Abstract

Background. Over the past fifteen years digital psychiatry and telemedicine have transformed access to mental health care. Yet under a single label coexist realities with very different levels of evidence.

Objectives. To distinguish critically what is demonstrated from what is promised — from teleconsultation to apps, from passive monitoring to generative artificial intelligence — assessing efficacy, ethical risks and access inequalities.

Methods. Critical survey of the international and Italian literature, anchored to verified primary sources.

Results. Synchronous telepsychiatry has efficacy comparable to in-person care for many conditions; the thousands of mental health apps are largely unvalidated, and their effects, where documented, are modest; generative artificial intelligence is promising but overhyped and raises specific risks; the digital divide threatens to widen the very inequalities digital care claims to reduce.

Conclusions. Digital psychiatry is a real tool where validated, a market bubble where not, and an equity paradox throughout. The right word is augmentation, not replacement: validate before adopting, integrate without delegating, and name the limits precisely.

Keywords: digital psychiatry, telepsychiatry, digital mental health, digital phenotyping, artificial intelligence, confabulation, digital divide, therapeutic alliance.

Introduction: a three-speed field

Digital psychiatry is no longer an experimental hypothesis but a structural reality of contemporary mental health care. The COVID-19 pandemic accelerated its adoption but did not invent it: its foundations had been laid over a decade of research and pilot programmes, and the emergency merely converted into practice what was already prepared on paper. Today, under the label of “digital psychiatry”, a very broad spectrum of tools is gathered together: synchronous telepsychiatry, real-time video consultations; therapeutic applications; passive monitoring through smartphones and wearable devices; and, at the most recent frontier, generative artificial intelligence systems.

To treat these tools as a single block is the error this review means to guard against. The field in fact proceeds at three different speeds, and with three different degrees of evidence. There is a mature, well-validated technology — teleconsultation — whose efficacy stands comparison with in-person care. There is a submerged continent of commercial applications, very numerous and for the most part devoid of any clinical validation. And there is a frontier — generative artificial intelligence — where enthusiasm runs far ahead of the evidence, and where new risks appear. Over all of this hangs a paradox: digital care promises to widen access, yet it risks widening the very inequalities it claims to close. Distinguishing these three speeds — and keeping an eye on the paradox — is the task that follows.

Synchronous telepsychiatry: the mature tool

On real-time teleconsultation the literature is solid and consistent. A now-classic review already concluded that telepsychiatry is effective across diagnostic categories, age groups and geographic settings, with outcomes comparable to those of in-person care (Hilty et al., 2013); and psychiatric video consultation has long been recognized as a reliable mode of care delivery, particularly valuable in rural or specialist-poor settings (Shore, 2015). The pandemic surge consolidated standards that were already under construction, and patient acceptability has proved high, with appreciation for convenience, privacy and the reduction of stigma.

It would, however, be naive to treat it as a perfect equivalent of presence. Clinicians report real limits: the assessment of non-verbal symptomatology, the building of the therapeutic alliance, and the management of crises and risk all require an adaptation of clinical skills, not their mere transposition onto the screen. The most honest formula, which the scientific societies themselves have adopted, is not “substitute” but “complement”: teleconsultation extends the reach and flexibility of care, it does not replace it. It is a mature tool precisely because we now know both its efficacy and its boundaries.

The submerged continent of apps: many, few validated

The picture changes radically when one moves from people — the clinician behind the screen — to software. The mental health applications available in the stores number in the tens of thousands, but only a minimal fraction rests on published, peer-reviewed evidence: the disproportion between the commercial offer and its validation is the most constant, and the most neglected, datum of the entire sector (Torous et al., 2021). This does not mean the apps do not work: the largest available meta-analysis, aggregating one hundred and seventy-six randomized trials, documents real effects on symptoms of depression and anxiety — but of modest magnitude, and heavily conditioned by adherence, engagement and context of use (Linardon et al., 2024).

Two cautions are in order. The first is the “digital placebo effect”: part of the perceived benefit arises from the very act of caring for oneself through a device, more than from the specific therapeutic content of the app (Torous & Firth, 2016). The second is that digital tools work best when integrated into stepped-care models and supervised by a clinician, not when used in isolation. The most rigorous synthesis of the field — an umbrella review of meta-analyses of randomized trials, conducted by an Italian group — confirms this prudent balance: digital interventions for mental health have a real but heterogeneous efficacy, and the methodological quality of the studies often remains the true limiting factor (Crocamo et al., 2025). Between the thousands of apps on the market and the few that are validated runs the distance between a marketplace and a medicine.

Digital phenotyping and passive monitoring

A distinct promise is that of passive monitoring, or digital phenotyping: the continuous collection, through smartphones and wearable sensors, of behavioural signals — sleep, mobility, language, social interaction — from which to infer psychic state. Thomas Insel described it as a “new lens” on mental illness, capable of a continuous, real-world assessment that the periodic interview cannot offer (Insel, 2017). It is a genuinely innovative prospect, and in principle valuable for catching early signals of relapse before they reach the clinical threshold.

Here too, however, the distance between promise and proof is wide. The individual variability of digital patterns is high, predictive algorithms must be validated on populations other than those on which they were built, and the implications for privacy — a continuous stream of intimate data collected passively — are among the most delicate in all of digital medicine. Digital phenotyping is a promising research frontier, not a mature clinical technology; to treat it as such would be premature.

The frontier of artificial intelligence: promise, risk, and a question of words

At the most recent and most clamorous edge of the field stand generative artificial intelligence systems: conversational chatbots, from the first automated agents for cognitive-behavioural therapy to the latest-generation language models. Expectations are enormous — automated triage, continuous support, personalization — but the evidence of clinical efficacy is still scarce and preliminary, and the more discerning literature urges a sharp separation of technological power from the demonstration of therapeutic benefit (Torous et al., 2025). To treat a chatbot as a psychotherapist is, as things stand, an unjustified leap.

And here a question of words must be raised that is anything but secondary, for it shapes the very understanding of the risk. Current usage calls “hallucinations” the errors by which language models generate false but plausible information. It is an inaccurate term, and it is worth saying so with clinical precision: a hallucination is a disorder of perception — a percept without an object — whereas what a language model does, filling a gap with plausible but groundless material, without perception and without intent to deceive, is structurally a confabulation, in the sense neuropsychology has given the term since Korsakoff. The distinction is not terminological pedantry. A system that confabulates does not “perceive” the false: it constructs it with the same fluency with which it constructs the true, and that is why its error is insidious — indistinguishable, on the surface, from a correct statement. In a mental-health context, where wrong information delivered with confidence to a vulnerable user can have grave consequences, adopting the exact term — confabulation, not hallucination — is the premise of an honest assessment of risk, not a lexical affectation. It is the reason why, on these pages, I have argued that the misleading construct of “hallucination” should be rejected in favour of the clinical model of confabulation (Bollorino, 2026a, 2026b, 2026c).

Ethics: alliance, data, accountability

The therapeutic relationship is at the centre of the ethical reflection on digital psychiatry. Research shows that the alliance can be preserved even in the remote setting, but that this requires intentional effort, clear boundaries and an adaptation to digital cues; the presence mediated by the screen alters the phenomenology of the clinical encounter — the perception of time, of space, of intimacy — and calls for new frameworks of empathy and containment (Gaggioli et al., 2022). This is not a technical detail: it is the relational heart of care, which no interface may take for granted.

To this are added the knots of confidentiality and accountability. The handling of sensitive data through third-party platforms raises questions of security and informed consent that regulations — from the European GDPR to local rules — address only in part. And the delegation of clinical functions to algorithms — triage, diagnostic orientation — imposes the principle that any system influencing clinical decisions be transparent, explainable and auditable, and that ultimate responsibility remain human. The ethics of artificial intelligence in psychiatry does not ask us to renounce the tools, but not to let the tools decide: the patient’s autonomy and the clinician’s contextual judgment are not automatable without loss.

The equity paradox: the digital divide

Here one touches the paradox that runs through the whole matter. Digital psychiatry is born with the promise of democratizing access to care; but, if designed without care, it risks reproducing and amplifying existing inequalities. Low-income populations, the elderly, rural dwellers and — a particularly cruel datum — people with severe mental illness face significant barriers: device availability, connectivity, digital literacy (Spanakis et al., 2025). Those who most need care are often those with the least access to the tools meant to provide it.

To this are added cultural and linguistic disparities: platforms and applications are largely designed for Anglophone and Western contexts, and the lack of cultural adaptation, translation and inclusive representation reduces effectiveness and increases dropout. The operational consequence is clear: digital mental health must be designed for equity from the outset — with inclusive interfaces, multilingual support, offline functionality and a participatory design that involves users and communities as co-authors of services, not as their mere recipients. An innovation that widens the divide is not progress: it is a new form of exclusion with an elegant interface.

Limitations and conclusions

This is a selective critical survey, not a systematic review: it has chosen the nodes that best illuminate the central question — the distance between proof and enthusiasm — leaving in shadow extensive territories, from virtual reality in psychotherapy to matters of economics and reimbursement. It shares their limits, beginning with the very rapid obsolescence of a literature that changes faster than research can validate it.

The conclusion, however, can be clear, and lies in refusing a false alternative: that between uncritical enthusiasm, which hails every app as a revolution, and reactionary refusal, which sees in the screen only a threat to the relationship. The real picture is more sober and more useful. Synchronous telepsychiatry is a mature, effective tool, to be integrated stably into practice. Applications are, for the vast majority, a marketplace in search of validation, and are to be treated with the same rigour demanded of any clinical intervention: validate before adopting. Generative artificial intelligence is a promise to be handled with caution and with an exact lexicon, one that calls confabulation what confabulation is. And the digital divide must be addressed as a constitutive part of the project, not as a defect to be corrected after the fact. The word that holds all this together is a single one: augmentation, not replacement. Digital technology can extend the clinician’s gaze and reach; it cannot, and must not, take their place. For at the centre of care, even when it passes through a screen, there remains what no algorithm can give: a presence.

References

American Psychiatric Association. (2021). Telepsychiatry toolkit. APA.

Bollorino, F. (2026a). AI: Il confabulatore perfetto. Genesi, fenomenologia e psicopatologia delle «allucinazioni» artificiali nel confronto col modello umano. Psychiatry on Line Italia.

Bollorino, F. (2026b). La mente di silicio: Anatomia, economia ed etica della confabulazione artificiale. Psychiatry on Line Italia.

Bollorino, F. (2026c). La sindrome di Geppetto: Illusione cognitiva dell’intelligenza artificiale. Psychiatry on Line Italia.

Crocamo, C., Palpella, D., Cavaleri, D., Nasti, C., Piacenti, S., Morello, P., … Carrà, G. (2025). Digital health interventions for mental health disorders: An umbrella review of meta-analyses of randomised controlled trials. The Lancet Digital Health, 7(8), 100878. DOI: 10.1016/j.landig.2025.100878

Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017). Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Mental Health, 4(2), e19.

Gaggioli, A., Riva, G., et al. (2022). Digital presence and the phenomenology of the clinical encounter. Frontiers in Psychology, 13.

Hilty, D. M., Ferrer, D. C., Parish, M. B., Johnston, B., Callahan, E. J., & Yellowlees, P. M. (2013). The effectiveness of telemental health: A 2013 review. Telemedicine and e-Health, 19(6), 444–454.

Insel, T. R. (2017). Digital phenotyping: Technology for a new science of behavior. JAMA, 318(13), 1215–1216.

Linardon, J., Torous, J., Firth, J., Cuijpers, P., Messer, M., & Fuller-Tyszkiewicz, M. (2024). Current evidence on the efficacy of mental health smartphone apps for symptoms of depression and anxiety: A meta-analysis of 176 randomized controlled trials. World Psychiatry, 23(1), 139–149.

Luxton, D. D., Kayl, R. A., & Mishkind, M. C. (2016). Privacy and security in telehealth. Telemedicine and e-Health, 22(3), 214–219.

Shore, J. H. (2015). Telepsychiatry: Videoconferencing in the delivery of psychiatric care. American Journal of Psychiatry, 172(3), 256–262.

Spanakis, P., Wadman, R., Walker, L., Heron, P., Mathers, A., Baker, J., … Peckham, E. (2025). The digital divide in people with severe mental illness: Lessons learned and challenges lying ahead. The British Journal of Psychiatry.

Torous, J., & Firth, J. (2016). The digital placebo effect: Mobile mental health meets clinical psychiatry. The Lancet Psychiatry, 3(2), 100–102.

Torous, J., Bucci, S., Bell, I. H., Kessing, L. V., Faurholt-Jepsen, M., Whelan, P., Carvalho, A. F., Keshavan, M., Linardon, J., & Firth, J. (2021). The growing field of digital psychiatry: Current evidence and the future of apps, social media, chatbots, and virtual reality. World Psychiatry, 20(3), 318–335.

Torous, J., Linardon, J., Goldberg, S. B., Sun, S., Bell, I., Nicholas, J., … Firth, J. (2025). The evolving field of digital mental health: Current evidence and implementation issues for smartphone apps, generative artificial intelligence, and virtual reality. World Psychiatry, 24(2), 156–174.

Vayena, E., Blasimme, A., & Cohen, I. G. (2018). Machine learning in medicine: Addressing ethical challenges. PLOS Medicine, 15(11), e1002689.

World Health Organization. (2021). Global strategy on digital health 2020–2025. WHO.

Yellowlees, P., & Shore, J. H. (2020). Telepsychiatry and e-mental health: Current and future directions. Current Psychiatry Reports, 22(12), 81.