Journal Club: Documentation Is Not Diagnosis: What AI Scribes Teach Us About the Future of Mental Health Care

Artificial intelligence is rapidly entering clinical practice through one of the least controversial doors in medicine: documentation. Ambient AI scribes promise to listen to the clinical encounter, reduce administrative burden, generate notes, and give clinicians back some of the time that electronic health records have taken away. In a healthcare system strained by workforce shortages, burnout, and growing demand, that promise matters. Documentation burden is not a minor inconvenience. It shapes the clinician’s attention, the patient’s experience, and the quality of the medical record.

But in mental health care, documentation is only the beginning. A recent study published in JAMA Psychiatry examined more than 20,000 primary care annual visit notes and compared visits using AI ambient scribes with visits using human scribes or no scribes. The findings were both encouraging and sobering. AI-scribed notes contained more documented neuropsychiatric symptom information across multiple domains. In other words, the technology appeared to capture more of the patient’s mental health experience in the clinical note.

Yet this increase in documentation did not translate into more documented psychiatric action. Visits using AI scribes were less likely to be associated with a depression-related diagnosis, antidepressant prescription, or mental health referral compared with contemporaneous visits without scribes (Castro et al., 2026).

This is not evidence that AI scribes worsen care. The study was retrospective, observational, and cannot prove causality. But it does raise a deeper question for the future of clinical AI: If AI helps us document more, why does that not necessarily lead to better recognition, diagnosis, or management? The answer may lie in one of the most important truths in psychiatry: documentation is not diagnosis.

More words are not the same as more understanding

Psychiatric diagnosis is not a simple act of recording symptoms. It is an act of pattern recognition.

A patient may report insomnia, fatigue, irritability, poor concentration, anxiety, sadness, appetite changes, substance use, restlessness, avoidance, shame, or emotional dysregulation. Each symptom matters. But none of them speaks for itself.

Insomnia may be a symptom of major depression, bipolar disorder, trauma, generalized anxiety, substance use, ADHD, pain, menopause, shift work, or an ordinary response to stress. Poor concentration may reflect ADHD, depression, anxiety, sleep deprivation, trauma, medication effects, substance use, cognitive decline, or burnout. Irritability may be developmental, relational, hormonal, neurobiological, traumatic, personality-related, or part of a mood episode.

The diagnostic meaning of a symptom depends on its timing, recurrence, intensity, context, developmental history, functional impact, associated features, family history, treatment response, and relationship to life events.

That is why psychiatry cannot be reduced to a transcript. A transcript tells us what was said. A note tells us what was documented. A diagnosis requires understanding what the information means. The JAMA Psychiatry study illustrates this gap. AI-scribed notes were longer and contained more text-derived neuropsychiatric symptom signal. But the clinical actions associated with depression did not increase. One interpretation is that AI scribes may help capture the surface area of distress without necessarily helping clinicians synthesize that distress into clinically actionable patterns. That distinction matters.

In mental health, the problem is rarely that no symptoms exist. The problem is that symptoms are fragmented across time, hidden behind somatic complaints, normalized by patients, deprioritized in short visits, or scattered across multiple clinical encounters. The challenge is not simply to hear more. It is to organize, interpret, and contextualize what is heard.

The limits of ambient documentation in mental health

Ambient scribes were not designed to perform psychiatric formulation. Their primary function is to document the visit. That alone can be valuable. Better notes can improve communication, continuity, billing, medico-legal clarity, and clinician experience.

But mental health assessment asks for more than a better note. A clinically meaningful psychiatric assessment often requires information that may not naturally emerge in a routine primary care visit: longitudinal symptom course, prior episodes, age at onset, trauma history, family psychiatric history, developmental history, menstrual or hormonal patterns, substance use trajectories, collateral information, functional impairment, risk factors, protective factors, and prior treatment response. Even when these elements are mentioned, they may not be integrated. A note may say that a patient sleeps poorly, feels anxious, has lost motivation, drinks more alcohol, and struggles at work. But the clinical question is: What pattern does this represent? Is this a depressive episode? An anxiety disorder? ADHD with secondary demoralization? Bipolar depression? Trauma-related hyperarousal? Substance-induced symptoms? Adjustment disorder? A medical or medication-related presentation? Something else? This is where documentation stops and clinical reasoning begins.

The JAMA Psychiatry findings should therefore not be read as a failure of AI scribes. They should be read as a sign that the next stage of clinical AI must move beyond transcription and note generation.

The future of AI in mental health will not be defined by who captures the longest note. It will be defined by who helps clinicians transform complex information into better clinical understanding.

Psychiatry needs structured context, not just captured conversation

One of the central lessons of mental health care is that the most important diagnostic information is often distributed. It may be distributed across:

• what the patient says today,

• what they forgot to mention,

• what they normalized because it has always been present,

• what appeared in earlier visits,

• what a family member or partner has noticed,

• what changed over months or years,

• what happens only during certain hormonal phases, seasons, stressors, or relational contexts,

• and what becomes visible only when symptoms are organized longitudinally.

This is why primary care clinicians face such a difficult task. They are often asked to recognize complex psychiatric patterns during brief, cross-sectional encounters filled with competing priorities. A patient may present with headaches, gastrointestinal symptoms, fatigue, pain, insomnia, or concentration problems, while the underlying psychiatric pattern remains only partially visible.

AI scribes may document more of the encounter. But if the encounter itself does not elicit the right information, or if the information is not synthesized afterward, better documentation may still leave the core problem untouched.

Mental health care therefore needs tools that improve the information environment before, during, and after the visit. That means structured intake. It means longitudinal history. It means collateral information when appropriate. It means measurement that supports—but does not replace—clinical judgment. It means organizing symptoms by time, context, severity, function, risk, and differential diagnosis. It means helping clinicians see the pattern, not just the paragraph.

From administrative AI to clinical intelligence

The first wave of healthcare AI has largely focused on administrative relief: documentation, summarization, inbox support, coding, and workflow automation. These are important problems. Clinicians need relief. But psychiatry and mental health care also need a second wave: clinical intelligence.

Clinical intelligence does not mean replacing clinicians. It means supporting the cognitive work that clinicians already do under difficult conditions. It means helping them gather better information, notice what may be missing, identify diagnostic possibilities, compare competing hypotheses, track risk, and make more informed decisions. This distinction is essential. A documentation tool asks: “What should the note say?” A clinical reasoning tool asks: “What does this pattern suggest, what else could explain it, what information is missing, and what should the clinician consider next?” Those are very different questions. The first improves the record. The second improves the conditions for judgment. For mental health, this difference is fundamental. Psychiatric diagnosis is not an output of transcription. It is a synthesis of narrative, chronology, function, risk, context, and clinical expertise.

Why this matters for health systems

Missed or delayed psychiatric diagnosis is not only a clinical problem. It is a system problem.

When mental disorders are not recognized early, patients may cycle through repeated visits, investigations, partial explanations, ineffective treatments, or fragmented referrals. Primary care clinicians carry much of this burden, often without enough time, specialty support, or structured tools.

AI scribes may reduce documentation workload, but they do not automatically solve the diagnostic complexity of mental health. The JAMA Psychiatry study suggests that even when psychiatric symptoms are more visible in notes, the pathway from documentation to action remains uncertain. For health systems, this points to a clear strategic lesson: documentation infrastructure and diagnostic infrastructure are not the same thing.

A health system that adopts ambient scribes may improve note generation. But if it wants to improve mental health detection, triage, risk recognition, and care planning, it may need tools designed specifically for psychiatric complexity. That means technology built around clinical patterns rather than generic documentation.

Toward the next standard in mental health assessment

At Aion, we believe the future of mental health AI will be defined by a shift from passive documentation to structured clinical synthesis. Elyx is being developed around that principle. The goal is not to automate diagnosis or replace the clinician. It is to support better assessment by helping organize the information that psychiatric reasoning depends on: symptoms, chronology, functioning, context, differential diagnosis, comorbidity, risk, protective factors, and longitudinal change.

This is where mental health AI can become clinically meaningful. Not by producing more text. Not by making the medical record longer. Not by turning every conversation into a polished note. But by helping clinicians identify the patterns that matter.

The recent evidence on AI scribes should encourage the field. It shows that AI can change documentation. But it should also humble us. More documentation does not automatically mean better diagnosis, better treatment, or better outcomes. The next frontier is not simply listening. It is understanding.

Reference

Castro, V. M., McCoy, T. H., Verhaak, P., Ramachandiran, A., & Perlis, R. H. (2026). Psychiatric documentation and management in primary care with artificial intelligence scribe use. JAMA Psychiatry, 83(3), 281–286.

Article - 2026 - Psychiatric Documentation and Management in Primary Care With Artificial .

Download • 599KB