What Your AI Note Taker Isn't Telling You About Liability

AI note taking software has become a routine part of practice for many clinicians, and the appeal is easy to understand. They save time and cut down on after-hours charting. But there's a gap between what these tools are marketed to do and what they actually create: a permanent, detailed record that behaves very differently from the notes clinicians have written by hand for decades.

A transcript is a different kind of record than a note

For most of the history of talk therapy, a session note has been a clinician's summary. It reflects professional judgment about what matters clinically: the presenting concern and the plan of care. A long pause, or a comment raised and then withdrawn, rarely makes it onto paper, because clinical judgment filters it out before anything gets written down.

An AI transcript doesn't filter. It captures every word spoken. That can be useful, but it also changes the legal character of the record. A clinical summary reflects professional judgment, and courts generally give that judgment some deference. A verbatim transcript reads more like a recording of an event, and it can be examined word by word, outside the context that made those words make sense in the room. A comment made in frustration, or a hypothetical raised and dropped—either one can be pulled from a transcript and read in a way the client never intended and the clinician would never have chosen to document.

A transcript also only captures what's said out loud. It can't record a silence that told the clinician something important, or a decision not to press on a topic that day. If that transcript later becomes evidence, it becomes the record, while everything the clinician actually noticed and responded to, but didn't say aloud, remains invisible to it. The written record and the clinical reality can end up telling two different stories, and only one of them is in the file.

When a clinician turns on an AI note taker, the tool listens to the session, converts the audio to a transcript, and then generates a note from that transcript. The clinician chose to use the software, but the transcript itself is an automatic byproduct of that choice, not a separate decision. In many cases, the clinician never even sees the full transcript. Their view is the finished note in the chart. But the transcript can still exist on the vendor's servers, and a subpoena or discovery request directed at the vendor can reach it independent of what's in the clinician's own records. A clinician can be entirely unaware that a full transcript of a session exists, and still be affected by what it contains.

It helps to be specific about what's actually at risk here, because Illinois mental health records already have real legal protection. A bare subpoena for a therapist's records doesn't work on its own; state law requires either a court order or the client's written consent before that protection can be pierced. The open question is whether a transcript sitting on a vendor's server gets treated the same way as a record in the therapist's own file. That protection was written around the therapist, not around a third-party AI company, and it hasn't been settled whether the vendor counts as close enough to the therapist for the same rule to apply, or as an outside party the protection doesn't reach.

Documentation differences become easier to question

Clinicians naturally document differently from client to client and session to session. That's normal, appropriate practice, reflecting real differences in what each client needs. A full transcript record, especially across many clients over time, creates something that didn't exist before: a searchable, comparable data set showing exactly how a clinician handled similar disclosures across different people.

If a clinician asked more follow-up questions with one client than another after a similar disclosure, or documented one client's risk assessment in more depth than another's, that difference used to live in clinical judgment and stayed there. With transcripts, the difference becomes visible and comparable in a way it wasn't before. In a licensing complaint or a lawsuit, an attorney may not need to prove intent, only point to the pattern and ask the clinician to explain it. There may be a sound clinical reason for treating two situations differently, but that reasoning now has to be reconstructed after the fact, in front of a judge or licensing board, rather than simply being part of how the clinician practices.

The added exposure for clients in vulnerable or targeted situations

A verbatim transcript can capture disclosures a client might never put in writing themselves: immigration status, substance use history, a name or pronouns used only in session, details of a custody dispute, or a health condition not disclosed elsewhere. A clinician's handwritten note would typically summarize what's clinically relevant and leave the rest out. A transcript retains all of it, indefinitely, on a vendor's server, in a form that can be breached, subpoenaed, or produced in response to a legal request the client never anticipated.

For clients from populations that face heightened legal or social risk, including transgender clients, immigrants, people with substance use histories, or anyone in a contested custody matter, there is a heightened concern. A subpoena, a data breach, or a records request tied to an employer-sponsored benefit plan can surface a transcript containing details the client shared in confidence, believing it would stay in the room. Clinicians serving these populations should weigh carefully whether a full transcript is worth the exposure it creates for the clients who have the most to lose if that record becomes public.

Isn't this the same as any other digitized medical record?

Not quite. A medical record hosted in an electronic health record (EHR), or sent through a secure clinical messaging platform, is still something the clinician wrote. The platform just stores or moves content the clinician already authored. That arrangement has decades of settled legal practice behind it—an EHR host generally isn't treated as an outside party whose involvement breaks confidentiality just by storing the file.

An AI note-taking tool works differently. It generates the transcript and the note itself, from raw session audio, and that content didn't exist in written form until the tool created it. Whether that active, generative role makes the vendor an extension of the therapist or an independent third party holding sensitive information hasn't been settled the same way an EHR host's role has. That difference, not the fact that the data happens to be digital, is where the added risk actually comes from.

The push toward storing everything, forever

A newer trend makes all of this more pressing. Some AI note-taking vendors are now marketing long-term memory features, where the system retains full session content indefinitely so the AI can reference something a client said months or years earlier, without the clinician having to re-enter it. The pitch is continuity: an assistant that “remembers” a client's history the way a longtime clinician would.

The appeal is real, but so is the tradeoff. Instead of a single session's transcript existing somewhere on a server, a client's entire multi-year treatment history exists there, in raw form. If that archive is ever reached by a subpoena, a data breach, or a records request, it isn't one session at risk. It's every session the client has ever had with that clinician, all at once, in the client's own words. For clinicians, that turns a single discovery request into exposure across an entire course of treatment rather than one appointment. For clients, it means a single point of failure can expose a history they never expected to exist outside the room at all. Before adopting a tool that advertises this kind of long-term memory, it's worth asking directly whether that feature can be turned off, and what happens to the archive if it is.

Steps that reduce risk

None of this means AI tools should be off the table. It means the record these tools create needs to be treated as a different kind of document, with different risks, than the notes clinicians have always written by hand.

A few steps worth taking regardless of which tool a practice uses:

• Start with retention. Ask the vendor directly whether raw audio and full transcripts are kept after the note is generated, and for how long, then push for the shortest window they'll commit to in writing. Don't assume “I never look at the transcript” means it doesn't exist somewhere on their servers.

• Ask specifically about long-term memory features. If a vendor advertises the ability to recall details from sessions months or years in the past, find out whether that feature can be disabled and what happens to the underlying archive if it is.

• The vendor agreement deserves as much scrutiny as the software itself. Look for an explicit promise that session data won't be used to train the vendor's models, a defined breach notification timeline, and a complete list of any subcontractors who touch the data.

• Get specific, written consent. Name the tool and explain what it does, rather than relying on a line in a general intake form. In Illinois, the Wellness and Oversight for Psychological Resources Act (Public Act 104-0054) spells this out directly. Section 15(b) provides that “no licensed professional shall be permitted to use artificial intelligence to assist in providing supplementary support in therapy or psychotherapy where the client's therapeutic session is recorded or transcribed unless” the patient is informed in writing “that artificial intelligence will be used” and “the specific purpose of the artificial intelligence tool or system that will be used,” and the patient “provides consent to the use of artificial intelligence.” The Act also defines what counts as valid consent, and what doesn't. Under Section 10, consent must be “a clear, explicit affirmative act” that is “freely given, informed, voluntary, specific, and unambiguous,” provided in writing, and “revocable by the individual.” The same section specifically excludes “the acceptance of a general or broad terms of use agreement,” a client merely “hovering over, muting, pausing, or closing a given piece of digital content,” or any agreement “obtained through the use of deceptive actions.” A checkbox on a standard intake form doesn't meet this bar.

• Match the record to the need. Consider which clients and which disclosures genuinely require a full transcript versus a clinical summary. The most protective note is often the one that captures what's clinically necessary and nothing more.

If any of this raises questions specific to your practice, that's a conversation worth having with your attorney and your malpractice carrier immediately. Don’t wait until a case is made against you to take precautionary measures regarding AI note taking.

---

This post is provided for informational and guidance purposes only and does not constitute legal advice. If you believe this guidance applies to your situation, we urge you to seek legal assistance from your professional liability insurance carrier or a licensed legal representative before taking any action.