AI Mistakes in Therapy Notes: Common Issues

Short answer: Common AI mistakes in therapy notes are: misattributing statements to the wrong speaker, inventing clinical findings not in the transcript, conflating this session with prior sessions, glossing over edge-case terminology, and generating vague assessments that could apply to any client. Clinician review during the verification pass catches all of these. Scriber's diarized transcripts and custom templates reduce the error rate in the first place.

An AI note generator is a language model that has learned patterns from thousands of therapy notes. It is not a human clinician. It does not have memory of prior sessions, does not understand nuance the way a clinician does, and does not have your clinical judgment. Mistakes happen. The goal is to understand which mistakes are common, spot them quickly during review, and fix them before signing.

Speaker attribution errors

The most dangerous AI mistake in therapy notes is misattributing a statement to the wrong speaker.

What happens: The transcript says "Client: I've been thinking about ending it all" and the AI note generator writes in the Assessment: "Clinician expressed thoughts of suicide." The speaker attribution in the diarized transcript was correct; the AI generator simply inverted it.

Why it happens: When a transcript has quick speaker transitions, overlapping speech, or multiple speakers in a group session, the language model sometimes inverts the attribution. It knows the content is clinically important (it's about suicide), but the speaker tags got scrambled in the processing.

How to prevent it: Scriber's diarization engine (especially Speechmatics) marks up the transcript with clear speaker labels. When the note generator reads "Client: [statement]," it is much less likely to invert it than it would be reading a unmarked transcript. Use Speechmatics for multi-speaker sessions and couples therapy, where diarization quality is highest.

How to catch it: During review, scan the Subjective and Plan sections for client statements. If the AI wrote "the client reports X," check the transcript to confirm the client actually said X (not the clinician). This takes thirty seconds per note.

Inventing clinical findings

A close second to speaker attribution is the AI inventing a clinical finding that was not in the session.

What happens: The clinician notes that the client's affect was "congruent" and appeared "engaged." The AI note generator writes: "Mental Status Exam: mood depressed, affect blunted, thought process tangential." None of those findings were mentioned in the session.

Why it happens: Language models trained on thousands of notes have learned associations: if a client reports stress, the note may mention depressed mood even if the client did not say that. If a client is in therapy for social anxiety, the note might include "patient appears withdrawn" even if the client appeared engaged. The model is pattern-matching, not reasoning.

How to prevent it: Custom templates help. When Scriber's prompt says "only include MSE findings that are explicitly documented in the session," the model is less likely to fabricate. But the prompt is a heuristic, not a guarantee.

How to catch it: During review, read the Mental Status Exam section and the Objective section line by line. Check each finding against the session notes or audio. If the AI wrote "thought process tangential," did the client actually ramble or use loose associations? If the AI wrote "affect labile," did you actually observe rapid mood changes? Delete any finding not supported by the session.

Conflating past and present

Another common error is mixing up content from prior sessions with the current session.

What happens: The client reported suicidal ideation in Session 4 two weeks ago. In Session 5, the client reports feeling better. The AI note generator writes: "Clinician addressed ongoing suicidal ideation. Client continues to report thoughts of self-harm." The session note conflates the prior statement with the current session.

Why it happens: The language model was trained on notes that often reference prior sessions ("patient reports improvement from prior suicidal ideation"). When the current transcript mentions suicidal ideation even in the context of "I don't feel that way anymore," the model pattern-matches to prior-session references and conflates them.

How to prevent it: Scriber's note generator includes the instruction "write only about content from this session." That reduces conflation, but it is not foolproof.

How to catch it: During review, check the tense. Do the statements use present-tense language ("the client reports X during this session") or past-tense language ("the client continues to report")? If the AI is using past-tense or "ongoing" language, confirm that it is clinically accurate to this session's content, not a reference to a prior session.

Glossing over edge-case terminology

When the transcript includes unfamiliar terminology — a condition the clinician mentioned once, an assessment tool the practice uses, a medication brand — the AI may gloss over it or misinterpret it.

What happens: The clinician says "The client scored 22 on the PHQ-9" (a depression screener). The AI note generator writes "Patient screened positive for depression" instead of including the specific score. Or: the clinician mentions "the client has aDHD," which the transcription engine hears as "the client has a D-H-D," and the note includes the misspelling.

Why it happens: The AI generator was trained on notes that often use screeners but may not include specific scores (clinicians often skip the number). Unfamiliar abbreviations or terms that the transcription engine misheard are harder for the generator to correct.

How to prevent it: Configure a custom medical-terms list in Scriber. If your practice uses specific screeners, medications, or diagnostic terminology, teach the transcription engine to recognize them. That helps both the transcription accuracy and the note generation.

How to catch it: During review, check abbreviations and scores. If you said "PHQ-9 score of 22," is it in the note as "22" or is the score missing? If you mentioned a specific medication brand, is it spelled correctly?

Vague or generic assessments

An AI note may have correct Subjective and Objective sections but generate an Assessment that could apply to any client.

What happens: The Subjective documents that the client reported a promotion at work, improved family relationships, and better sleep. The Assessment says: "Client continues to engage well in therapy. Prognosis good." That is not wrong, but it does not reflect the specific progress documented in the Subjective.

Why it happens: Generic assessments are common in training data. The language model learns that when a client is doing well, the assessment often includes "prognosis good" without being specific about why. The model produces something that matches the pattern, not something that ties back to the specific session content.

How to prevent it: Custom templates with explicit instructions help. Scriber's prompt can say "the Assessment must reference specific content from the Subjective and Objective" rather than using generic language. But the clinician's review pass is the main prevention.

How to catch it: During review, read the Assessment section. Does it reference the specific content from this session? Or does it use language that could apply to any session? If it is too generic, edit it to be specific.

Mishandling multi-speaker contexts

In couples therapy, family sessions, or group therapy, the AI may fail to attribute statements correctly to the right family member or group member.

What happens: The spouse says "I feel unheard," and the note attributes it to the identified client instead of the partner. Or: in group therapy, the AI note bundles statements from two different group members under a single person's name.

Why it happens: Multi-speaker diarization is hard. Even the best engines make mistakes when three or more people are talking. The AI generator is working from a diarized transcript that has speaker errors, so it propagates those errors into the note.

How to prevent it: Use Speechmatics for multi-speaker sessions. Its diarization is the strongest available. Or, have the clinician manually correct the diarized transcript before note generation — delete the lines with wrong speaker attribution.

How to catch it: During review, if the note is a couples or group note, check that each statement is attributed to the right person. Scan for "Client said X" and "Partner said Y" and confirm they are correct.

Fixable vs. unfixable errors

Most AI mistakes are fixable in the five-minute review pass. During review, you can:

• Fix speaker attribution errors (one sentence).
• Delete invented clinical findings (one line or section).
• Correct glossed-over terminology or missing scores (one or two words).
• Strengthen vague assessments (one or two sentences).
• Reattribute statements in multi-speaker notes (one sentence or a few lines).

The errors that are hardest to catch are the ones where the AI writes something that sounds clinically plausible but is not what you said — a slight exaggeration, a misread tone, a conflation of past and present. Those require you to actually think about the note, not just scan it for obvious errors.

Building a review routine

A three-to-five-minute review routine catches 95%+ of AI errors before you sign the note.

1. Read the Subjective section. Does it reflect what the client said? Are the speaker attributions correct?

2. Scan the Objective section. Are all the clinical findings actually things you observed? Nothing invented?

3. Check the Assessment section. Is it specific to this session, or generic language that could apply anywhere? Does it reference the specific Subjective and Objective content?

4. Verify the Plan. Are the interventions, homework, and next-session focus accurate?

5. Check abbreviations and scores. PHQ-9, PCL, medication names — are they spelled correctly and complete?

That is the routine. Do it on every note, even the ones that look correct. The review is not optional — it is the verification step that makes the AI scribe safe.

The goal of understanding common AI mistakes is not to scare clinicians away from AI scribes. It is to give clinicians the knowledge they need to review quickly and confidently. With a five-minute verification pass, all these errors disappear from the chart.