New technology, old problems

Despite a rapidly expanding body of research on artificial intelligence applications in mental healthcare, artificial intelligence has yet to produce clinical decision-support tools that meaningfully transform psychiatric practice. ^{Reference Sun, Lu, Shao, Han, Xia and Zheng2} By contrast, clinical artificial intelligence has succeeded chiefly in image-based specialities, ^{Reference Lim, Regillo, Sadda, Ipp, Bhaskaranand and Ramachandra3,Reference Spratt, Tang, Sun, Huang, Chen and Mohamad4} driven by the maturity of computer vision, established frameworks for aggregating massive data-sets and the high signal-to-noise ratio inherent in pathological images.

Psychiatry occupies a distinctive position in medicine. The aetiology of most mental disorders remains poorly understood, and the field lacks robust biomarkers for diagnosis, prognosis or treatment monitoring. Clinical practice instead relies heavily on subjective patient reports and clinician judgement; validated psychometric instruments help but are inconsistently applied in routine care, ^{Reference Gilbody, House and Sheldon5} and diagnostic criteria for specific diagnoses have moderate test–retest reliability. ^{Reference Xie, Nordgaard, Sheldrick, Ahmad, Gomes and Frey6}

These limitations pose fundamental challenges for data-driven approaches. Artificial intelligence models require reliable ground-truth labels, yet psychiatric diagnoses and symptom scores are heterogeneous, often imprecise and linked only weakly to underlying biology. Shifting to ‘harder’ proxy outcomes, such as hospitalisation or sickness absence, may partially mitigate this issue but captures only the coarse aspects of illness. Crucially, the proportion of clinically meaningful variance explained by candidate biomarkers in psychiatry remains far lower than in most other medical disciplines. Whereas an ophthalmologist can confidently diagnose severe diabetic retinopathy from a single retinal image, a psychiatrist cannot distinguish a magnetic resonance imaging scan of a patient with established schizophrenia from that of a healthy control: the subtle structural differences emerge only through large-scale statistical aggregation. Compounding this problem, the infrastructure required to collect large, longitudinal, high-quality data-sets is not routinely embedded in psychiatric care. Together, these constraints limit the scope for near-term artificial intelligence-driven transformation.

Data sources for psychiatric artificial intelligence

Because modern artificial intelligence systems are data-intensive and particularly effective with unstructured data such as text or images, it is essential to consider which data sources can realistically support progress.

Large international initiatives aim to assemble deeply phenotyped cohorts incorporating genomics, neuroimaging and biochemical data, with artificial intelligence invoked to integrate these high-dimensional data-sets and uncover subtle predictive signals that might elude human clinicians. Important caveats apply, however. Current artificial intelligence systems lack explicit causal frameworks and may exploit spurious correlations or encode existing biases, particularly affecting minorities. ^{Reference Hofmann7} Explainability remains limited, and models struggle to clarify pathophysiological mechanisms. Calibration is a further concern, because poorly calibrated predictions can misrepresent true clinical risk. Finally, models reliant on data collected in resource-expensive research settings may prove clinically impractical if their marginal benefit does not justify the cost of deployment.

Beyond traditional biological data, the Internet of Things (wearable devices in particular) offers an increasingly important source of information. Smartphones and smartwatches are widely adopted and generate continuous, longitudinal data with minimal user burden. Although such devices can be incorporated into research studies, far larger data-sets already exist within the general population, held by private companies. ^{Reference Narayanswamy, Xu, Heydari, Abdel-Ghaffar, Guerard and Vaillancourt8} These data are not readily available to academic researchers and typically lack detailed psychiatric phenotyping, making them arguably more suited to the study of mental well-being or dimensional traits than categorical diagnoses. Relatedly, the emergence of large pre-trained models has lowered the barrier to using such data; models pre-trained on large, generic data-sets can now be adapted to smaller, domain-specific cohorts through fine-tuning, a paradigm popularised by LLMs but increasingly feasible for other modalities, including wearable-derived actigraphy. ^{Reference Corponi, Li, Anmella, Valenzuela-Pascual, Mas and Pacchiarotti9}

Scientific progress and integrity

Scientific progress depends not only on technology but also on how research itself is conducted. Open data access, within ethical and legal constraints, has driven much of artificial intelligence’s progress and should be prioritised in psychiatric research too. Consensus on which tasks are clinically meaningful, for example, detecting at-risk populations from electronic health records, ^{Reference Crowley, Parkin, Rocheteau, Massou, Friedmann and John10} coupled with curated open-access datasets, would enable meaningful model comparison and cumulative progress.

Artificial intelligence is increasingly utilised as a tool supporting researchers, fundamentally reshaping science and not always for the better. Under constant pressure to publish, researchers increasingly use LLMs as a shortcut, effectively treating fallible models as authoritative sources. Hallucinated citations (references to papers that never existed) are the most visible symptom of this practice ^{Reference Resnik and Hosseini11} but are relatively easy to catch. Far more insidious is LLM-assisted generation of code or statistical methodology, which is time-consuming to verify and demands technical expertise that many peer reviewers lack; a flawed pipeline can pass review looking polished and complete, wasting far more time than a hallucinated reference ever would. The growing use of LLMs within peer review itself raises questions about accountability, ^{Reference Naddaf12} with the looming possibility of a large portion of research being both generated and evaluated by artificial intelligence. Mitigating these risks will require setting clear standards: mandatory code and analysis pipeline-sharing, disclosing artificial intelligence assistance with explicit boundaries on permissible use and reviewing capacity built specifically to scrutinise artificial intelligence-assisted work.

LLMs and mental health

Following their public release in 2022, LLMs rapidly entered mainstream use and many individuals now turn to these systems for emotional support outside any formal healthcare context. Researchers, meanwhile, are testing their capabilities in psychotherapy and triage. Given the scale of unmet need and the limited availability of trained therapists, the appeal is intuitive but the evidence base remains preliminary and mixed.

A randomised trial has suggested benefit from a chatbot built on a language model specifically trained to discuss mental health symptoms. ^{Reference Heinz, Mackin, Trudeau, Bhattacharya, Wang and Banta13} However, claims that LLMs will soon match human psychotherapists are not supported by the evidence: one study found that interacting with a random human peer was significantly more effective at reducing psychological distress and loneliness. ^{Reference Li, Folk, Singh, Ungar and Dunn14} A large observational study, meanwhile, reported a positive association between LLM use and depressive symptoms; ^{Reference Perlis, Gunning, Uslu, Santillana, Baum and Druckman15} causality cannot be inferred from this design, but the finding underscores the need for caution rather than enthusiasm.

Safety concerns compound this uncertainty. Anecdotal reports and case series describe the exacerbation of delusional thinking in vulnerable individuals and, in rare cases, associations with suicidal ideation or behaviour. ^{Reference Clegg16} These are no longer merely clinical concerns: high-profile litigation has targeted major technology firms over suicides linked to conversational artificial intelligence, ^{Reference Tabachnick17} and legislation such as California’s SB 243 has begun establishing legal boundaries for artificial intelligence-driven mental health interactions. ¹⁸

Beyond individual harms, there are structural risks. From a sociological standpoint, reliance on conversational agents may displace meaningful human relationships, which are central to mental well-being, ^{Reference Holt-Lunstad19} thereby compounding the very psychosocial vulnerabilities that drive psychiatric distress. Sycophantic LLM behaviours (the tendency to reinforce a user’s existing beliefs and prioritise flattery over honesty) may further decrease prosocial intentions and promote dependence. ^{Reference Cheng, Lee, Khadpe, Yu, Han and Jurafsky20}

Against this backdrop, LLMs are probably better understood not as a substitute for therapy but as a potential tool for managing demand and waitlists, and, more prosaically, administrative work. Ambient artificial intelligence scribes are already the most widely adopted LLM application in healthcare, freeing clinical time and cognitive resources by reducing after-hours documentation burden without compromising billing accuracy. ^{Reference Shah and Garcia21} Nevertheless, routine reliance on artificial intelligence-generated notes can erode independent clinical formulation, prematurely shaping a trainee’s reasoning before they can critically evaluate the output. Used judiciously, however, ambient documentation remains a grounded, non-aspirational application that empowers rather than supplants human practitioners.

Implications for research and policymakers

Artificial intelligence is not a magic solution to psychiatry’s intrinsic challenges: the remarkable capabilities of LLMs and the richness of wearable data do not, by themselves, circumvent limited mechanistic understanding, heterogeneous clinical labels or the absence of validated biomarkers. Experience from other medical disciplines suggests that clinical impact emerges only when artificial intelligence is developed alongside these fundamentals, and alongside how the research itself is conducted. This means greater openness, shared benchmark data-sets, closer collaboration between computer scientists and mental health researchers, rigorous standards for artificial intelligence-assisted analysis and sustained investment in infrastructure for large, multimodal, longitudinal studies.

These are problems for the research community to solve. What happens once solutions exist is a question for policymakers. Given the intrinsically interpersonal vocation of psychiatry, the goal of artificial intelligence should not be to replicate human judgement but to augment it: supporting clinicians in remote monitoring, assisting objective measurement of treatment response and helping personalise care. Such aims are not self-enforcing and could remain rhetoric. Under pressure to cut public spending, governments may have little incentive to add artificial intelligence capacity on top of existing staff, since doing so would mean spending more, not less. In this context, a possible outcome could be thinning of clinical ranks, leaving fewer staff to supervise systems that outpace human review: the so-called ‘reverse-centaur’ setting, ^{Reference Doctorow22} where the human works at the machine’s pace rather than the other way around. Vigilance for the rare error tends to degrade precisely as artificial intelligence grows more reliably correct and, when something goes wrong, responsibility could be placed on the nearest human operator. A hospital can buy an artificial intelligence triage tool and cut staff in the same budget cycle. Whether that happens is a choice for policymakers, not something the model architecture determines.