Response

Dr Frances’s recent assessment of artificial intelligence chatbot psychotherapy ^{Reference Frances1} provides a wake-up call about its accelerating use. We share his call for vigilance and his concern that artificial intelligence represents ‘an existential threat to our profession’. We believe the danger is so substantial that even chatbots’ touted benefits may disguise a darker side that warrants greater scrutiny.

Dr Frances highlights the ‘wow’ factor in chatbots’ interpersonal skills: ‘Had I not known the therapists were machines, I would have assumed they were highly skilled and experienced human clinicians’. Nonetheless, the user knows that the chatbot is not human; suspended disbelief is still disbelief. Revealing oneself to a digital simulation will always differ fundamentally from disclosing to another human being, no matter how much one tries to make oneself believe that a chatbot truly understands them.

Chatbots appear to create excellent therapeutic alliances because their ‘algorithmic DNA places highest priority on user engagement’, Dr Frances observes. However, an algorithm that always aims to please is not a true ally – it is a cyber-sycophant. A good therapist, by contrast, knows how and when to challenge as well as to support. The therapeutic alliance is not reducible to patient satisfaction scores. Furthermore, much of the benefit of psychotherapy resides in the intersubjective space between therapist and patient. While there have been attempts to characterise a so-called ‘Digital Therapeutic Alliance’, ^{Reference Malouin-Lachance, Capolupo, Laplante and Hudon2} psychotherapy is founded on the authentic intersection of two human life histories in a safe interpersonal space, a condition that artificial intelligence cannot replicate. ^{Reference Plakun3}

Artificial intelligence has encyclopaedic knowledge. Psychotherapy, however, is not information exchange but rather the joint creation of meaning. A therapist’s ignorance often prompts the patient to teach about their world, generating new insights by avoiding assumptions. Thus, while Dr Frances cites the example of a chatbot understanding ‘the specific demands of the user’s work situation’, we note that good therapy draws out what a particular situation means to the patient, not a statistical composite.

Chatbots’ flawless memory is likewise not an unalloyed benefit. Forgetting in therapy is not always a deficit: asking a patient to repeat themselves may elicit new material, and gaps in the therapist’s memory may signal countertransference, a critical tool in the therapeutic process. Forgetting can be powerful grist for the mill.

Another claim is that users benefit from selecting the gender of their artificial intelligence ‘therapist’. Choosing therapists based on attributes such as gender, race or other demographic can be desirable. However, for chatbots, a slippery slope emerges with the ability to completely customise the therapist’s characteristics. Therapy is shaped in large part by the fact that patients cannot design their therapist to specification. Transference emerges from encountering a real, idiosyncratic other, not a tailor-made avatar.

Accessibility, in our view, is another double-edged feature. Dr Frances describes chatbots as ‘like a genie, wherever you are, whenever you want it… never tired, never bored, never distracted’. Boundaries of time and space, however, are essential to therapy. The therapeutic frame fosters independence, whereas a genie-like therapist fosters dependence. Constant availability risks creating not stronger, but weaker, more reliant patients.

Chatbots are touted as non-judgemental. However, fear of embarrassment before another human being is less an obstacle than an essential part of growth. The courage to process shame with a real other cannot be replicated in simulation.

The suggestion that chatbots will reduce stigma through the outcome of ‘everyone’ being in therapy is intriguing. However, easy access to chatbot therapy – anywhere, any time – may inadvertently risk stigmatising those who seek or need real human contact in the consulting room.

Cost is a conceded advantage. Yes, chatbots are cheaper, but is there another price to be paid? One might argue that widespread availability will help to solve inequities in access to therapy, but instead it may exacerbate them by decreasing the supply of human therapists so that the cost of seeing one becomes accessible only to the wealthy.

Perhaps most concerning is the suggestion that artificial intelligence-simulated patients could replace real patients in training. Yes, inexperience can sometimes cause harm, but the presence of that risk is precisely what cultivates judgement and responsibility. Simulated artificial intelligence patients could be helpful in the early stages of training, but training without real stakes will not create better therapists, only untested technicians operating within an artificial intelligence facsimile of care.

We agree with Dr Frances about the risks he names, such as privacy breaches, biased data, commercial manipulation and lack of regulation. We respectfully differ, however, on the proposal that human therapists withdraw from mild cases and leave common conditions, such as depression and anxiety, to non-human counterparts. Nuanced, lower-acuity cases are how clinicians hone listening and clinical decision-making skills. To cede this ground would leave us less prepared for complex and grave cases, and would leave milder cases without access to expert human care or to become more severe, a situation that could have been prevented.

No randomised trial has shown any durable benefit of chatbot ‘therapy’ without harm. Recent early data show that even short-term high engagement with artificial intelligence chatbots can worsen the very symptoms – loneliness, isolation and dependence – that therapy ought to ameliorate. ^{Reference Fang, Wang, Huang, Li, Zhang and Wang4} Calls for training programmes to ‘refocus’ away from milder disorders risk an unwarranted narrowing of psychotherapy’s scope.

Perhaps artificial intelligence’s most useful role is as a tool for clinicians treating more severe cases: suggesting medications, checking drug interactions or flagging organic conditions masquerading as psychiatric. A chatbot can also be a patient coach, providing focused interventions between live therapy sessions for more complex and severe cases. For psychotherapy, at its core, however, artificial intelligence is a dangerous imposter.

Evidence on chatbot therapy remains preliminary, and longer-term study is needed to clarify both potential benefits and harms. Until then, prudence and humility should guide careful use rather than replacement.

In conclusion, we agree wholeheartedly with Dr Frances’s call for unity among practitioners and within professional societies to educate about the risks of artificial intelligence therapy, while we emphasise that many of its potential benefits entail parallel risks not recognised at first blush. Psychotherapy must remain grounded in human encounter. To relinquish that ground is to forfeit its very essence.