Is AI note-taking safe in therapy?

Is AI note taking safe in therapy?

When therapists ask whether AI note-taking is "safe," they are usually asking two different questions at once. The first is technical: does this tool protect client data the way a Business Associate Agreement and HIPAA require? The second is clinical: does relying on AI to draft a note change the quality, defensibility, or integrity of the documentation itself? Both questions matter, and conflating them is where many practices get into trouble.

The honest answer is that AI note-taking can be safe, but safety is not a property of the software alone. It is a property of how a practice selects, configures, supervises, and audits the tool over time. That distinction matters more than any single vendor's compliance badge.

The Documentation Burden That Made AI Attractive in the First Place

It helps to understand why practices are considering these tools at all. Documentation has become one of the heaviest, least sustainable parts of clinical work, a load we've traced in detail in how administrative burden is breaking mental health care. Even so, clinician adoption of AI has run well ahead of willingness to let it near the clinical record, and that hesitation is reasonable.

A progress note is not just an administrative artifact. It is a legal record, a clinical reasoning trail, and in many states a document with heightened statutory protection because it touches mental health and substance use treatment. Any tool that touches it inherits that weight, regardless of how polished its interface looks.

Do AI Scribes Actually Save Time? What the Evidence Shows

The efficiency case for AI scribes is real, but it is more modest than marketing pages suggest. A large multisite study following roughly 1,800 clinicians found that those using AI scribe technology saved about 16 minutes of documentation time and spent 13 fewer minutes in the medical record for every eight hours of patient care ^[1]. That is a meaningful reduction over a full caseload, but it is not the dramatic transformation some vendors imply, and the same research noted inconsistent use patterns across clinicians, meaning the benefit depends heavily on how a tool is integrated into daily workflow rather than the tool's existence alone.

There is also emerging evidence on burnout specifically. A qualitative study of ambient AI scribes in psychiatric consultations found that clinician burnout dropped significantly within the first month of adoption ^[2]. For practice owners managing retention and clinician wellbeing, that is a legitimate operational argument for exploring these tools, separate from any efficiency argument.

Where the Real Safety Risk Lives: Clinical Judgment, Not Just Data Security

Most compliance checklists focus on encryption, data retention, and signed agreements. Those are necessary, but they are not where most clinical risk actually lives. The more consequential risk is subtler: what happens when a clinician starts trusting an AI-generated draft more than their own clinical reasoning.

A pilot study of AI scribe use among consultation-liaison psychiatrists found that these tools were reliable for capturing patient history and prior treatment information, but researchers concluded that clinicians should continue to author their own mental status exams, assessments, and treatment plans directly, because those sections reflect the clinician's independent judgment rather than a summarized transcript ^[3]. In other words, the parts of a note that carry the most clinical and legal weight are precisely the parts AI is least equipped to generate safely.

This has a direct operational implication for practice owners: if a tool auto-populates assessment or treatment plan sections without requiring substantive clinician revision, that is a workflow design flaw, not just a convenience feature.

Who Owns the Progress Note? The Accountability Gap

Regardless of which tool a practice adopts, professional responsibility for the record does not transfer to the software. The American Psychological Association's ethical guidance for AI in clinical practice is explicit that AI should augment, not replace, human decision-making, and that psychologists remain responsible for final decisions rather than accepting AI output uncritically ^[4]. The diagnosis and the clinical formulation are yours, and no AI tool changes that professional obligation, regardless of how clinically fluent its output sounds.

This has practical consequences for practice policy. Every AI-generated note should be treated as a draft requiring clinician review before it becomes part of the permanent record, not a finished product. Practices that skip this step are not saving time, they are transferring risk from the clinician's calendar to the clinician's license.

Transparency with clients matters here too. Recent APA reporting on AI adoption in clinical settings recommends that clinicians disclose how they use AI tools directly to clients, treating a notes-summarization tool the way one might treat feedback from a supervisor rather than as an authoritative record of the session ^[5]. That disclosure is not just an ethical nicety. It protects the therapeutic relationship and gives clients the ability to raise concerns before a tool becomes routine.

A Practical Evaluation Framework for Practice Owners

Rather than asking "is this AI tool HIPAA compliant" as a yes/no question, practice owners are better served evaluating tools across several dimensions simultaneously:

• Data handling: Does the vendor offer a signed Business Associate Agreement, and does it explicitly state whether session data is used to train models?
• Clinical accuracy: Does the tool distinguish between sections it can draft reliably (history, subjective report) and sections that require full clinician authorship (assessment, treatment plan, risk formulation)?
• Review workflow: Does your practice require every AI-generated note to be reviewed and edited before signing, and is that requirement documented in policy?
• Client transparency: Do clients receive disclosure about AI use as part of informed consent, with a documented option to decline?
• Audit readiness: Can you demonstrate, if audited, that a human clinician exercised judgment on every note, not just that a tool produced output?

None of these questions have a universal right answer. They depend on your practice's client population, state-level record protections, and clinical specialization. Where state record protections are in play, run your documentation policy past qualified legal counsel before you finalize it.

This checklist is the documentation-specific slice of a broader discipline. For the wider landscape, from scribes to client-facing chatbots, our clinician's framework for evaluating AI mental health tools applies the same questions across every category.

The Bottom Line

AI note-taking is not inherently unsafe, but it is also not automatically safe simply because a vendor claims HIPAA compliance. The evidence suggests real efficiency gains and genuine burnout reduction are possible, but only when the tool is deployed with clear boundaries around what it drafts, mandatory clinician review, and transparent disclosure to clients. The practices that will benefit most from these tools are the ones that treat AI as a supporting instrument in an accountable clinical workflow, not as a replacement for the judgment that makes therapy therapy in the first place.