutic bonds with AI agents. For instance, human minds naturally infer emotion and intentionality from limited data-points,^{Reference Heider and Simmel10,Reference Scholla and Tremouletb11} and derive comfort from^{Reference Bemelmans, Gelderblom, Jonker and de Witte12} and even empathise with^{Reference Suzuki, Galli, Ikeda, Itakura and Kitazaki13} minimally animate objects. Evidence suggests that people can be more honest with computers than they are with humans.^{Reference Lucas, Gratch, King and Morency14} The intelligent clinician-in-the-app might even foster trust by accurately reporting confidence intervals on its own predictions. Taken together with the creation of highly contextualised models of a person's thoughts and behaviours, it seems very likely that people will readily experience an AI clinician as genuinely caring and understanding. Availed with sufficient data, a future AI will build a deep enough model of a person's responses that its understanding of them surpasses that of their psychiatrist.

We emphasise that the succession of human psychiatrists by AI agents is not without risk. Unsupervised machine-learning could lead to insidious exacerbations of existing biases.^{Reference Bolukbasi, Chang, Zou, Saligrama and Kalai15} Value misalignment – the degree to which the goals of AI systems fail to overlap with our own – is a major potential hazard.^{Reference Bostrom16} In psychiatry this problem is prefigured by existing disagreements as to what the objective of mental healthcare is,^{Reference Szasz17–Reference Moutoussis, Story and Dolan19} in other words, how people ought to live. These are political questions and will remain a matter for human discourse. As such, we must qualify our argument with the suggestion that the process of psyche-iatros (mind-healing) in a broad sense will retain a human component, even in a post-human doctor world.

Perhaps human patients will want human doctors, with all their quirks and comparative ineptitude, simply because they are human. Cynically, if silicon shrinks are cheaper and measurably as effective as human psychiatrists, they may find mass employment by economic demands alone. More positively, AI promises considerable advantages for patients. Rather than straitjacketing complex individuals into diagnostic categories and doling out treatments based on generic guidelines, AI offers truly individualised care. Furthermore, unlike the hugely fragmented pathways of care patients are currently subjected to, a personal AI clinician would be available more or less anywhere (from primary care to the in-patient ward), at any time (cradle to grave, night or day). Over time then, patients' trust could be earned and maintained, rather than being repeatedly fractured with every handover. Thus, people may well find AI-led care to be paradoxically more humane than the status quo of psychiatry today; in their desire to be understood and treated to the best of scientific understanding, they will willingly choose AI over its flesh-and-blood alternative.

Christian Brown and Giles William Story

Against

Will the art of mind-healing eventually ditch ‘the human middleman’ and come to rely exclusively on data-driven AI that patients will pour their hearts out to and then optimise their mental health? Do not count on this rosy, tech-driven, mental health future coming to pass anytime soon.

Although we agree that AI has the potential to significantly transform the technical aspects of psychiatry by discovering hidden patterns in rich, longitudinal, multimodal data, which can be predictive of individual risks and outcomes,^{Reference Bzdok and Meyer-Lindenberg20} we are less convinced that AI-based tools will ever be able to serve as the treatment of choice for the vast majority of humans in distress, at least in a way that fulfils the purpose of the psychiatrist. Of course, many humans find it easier to be more honest with computers about certain things; however, this hardly indicates that the encounter is serving the best interests of the patient, except in a strictly information-gathering context. However, this is a tiny fraction of a psychiatrist's role and one often served by less-trained individuals or by intake forms. Once computers start responding more like a human concerned for that individual's long-term well-being, humans may well become more circumspect about what they reveal to computers.

Starting from a technical point of view, there are still many challenges for AI to overcome. Currently, the quality (and quantity) of data available to train AI models in psychiatry is very limited, partly driven by untested or even invalid assumptions about the structure of mental illness. As an example, the performance of neuroimaging-based, diagnostic AI models tends to decrease as sample sizes increase (for examples, see Varoquaux^{Reference Varoquaux21} and Janssen et al ^{Reference Janssen, Mourão-Miranda and Schnack22}), indicating that the models do not perform well for highly heterogeneous large samples. One important reason for this lack of clarity in such studies focused on using AI for diagnostics is the extent of comorbidity in psychiatry, and the limitations of current categorical classifications.^{Reference Insel, Cuthbert, Garvey, Heinssen, Pine and Quinn23} Therefore, it is very unlikely that AI systems will be successful in solving any diagnosis task because psychiatrists often disagree and therefore there is no ground truth to measure model performance against.

Furthermore, AI is very different from human intelligence. AI systems often perform badly when presented with data that is very different from the training data or with a new task (different from the one it has been trained to solve). Although AI systems can be trained to detect novelty (‘anomaly or outlier detection’^{Reference Hodge and Austin24}) and transfer knowledge to solve related problems (‘transfer-learning’^{Reference Pratt and Thrun25}), making accurate predictions when confronted with uncommon patterns or new tasks is incredibly difficult for an AI system. For example, an ‘AI psychiatrist’ would not know what to do if a patient were to present with a behaviour completely different from every behaviour it has ‘seen’ before. In such situations, a human psychiatrist is able to take in a far broader and real-world relevant set of observations, and then draw on their own idiosyncratic knowledge base to contextualise those observations, applying a moment-by-moment solution that mutually optimises for more than just the treatment outcome (e.g. aspects of that individual's current life situation that the AI never sees and thus never models). This ability of human intelligence to draw on ‘common sense’ when needed, and uncommon sense when especially needed (i.e. that of specialists with their unique training data), means that humans will almost certainly remain critical for managing and interacting with even relatively simple psychiatric cases for the foreseeable future.

Looking ahead, we acknowledge that AI systems may eventually learn to incorporate many of these contextual variables and become more efficient in transfer-learning across different tasks. And yet, AI systems will require more, better data than currently exist to solve any real-world clinical task faced by psychiatrists. For most of these points in the modern psychiatry workflow (e.g. initial evaluation, establishing a trust-based working relationship, differential diagnosis, treatment selection and side-effect and efficacy monitoring) very few if any studies have attempted to measure and compare clinical behaviours that distinguish adequate from inadequate clinical performance. Although several large-scale research efforts, such as the UK Biobank, are collecting detailed information from thousands of individuals, including many demographic, lifestyle, behavioural and imaging-related factors, these data barely scratch the surface of what a psychiatrist would need to perform any of the above tasks competently, let alone provide comprehensive care of a single patient.

Some have even suggested that the lack of group-to-individual generalisability, which is not unique to psychiatry, threatens the entire enterprise of human-subject research^{Reference Fisher, Medaglia and Jeronimus26} because most human-derived measures fail to account for important within-person variances that are critical to their interpretation. Therefore, to truly succeed at contextualised AI – able to provide nuanced care to complex, evolving people embedded in complex, evolving environments – these systems would require data from hundreds or potentially thousands of patients and psychiatrists, gathered across dozens of repeated encounters, and combined with standardised ways to assess observed outcomes to validate any autonomously discovered relationships to establish trust in the algorithm. These data-sets are possible, and indeed are starting to be collected, but they too require tremendous human capital to collect and understand.

And no matter how sophisticated such systems become, psychiatry will always be about connecting with another human to help that individual integrate all the conflicting signals they are receiving and make the best choices for their life situation. So much of addressing mental health is dealing with challenging, embarrassing issues that people often do not admit even to themselves, and only begin to comprehend as layer upon layer of meaning and data are exposed by skilled professionals. Human psychiatrists (and psychologists) are the ones in the loop who that individual will not be able to fool, and will know how to respond when things go sideways, which they often do. In these cases, how many would feel safe with an autonomous agent helping them and their family make the best use of available resources and data during a complex mental health challenge, and not inadvertently make things worse?

Clinicians nowadays rarely rely on strict categorical diagnostic systems, because they already know to treat constellations of symptoms and avoid side-effect profiles that will intersect with real-life challenges like cost and access to follow-up care. It would seem that entrusting an AI with all this complexity and uncertainty has no less potential for abuse and eventual backlash than any of psychiatry's past treatment approaches (e.g. straightjackets) and diagnostic schemes that eventually were discarded, and even reviled, precisely because they had failed to capture relevant individual nuance and thereby eroded patient trust. This meta-psychiatry problem is a social one and probably one AI systems cannot and should not solve. Human psychiatrists, along with the people they treat, clearly will first need to agree on which problems are even worth addressing with AI, to move forward in a way that is both pragmatic, ethical, stakeholder-driven and iterative by design. The power of iteration is one place humans could stand to learn a great deal from artificial systems, but with regard to learning what it means to be ‘humane’, we humans will be the ones doing the teaching for a good long while.

Janaina Mourão-Miranda and Justin Taylor Baker

For: rebuttal

Baker and Mourão-Miranda argue that psychiatry presents significant technical challenges to the AI practitioner of today, and we do not disagree. However, we take the view that there is nothing so mysterious about human behaviour and social interaction that it will never be possible to simulate these by artificial means. As such, the question of technical capability is, for many, not one of if, but when. We turn, therefore, to our opponents' non-technical claims made against AI psychiatry in general.

They assert that ‘psychiatry will always be about connecting with another human’. If AI has too little understanding of humans, they point out, there is the potential for people to ‘be able to fool’ the artificial clinician, or for the silicon shrink to ‘inadvertently make things worse’. With too much understanding, however, AI could arouse suspicion. Of course, similar concerns arise with human clinicians. Indeed, this re-telling of Goldilocks, in which people want their AI psychiatrist to have just the right level of intelligence, reveals a human tendency to prefer their own kind a priori, notwithstanding how sophisticated the alternative might be. Among other factors, an AI psychiatrist's authenticity is limited by its lack of a human body, and who knows whether we will ever fully confide in those who are not made of flesh and blood like ourselves? If we cannot, and continue to nurse our anthropocentric distrust, we stand to miss out on great therapeutic potential.

Our colleagues highlight that AI psychiatry has the potential for abuse and backlash (à la the straitjacket). We would go further and acknowledge the immensely serious concerns of many within the world of AI ethics, that these new technologies could give rise to devastating, perhaps existential consequences. If human intelligence is what afforded us dominion over nature, superintelligence should rightly be regarded with caution. Baker and Mourão-Miranda highlight, in particular, a social and ‘meta-psychiatric’ problem, which they assert that AI should not solve, with humans remaining at the helm. We do not see the problem in such binary terms.

There is no clear point at which prevention of mental illness departs from improving the well-being of the population. Similarly, the practice of psychiatry on an individual level affects the culture of psychiatric practice, which in turn exists in equilibrium with wider societal values. We acknowledge, therefore, that AI practitioners will have effects far beyond the clinic room, becoming active participants in our ever-shifting ethical dialogue. Neither humans nor computers ought to entirely dictate this dialogue; rather, various agents will contribute to varying degrees, in an interplay where humans remain able to exert their diverse priorities and values. We do not anticipate a future of ethical homogeneity, artificial or human.

In conclusion, we emphasise that the advance of AI psychiatry is inexorable, and for its advantages (therapeutic and economic), people will readily choose to use it. As such, we should stop debating whether it will become sufficiently sophisticated to replace psychiatrists, and instead turn our focus on how best shape this future, and what kind of ethical and regulatory systems are needed to prevent disaster.

Christian Brown and Giles William Story

Against: rebuttal

Brown and Story emphasise the inexorable nature of AI eventually replacing psychiatrists, and again we do not disagree that some forms of AI-driven mental health solutions will be integral to the future of mental healthcare. However, we are not aware of any evidence to suggest the true effectiveness of human- versus computer-mediated psychiatric disease management is likely, in the end, to come out in favour of the AI agents outperforming actual humans in the complex task of managing humans, or interacting with them at the point of encounter to achieve optimal behavioural health outcomes. It seems to us far more likely, when all is said and done, and the right data are collected and analysed to properly assess the value of both the in-person and remote encounters, that the data will show that computers alone (and even computer-assisted virtual encounters) are not nearly as effective at getting the job done as well-trained humans – at least not today's computers, with their limited interfaces and access to human-relevant metadata. But we agree much of this will change and is changing rapidly.

After all, many individuals do seek out indirect human interaction to alleviate their suffering, including online text coaches and therapists (Crisis Text Line, 7cups, Woebot, etc.), and there is clearly a market for such platforms for many individuals who cannot or prefer not to seek care through conventional means. But again, we strongly suspect, based on the underlying phenomenology of depression and other severe psychiatric conditions, that in-person contact will continue to represent best clinical practice for the vast majority of psychiatric interactions, and that despite the many exciting alternative forms of care delivery designed to improve access and standardise quality, humans will continue to be favoured over less direct, less humanising alternatives. Indeed, we are especially concerned with unintended consequences of expanding access to AI-based platforms that are relatively untested, as they could exacerbate existing disparities by providing those with fewer means access to only lower-quality treatment options. We think systems can be designed to avoid this scenario; however, for many individuals who might have been helped by a human with even modest training, the prospect of inexpensive AI-based systems that become the only available option for those in need raises significant ethical concerns that must be addressed.

Given our contention that humans in distress will always be best served by meaningful interactions with other humans, we cannot ignore the potential for immersive technologies like virtual reality and augmented reality to help facilitate human interactions at a distance. With virtual and augmented reality capabilities built into modern communication devices, immersive technologies are increasingly culturally resonant. As such, over the 3–5 year horizon (and certainly within the next decade), realistic in-home agents that replace two-dimensional video conferencing with three-dimensional virtual reality and/or augmented reality interactions with distant individuals will become a possibility. Here, we are not talking about interacting with virtual avatars,^{Reference Morency and Rizzo27} but rather connecting with people at distant locations in ways that feel real, allowing people to experience deep personal connections with others without technology getting in the way. Just as the first people to experience a musical recording or a film in their own home in ways that felt ‘just like the real thing’, we suspect humans will one day be able to experience one another's presence at a distance in ways that no longer feel degraded. Although we are not there yet, the technical advances needed to realise this are knowns rather than unknowns, and many are well in process, the remaining engineering and cultural obstacles to achieve embodied interactions with distant humans in realistic virtual or augmented environments is so close at this point as to be both exciting and terrifying, given how it could affect humanity.

In conclusion, although in a narrow sense we stand by our assertion that human psychiatrists will never be replaced by AI, the technical advances provided by AI will inevitably transform many aspects of psychiatry, particularly in the identification of new biomarkers and approaches to patient stratification, which will likely lead to development of new and more effective therapies. And yet in the end, until technology can enable new ways of connecting with another human being that do not seem degraded by the human–computer interface, psychiatrists will probably remain most effective at the human–human interface.

Janaina Mourão-Miranda and Justin Taylor Baker