- clinical AI
- clinical documentation
The note that sounds right and says nothing: reviewing an AI draft in 90 seconds
An AI-drafted note rarely fails with an absurd sentence you would catch at a glance. It fails as a polished note that could belong to any client, that leaves out the one thing that mattered, or that turns your hunch into a fact. Wherever you practise, the note you sign is yours, so the review cannot be a read-through. It is four questions, asked in order. Is there an observation that could only have come from this session? Is your hypothesis marked as a hypothesis? Was anything asserted that the client did not say? Is the plan specific enough to act on next week? Before you open the draft, take ten seconds to say what mattered in the session. If the draft passes all four, you sign in ninety seconds. If it fails any, the review stops being a ninety-second job, which is the point. The evidence below is international; the worked example of how a records rule separates form from content is Mexico’s, because its standard puts the two in different clauses. Check the documentation rules where you practise.
What circulates: notes that are “comprehensive and reliable”
NovoPsych, which sells an AI note assistant, puts the pitch plainly on its page about AI in psychology (accessed 8 October 2026): “Manual note-taking often results in incomplete or inaccurate records due to human error, such as missed information or misinterpretation.” AI tools, it goes on, “can also identify and correct common errors, ensuring that notes are comprehensive and reliable.” The Spanish version carries the same promise into the market this blog writes for: «lo que garantiza que las notas sean exhaustivas y fiables» (“which guarantees that notes are exhaustive and reliable”). If notes come guaranteed, review is left as a final formality on a text that arrives correct.
No study we could locate supports “ensuring”. The closest to your work is TN-Eval (ACL 2025). Nine licensed therapists rated, blind to authorship, SOAP notes for 50 English-language conversations transcribed from motivational-interviewing demonstration videos; some notes were written by therapists, others by two language models. On a 1-to-5 impression scale, the therapists rated the model notes as more faithful to the conversation (4.68 and 4.90) than the human notes (4.43). When they checked sentence by sentence against the transcript, the order reversed: a faithfulness score of 87.0 for therapist notes, 68.5 and 71.8 for the models. The authors’ explanation: the models’ fluent writing “likely biases human annotators, thus conflating fluency with accuracy.”
The same study requires a caveat. The model notes covered more rubric items than the therapists’ notes, which were shorter. The problem is not that AI forgets everything; it is that what it writes reads as true whether or not it is, and a review by impression rewards exactly that.
Why reading it through fails
Fluency. That is the TN-Eval finding, and it is not isolated: in the commercial scribe study covered in what AI can and cannot do in psychotherapy, blinded reviewers preferred the AI notes even though they carried more hallucinations.
Anchoring. In a simulation study in The Lancet Digital Health (Chen and colleagues, 2024), six radiation oncologists answered simulated patient messages, first on their own and then by editing GPT-4 drafts. Their edited answers resembled the drafts more than what they had written unaided. The authors conclude that physicians “might not simply use LLMs to better phrase their own assessment, but instead adopt the assessment by the LLM.”
Time pressure. The systematic review by Goddard, Roudsari and Wyatt (2012, 74 included papers, focused on healthcare and all predating language models) named workload, task complexity and time constraint as mediators of automation bias, the tendency to over-rely on the machine; among the mitigators, emphasising the user’s accountability.
Hence the honest scope of the title. Ninety seconds is not a measured figure: no study we found measures how long a good review takes. It is a budget, and a dangerous one if spent reading, because haste feeds the bias. It works only if every second goes to a question the text cannot answer for itself.
Before you open it: ten seconds without the draft
Say, or jot in one line, what mattered in the session: the new fact, the moment something shifted, what was left pending. If you keep a session record, that is your input; if not, your same-day memory. Do it before reading, not after. No study we found shows this step is enough against anchoring: in Chen’s, the physicians had already answered each message on their own and, editing the draft, still ended up closer to it than to their own answer. So the line does not stay in your head: it is the yardstick you hold the text against, point by point.
It is also what protects you from the most frequent error, which is not the visible one. In the study by Asgari and colleagues (npj Digital Medicine, 2025), clinicians annotated 12,999 sentences in 450 notes generated with GPT-4 from simulated primary-care consultations. There were 191 hallucinated sentences and 1,712 consultation sentences the notes left out. Counting only major errors, those that could affect diagnosis or management if left uncorrected, there were 286 omissions against 84 hallucinations. In a proof-of-concept study from Charité in Berlin, with six simulated psychiatric interviews, the AI reports averaged 33.5 false negatives (findings present in the interview and absent from the report) against 6.3 for human reports; false positives (13.2 against 7.0) did not differ significantly with only six interviews. An omission leaves no trace on the page. You find it only if you knew what to look for before you started reading.
Three reading caveats. None of these numbers comes from real psychotherapy sessions. Asgari’s authors work for Tortus AI, and the maker of the software tested at Charité funded that study. The samples are small or simulated. They tell you where to look, not how many errors your draft contains.
1. Is there an observation that could only have come from this session?
The substitution test: cover the name and ask whether the note would be true of another client you saw this week. A typical draft (fictional):
“S: Reports feeling better overall, although some concerns about the work environment persist. O: Punctual, well groomed, oriented ×3; euthymic affect, coherent speech. A: Shows progress in managing anxiety and greater awareness of thought patterns. Denies suicidal ideation. P: Continue the therapeutic process, reinforce coping strategies and follow up on short- and long-term goals.”
No sentence has a typo and none survives substitution. It is the AI version of the cloned note: it does not copy last week’s note, but it produces anyone’s. The Charité team described the mechanism: the model condensed details into broad statements (“Often feels lonely”) and dropped time qualifiers, plausible compressions that left “superficially adequate summaries”.
Mexico’s clinical-record standard, NOM-004-SSA3-2012, makes the distinction explicit: it puts form and content in separate clauses. Clause 5.11 requires notes written «en lenguaje técnico-médico, sin abreviaturas, con letra legible, sin enmendaduras ni tachaduras» (“in technical-medical language, without abbreviations, legibly, without corrections or crossings-out”); a model meets that effortlessly. Clause 6.2.1 requires the progress note to describe the «Evolución y actualización del cuadro clínico» (“course and update of the clinical picture”), and clause 5.16 requires that, in independent outpatient psychology, progress notes be adjusted «a la naturaleza de los servicios prestados» (“to the nature of the services provided”). A fluent note satisfies the form of 5.11 and leaves 6.2.1 empty. Whatever your jurisdiction’s rule, the distinction is worth applying.
What passes is one fact that could not sit in another note: a score with its comparison (“GAD-7: 11; two weeks ago, 15”), a verbatim phrase in quotation marks, a behaviour with its moment. If the draft has none, check your record. If it was there, the model dropped it and you restore it. If it was not, the problem started before the draft.
2. Is your hypothesis marked as a hypothesis?
In the Assessment, a draft can state in the indicative what you only suspected: “the anxiety stems from the conflict with her boss.” Asgari and colleagues classed 10% of hallucinations as causal, links the consultation never established, and found major hallucinations most often in the Plan and next in the assessment section. Add anchoring and the risk is concrete: signing as yours a formulation you never made.
The review has two moves. First, every interpretive sentence must be attributable to you and carry its marker: “working hypothesis”, “appears”, “to explore”. What the client reports carries its own: “reports”. That is the S-versus-A boundary from the SOAP notes guide. Second, if the draft’s hypothesis is not yours, delete it. Do not soften it with “possibly”: a hypothesis you never held, hedged, is still not yours.
3. Was anything asserted that the client did not say?
The most dangerous sentences are negative ones. In Asgari’s study, 30% of hallucinations were negations, and the authors call them the most concerning: they appeared mostly in the plan and “contradicted what was said during the consultation.” In a therapy note the canonical case is “denies suicidal ideation”. If the client spoke of wanting to die, that sentence contradicts the session, which is exactly the negation Asgari measured; if you did not ask, it documents an assessment that never happened. Either way it is the first one anyone would read if something went wrong. Delete it. And if the session carried a risk signal, that note is not a ninety-second review: write it with the care of documenting a suicide risk assessment.
Next, observations nobody made. A model working from audio or from your text has no eyes. In TN-Eval, the Objective section of the model notes scored 49.0 and 60.4 for faithfulness, against 85.1 for therapists. “Well groomed, oriented ×3, euthymic affect”: if your record does not say it, it is invented, however plausible it sounds.
Last, attribution. What the client says about other people remains their account: “reports that her partner belittles her”, not “her partner belittles her”. If the draft dropped the reporting verb, put it back.
4. Is the plan specific enough to act on next week?
The Plan is where models invent most. In Asgari, major hallucinations appeared most often there, and fabrications, the most common type, clustered there. In TN-Eval, Plan faithfulness was 46.6 and 43.8 for the two models against 78.2 for therapists, and plans across all the models the study tested averaged 78.5 words, against 29 for therapists: they carried short- and long-term goals where therapists typically wrote one sentence about the next appointment. If the session did not end with an explicit plan, the template field does not stay empty: the draft may supply a plausible one.
The test: what will you do differently next session, and what did you ask the client to do? “Continue the therapeutic process and reinforce coping strategies” answers neither. Mexico’s clause 6.2.6 asks for «Tratamiento e indicaciones médicas» (“treatment and medical instructions”), adjusted by 5.16 to a psychology service, and a plan that passes looks like this: “two pending calls before Friday; return to the conversation with her boss next session; repeat the GAD-7 in two weeks.” If it does not match what you agreed, what you agreed wins. And if note after note fails here, the gap is in the treatment plan that should be holding them up.
When the review does not fit in ninety seconds
Failing any of the four questions changes the job: you are no longer reviewing, you are rewriting that section from your record rather than patching the sentence.
Some sessions never enter the budget: those that touched risk, the first session, a crisis, and any note that will leave your practice for a court, a school or an insurer. You write those.
If the same question fails note after note, the problem is the input or the tool. The APA’s 2025 ethical guidance on AI, which states that it does not represent official APA policy, asks psychologists to critically evaluate AI-generated content “both at the start of use and in ongoing applications”.
The signature closes the loop. Mexico’s clause 5.10 requires every note to carry the «nombre completo de quien la elabora, así como la firma autógrafa, electrónica o digital» (“full name of whoever prepares it, and a handwritten, electronic or digital signature”). The APA guidance says the same in ethical terms: “Psychologists remain responsible for final decisions and must not blindly rely on AI-generated recommendations.” In gesell.ai a draft does not sign itself: signing requires your professional licence number, locks the note, and any later correction is added as an amendment.
The card: four questions and what each one catches
| Question | What it catches | Where it was measured | If it fails |
|---|---|---|---|
| 0. What mattered? (before reading) | Omissions | Asgari: 286 major omissions, 84 major hallucinations | If the draft lacks it, you write it |
| 1. An observation from this session? | The generic note | Charité: “superficially adequate summaries” | Restore the fact from your record |
| 2. Hypothesis marked? | Someone else’s conclusion | Chen: “adopt the assessment by the LLM” | Mark yours, delete what is not yours |
| 3. Anything the client did not say? | Negations, observations without eyes | Asgari: 30% negations; TN-Eval, Objective: 49.0 and 60.4 | Delete it; if there was risk, you write it |
| 4. Does the plan work next week? | The invented plan | TN-Eval, Plan: 46.6 and 43.8 against 78.2 | Replace it with what you agreed |
References
- Shah, Xu, Liu, Burnsky, Bertagnolli and Shivade — TN-Eval: Rubric and Evaluation Protocols for Measuring the Quality of Behavioral Therapy Notes (ACL 2025, Industry Track)
- Asgari et al. — A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation (npj Digital Medicine, 2025)
- Gülegen, Haaf, Schlüßler and Köhler — AI-generated documentation of psychiatric interviews: a proof-of-concept study (Frontiers in Psychiatry, 2026)
- Chen et al. — The effect of using a large language model to respond to patient messages (The Lancet Digital Health, 2024)
- Goddard, Roudsari and Wyatt — Automation bias: a systematic review of frequency, effect mediators, and mitigators (JAMIA, 2012)
- American Psychological Association — Ethical Guidance for AI in the Professional Practice of Health Service Psychology (2025)
- Diario Oficial de la Federación — NOM-004-SSA3-2012, Del expediente clínico (Mexico)
Ready to spend less time on documentation?
gesell.ai turns your session notes into professional clinical documentation — SOAP, DAP, BIRP, or GIRP — and generates treatment plans with SMART goals.
Try gesell.aiHave a question about your practice?
How to document a case, which format suits you, or how Gesell fits your practice: write to us and we will answer.
About the author
Gesell Team
Clinical and product content written by the gesell.ai team together with certified clinical psychologists.