What psychiatrists and standardised patients told researchers when an ambient AI scribe joined the consultation.
Documentation burden is a significant contributor to psychiatric clinician burnout (Domaney et al., 2018), and administrative work consumes a substantial share of physicians' working hours (Woolhandler & Himmelstein, 2014).
A psychiatric consultation demands attention on several fronts at once: what is said and what is not, emotional tone, risk, the therapeutic relationship, and the clinical formulation. It is also a setting where recording a consultation carries particular weight.
Ambient AI documentation is moving from pilots into routine care. A recent systematic review points to reduced documentation workload and improved clinician workflow, while noting that the evidence base is still limited and that note quality varies by tool and setting (Sasseville et al., 2025). How clinicians and patients experience these tools in practice, particularly in psychiatry, has been far less studied.
Because this is exploratory, simulation-based work, we read the themes as hypotheses and implementation cautions, not confirmatory evidence.
The patient voice here comes from trained standardised (simulated) patients, not people receiving psychiatric care. It is valuable for surfacing risks and hypotheses, and not a substitute for lived experience.
Clinicians described more authentic human connection, and simulated patients noted more eye contact and faster recognition of non-verbal cues. Cues missed under keyboard documentation were "picked up on the first time" with the scribe.
Clinicians framed the impact in terms of their own mental health, describing cognitive and emotional relief. One described being able to ask about trauma history in a way they otherwise might not have.
Clinicians valued the system translating colloquial language into psychiatric terminology, and prompting for assessments not yet covered, for example flagging that forensic history had not been assessed.
Participants noted instances of hallucination, which prompted healthy vigilance. Clinicians adapted, for example by speaking non-verbal cues aloud so the scribe would capture them. Concerns remained about reliability in noisy environments.
The theme that sets psychiatry apart. Recording carries particular stigma, and capacity to consent varies. Participants drew a clear boundary: comfort with AI for documentation, but not for diagnosis or medication decisions.
Both groups expressed a strong preference for AI-assisted consultations, conditioned on transparent consent, human oversight, and the ability to customise to preserve each clinician's voice.
Collectively, the study suggests ambient AI scribes may support more present, patient-centred psychiatric consultations by reducing documentation burden, while raising consent, trust and specialty-specific considerations that require careful handling.
"Ten minutes in, it was like talking to a friend."
A standardised patient, describing the AI-assisted consultation. In psychiatry, presence is not a side effect of the tool; it is part of the treatment.
The themes already carry interpretation, so we do not restate them. These four lessons set out the practical implications.
This is the lesson most specific to the specialty, and the one generic AI-scribe guidance serves least well. In medicine, the bar for safety rises with the stakes, and few settings raise those around recording and consent as sharply as psychiatry. Stigma, the sensitivity of disclosure, and variable capacity to consent mean consent cannot be a one-size-fits-all process.
Transparent, plain explanation before recording; a genuine ability to decline or pause; explicit attention to capacity; and a firm boundary that the tool documents the encounter and does not make diagnostic or medication decisions.
The strongest signals were not about speed. They concerned what changed when documentation was handled ambiently: greater attention in the room, the capacity to ask more difficult questions, and cognitive relief clinicians described in terms of their own mental health. In psychiatry these matter profoundly in their own right, and they are easily missed if implementation tracks only time saved.
Implementation measures what changes in the consultation, such as presence, rapport and the ability to stay with difficult material, and treats clinician wellbeing as an outcome worth protecting.
Clinicians valued terminology translation and prompts for assessments they hadn't covered. They also noticed hallucinations, and were unanimous that human oversight is essential. Notably, visible imperfection helped keep reviewers engaged rather than complacent. Many apparent errors reflect questions of context the system could not resolve, and it is the clinician, not the tool, who judges whether an output belongs in the record.
Key details are easy to verify and correct, the workflow keeps the clinician reviewing rather than rubber-stamping, and the system stays out of decisions that belong to the clinician.
The study's most important caveat is structural: the patient voice here is simulated, not lived. That makes the findings strong for generating hypotheses and weak for concluding anything about real patients. Trust in a setting like this is earned with evidence from that setting, and re-earned as use continues.
Simulation work is followed by careful real-world evaluation with actual psychiatric patients, templates are specialty-specific rather than generic, and vendors support that evidence path rather than shortcutting it.
The questions to ask a vendor go beyond accuracy and speed. This study is useful precisely because it surfaces the experiential and consent issues that demos rarely show.
How does the workflow support transparent consent, the ability to decline or pause, and attention to capacity? How is sensitive material handled?
Are there psychiatry-specific templates and terminology handling? Can the tool keep non-clinical therapeutic exchange out of the note appropriately?
Does the interface keep the clinician engaged in review and make key details easy to verify and correct? How are possible hallucinations surfaced?
Is it clear the tool documents the encounter and does not make diagnostic or prescribing decisions?
Will the vendor support real-world evaluation with actual patients, share quality information, and respond to independent research?
Do the presence and wellbeing effects persist beyond initial novelty?
How do these experiences hold up with patients actually receiving psychiatric care, rather than simulated patients?
Does the picture change across multiple centres and real-world conditions?
How are consent and capacity best handled in routine psychiatric practice?
Lyrebird funded this independent research because the experiential questions in psychiatry, such as presence, trust, consent and stigma, warrant serious study, including where the findings are not straightforward.
We stand by the principle that human judgement remains the gold standard for assessing the quality of a clinical note, and that oversight matters most where the stakes are highest. Our framework for clinical note quality evaluation reflects that. Results from simulated consultations do not transfer directly to live psychiatric care, and we make no claim to have resolved the consent and trust questions this study raises.
This analysis was prepared by the clinical and research leadership team at Lyrebird Health, who are committed to objective interpretation of research findings and transparent discussion of both benefits and limitations.
Lyrebird is the clinical AI platform for Australian clinicians. Ambient scribing is one core feature; in psychiatry that means specialty templates, oversight built into review, and consent handled with care.
