The Psychiatry Note Is Not a Transcript

The opening scenario is fictional. It illustrates a documentation problem and does not reproduce a patient encounter, a complete suicide risk assessment, a clinical formulation, or a disposition decision.

Imagine that a patient says, "I do not care whether I wake up tomorrow." She denies a plan or intent to harm herself. After an assessment that is not reproduced here, the psychiatrist documents a passive wish for death, the relevant protective factors, and the agreed follow-up.

An AI scribe drafts a cleaner sentence: "Patient denies suicidal ideation."

Nothing in that sentence looks invented. The error is compression. A denial of plan and intent has become a denial of the thought itself. The draft sounds ordinary, and that is what makes it dangerous. A covering clinician who reads it later receives a different clinical fact from the one the psychiatrist assessed.

The words were captured. The note is still not what happened.

An ambient scribe and a psychiatrist are doing different work. The scribe turns speech into a draft. The psychiatrist decides what the exchange meant, who supplied the information, what belongs in the record, and how the evidence supports the assessment and plan.

In most of medicine, compression is a feature. A useful note cannot reproduce every word. In psychiatry, however, compression often carries clinical reasoning. "No intent" is not "no thought." A patient's report is not a family member's report. A provisional explanation is not a diagnosis.

The emerging evidence supports neither blanket enthusiasm nor blanket rejection. In a 2025 randomized trial involving 238 outpatient physicians across 14 specialties, one of two ambient scribe products reduced time in the note compared with usual care, while the other did not produce a statistically significant reduction. Both product groups showed possible improvement on some secondary measures of work experience, although the authors said those findings need confirmation in larger multicenter trials. Clinicians reported occasional inaccuracies. The authors concluded that these require ongoing vigilance.

A Veterans Health Administration study, published online April 17, 2026, and in the June 2026 issue of Annals of Internal Medicine, compared 11 ambient scribe tools with 18 human note takers. Thirty blinded raters assessed notes from five standardized primary care cases using a modified Physician Documentation Quality Instrument. Human-generated notes scored higher in all five cases.

Neither study was designed to answer the central psychiatry question. They do show why "AI scribe" should not be treated as a single intervention with a single quality profile. Product, setting, task, and review process matter.

Psychiatry-specific evidence remains thin. A 2026 exploratory qualitative study involved eight psychiatrists and four standardized patients after simulated psychiatric consultations. Participants saw the possibility of less documentation burden and greater presence in the encounter. They also identified conditions for adoption that included transparent consent, clinician training, human oversight, specialty-specific templates, and evaluation with real patients.

The limitations are substantial. This was a small, single-center simulation in Dubai, and it was the participants' first exposure to the technology. AI-assisted consultations always occurred second, which the authors said confounded the findings with practice effects and scenario familiarity. The research team knew the parent study's positive findings, which may have influenced interpretation. The study also received funding support from AI scribe vendor Lyrebird Health, although the authors reported that the funder was not involved in the study or publication decisions. It is useful as a map of questions, not proof that any product is safe or effective in routine psychiatric care.

For a practice evaluating an AI-generated psychiatric note, four elements deserve deliberate review.

Psychiatric language is full of distinctions that ordinary summarization can erase.

"I sometimes wish I could disappear" is not interchangeable with a documented intent to die. Feeling watched is not automatically a fixed delusion. Sleeping four hours because of a new baby is not the same history as a reduced need for sleep. "My partner thinks I drink too much" is not a direct admission by the patient.

A transcript can preserve every sentence and still fail as a clinical record. The reviewer has to ask what transformation occurred between the exchange and the draft.

Did the system turn a quotation into a clinical label? Did it convert a denial of one part of a phenomenon into a denial of the whole phenomenon? Did "possible," "consider," or "unclear" become a confident conclusion? Did it merge a historical symptom with a current one? Did it smooth over a contradiction that the next clinician needs to see?

The psychiatry simulation produced a useful edge case. A clinician asked, "could you hear any voices?" The scribe documented that the patient "denied auditory hallucinations." The clinician valued the translation into psychiatric terminology. The authors then raised the other side of the problem: polished terminology can imply that a particular symptom was explicitly assessed, even when the question took a different form or was not asked. The lesson is not that translation is always wrong. It is that the reviewer must distinguish faithful compression from inferred examination.

Certainty can disappear during compression. Generated prose tends to sound complete even when the source material was tentative. Review the words that carry degree and time: sometimes, recently, previously, possibly, denies, reports, appears, and according to. Small changes in those words can change the clinical meaning of a sentence.

Psychiatric encounters may contain several voices: the patient, a relative, a therapist, a prior record, and the clinician. Capturing sound does not guarantee that the draft will preserve who said what.

"The patient has become increasingly isolated" may be the patient's report, a relative's concern, a conclusion in an earlier note, or an inference introduced by the draft. Those sources are not interchangeable.

Attribution also matters in the mental status examination. A generated note should not turn a domain that was not assessed into a normal finding. In the psychiatry study, one clinician reported that the tool recorded a standardized patient's affect as "euphoric" when it was not. That was not compression. It was an unsupported clinical finding.

The same study showed the useful reverse case. The tool marked that forensic history had not been assessed. A truthful omission flag may help a clinician notice a gap. It does not mean the domain was assessed, and it should not become a normal finding. A limited telehealth observation should not read like a comprehensive examination. The note should distinguish the patient's account from the psychiatrist's observation and interpretation.

Review attribution where the note addresses adherence, substance use, aggression, functioning, symptoms, and safety. For every material statement, ask: Who supplied this information? What did the clinician directly observe or assess? If the note does not make the answer clear, the sentence is not finished.

An ambient system may capture more than a psychiatrist would ordinarily chart. That creates a selection problem, not an obligation to preserve everything.

Psychiatric conversations can include trauma, family conflict, sexual history, substance use, abuse, legal concerns, and thoughts of suicide or violence. Some details are necessary to understand diagnosis, risk, treatment, or follow-up. Others are incidental or excessively detailed for the record.

Ask whether each detail serves a legitimate clinical purpose in this note. Review intimate narrative, long quotations, third-party information, speculative language, and details whose relevance is unclear. Do not let aggressive summarization remove context needed to understand the assessment or plan.

This is a clinical judgment, not a universal deletion rule. The right level of detail depends on the encounter, the record's purpose, applicable requirements, and the psychiatrist's judgment. The scribe can propose text. It cannot make the inclusion decision.

The assessment and plan are not clerical leftovers. They are where the psychiatrist connects the history, examination, uncertainty, risk, diagnosis, treatment, and follow-up.

Read those sections as if the rest of the draft were hidden. Do they reflect the reasoning actually used? Does the plan match what was discussed? Are changes, monitoring, referrals, follow-up, and safety-related actions stated with the correct scope? Did the draft introduce a recommendation or omit a decision?

Fluent prose can conceal a missing inferential step. A list of symptoms followed by a diagnosis is not a formulation. A risk label is not its supporting assessment. A templated plan is not evidence that options, tradeoffs, or instructions were discussed.

If the assessment could have been generated from a generic template without this encounter, it is not ready to sign.

The four elements above describe the note. The next four questions describe the system around it.

Automatic signing should not be possible. Automatic filing also deserves scrutiny if filed text can be treated as final, copied forward, released, or acted on before review. The workflow should identify which clinician owns the draft, what status it carries, and what act converts it into an attested note.

"The clinician will review it" is an intention. The practice needs a visible control.

The practice should define its disclosure, consent, and refusal workflow before the first recorded encounter. Requirements vary by jurisdiction, setting, technology, policy, and data handling. A general intake notice may not answer what the patient reasonably wants to know at the point of use.

Patients should receive a clear explanation of what the tool does, what it captures, how the output will be used, and what alternative is available if they decline. The clinician must know what to say, and the practice must be able to honor the answer.

The simulation illustrates the gap between a consent principle and a working process. When the eight psychiatrists were asked whether they had sought permission to record, only two said they had. That is not an estimate of consent practice in routine psychiatry. It is a concrete warning that a recommended safeguard can disappear unless the workflow makes the act visible.

A general accuracy claim is not a psychiatric validation. Test the deployed product, configuration, microphone setup, template, and integration against the encounters it will actually face. Include negation, uncertainty, multiple speakers, interruptions, changes over time, sensitive detail, and a difference between what a patient says and what a clinician observes.

The VHA comparison used simulated primary care encounters. A psychiatric practice should test under the conditions its own notes must survive, including collateral speakers, emotionally charged disclosure, long narrative answers, and changes in certainty during the encounter.

Audit more than grammar. Look for changed meaning, unclear attribution, unsupported certainty, inappropriate inclusion, important omission, and a plan that does not reflect the encounter. Track corrections and patient concerns. Define what would pause use while the practice investigates.

If the product offers risk flags or prompts, evaluate them separately. A flag may draw attention to text. It is not a psychiatric assessment, and it should not be allowed to create the appearance that a risk domain was assessed when it was not.

When the HIPAA Rules apply, a third-party service that creates, receives, maintains, or transmits electronic protected health information on behalf of a covered entity will generally require an appropriate business associate agreement. HHS Office for Civil Rights guidance states that a covered entity or business associate that uses a cloud service provider to process or store electronic protected health information without a HIPAA-compliant business associate agreement is in violation of the HIPAA Rules.

OCR's business associate guidance lists an independent medical transcriptionist, or the vendor of an app providing transcription services to a physician, among its examples of business associates. Its AI example is narrower: a third-party vendor chatbot on a patient portal handling symptom assessment, reminders, and scheduling. OCR has not addressed ambient scribes by name.

A contract described as "HIPAA compliant" is not the end of the review. The practice should map what is captured, where it is sent, who can access it, how long it is retained, how it is deleted, whether subcontractors receive it, and whether the vendor may use it to improve or train any system. Those questions may be addressed across the business associate agreement, service terms, privacy terms, security documentation, and negotiated contract language.

Do not place identifiable patient information into an unapproved public AI service.

Ambient documentation may reduce clerical work and return attention to the encounter. The randomized evidence suggests that benefit is possible, but not uniform across products. Early psychiatry-specific work suggests that clinicians and simulated patients can see value, while also asking for consent, oversight, training, and specialty-specific evaluation.

Those are implementation questions. They do not change the act of attestation.

The psychiatrist's signature should mean more than "I scanned what the system produced." It should mean that the final note preserves the patient's meaning, distinguishes its sources, includes clinically appropriate detail, and records the clinician's own reasoning.

An AI scribe can draft a document. It cannot close the distance between a transcript and a psychiatric record. That remains the psychiatrist's work.

This article provides general educational information for psychiatric clinicians and practice leaders. It is not clinical, legal, privacy, compliance, or patient-specific advice. Requirements concerning consent, recording, disclosure, health information, retention, and documentation vary by jurisdiction, setting, contract, and technology.