AI notes for psychologists: what they can and cannot do safely

AI-assisted notes can help turn material from a session into a more organised draft. That ability can be useful, but it does not make the system an independent source of clinical truth or move responsibility away from the professional.

The important question is not simply whether a tool can generate text. It is which parts of the work it can support, where its limits begin, and what conditions should be in place before a draft becomes part of the clinical record.

AI can support a draft, not decide the case

Within a carefully designed documentation workflow, AI may help to:

• Organise information from a session.
• Propose a clear structure for the record.
• Summarise themes without turning the document into a transcript.
• Adapt the draft to the language of a therapeutic orientation.
• Recover previously documented context within an authorised clinical environment.
• Surface material the clinician may want to review, such as unfinished themes or changes from earlier sessions.

These functions can make a draft easier to review. Their value depends on the quality of the source material, how the system is configured, and the review that follows.

A tool was not present in the session in the same way as the therapist. It does not know the full relational context, which silences mattered, or which formulation best fits the developing process.

What an AI-assisted note cannot know on its own

Text can sound convincing and still be wrong. Fluency does not guarantee accurate facts, attribution, or interpretation.

AI should not be treated as an authority that can:

• Confirm that a transcript or summary accurately reflects the meeting.
• Decide which material is clinically relevant.
• Turn an observation into a diagnosis.
• Establish a closed formulation of the case.
• Assess risk autonomously.
• Determine the meaning of an ambiguous statement.
• Decide what information must be retained as a legal matter.
• Replace consent, the therapeutic frame, or professional decisions about data processing.

It can propose. The clinician still needs to check, contextualise, and decide.

When an inference becomes a fact

One of the more consequential errors occurs when a draft states an interpretation with too much certainty.

From a conversation about a recent relationship, a system might write:

“The client has abandonment fears and emotional dependency.”

That statement may go beyond what took place. A more careful version might say:

“When discussing a recent relationship, the client describes intense concern about possible distance. The session explores how they respond when the other person appears less available.”

The second version preserves clinically relevant material without turning one possible reading into a conclusion. This is not merely a stylistic correction. It is part of clinical judgment.

Health data: the tool matters as much as the text

Therapy sessions and psychological records contain especially sensitive health data. Before clinical information enters a system, the clinician should understand what happens to it.

It is not enough for a tool to say that it uses encryption or is “GDPR compliant”. The professional needs to understand the actual flow: who processes the data, for what purpose, where it is handled, how long it is retained, and which controls apply.

Useful questions for a provider include:

• What data-processing role does each party hold?
• Is there an appropriate data-processing agreement?
• Where is the data processed and stored?
• Who can access the information, and how is access controlled?
• How long are sessions, transcripts, and drafts retained?
• Can the clinician request deletion?
• Is clinical data used to train models?
• How are security incidents handled and communicated?
• What happens to the information when the service relationship ends?

For more detail on these questions, see our guide to GDPR and clinical data for psychologists.

Minimise before automating

Using AI does not change a basic principle: only the information needed for the clinical and documentation purpose should be processed.

Before adopting a tool, it is worth asking:

• Does it need the complete audio, or can it work with less information?
• Does it require identifying details, or can references be separated?
• Does it retain copies that are no longer needed?
• Does the draft include intimate detail without a clinical purpose?
• Can intermediate material be removed once the final record is complete?

Automating excessive documentation does not make it safer. Clinically useful documentation still requires precision and restraint.

Human review is not a final formality

Reviewing a generated note is not only a matter of correcting words. The clinician should check:

Facts and attribution. What happened, who expressed each idea, and whether the sequence is correct.

Clinical relevance. What belongs in the record and what can remain outside it.

Observation and interpretation. Whether the text distinguishes what the client said, what was observed, and what is being considered as a hypothesis.

Risk and protection. Whether important information has been omitted, exaggerated, or removed from context.

Therapeutic orientation. Whether the draft reflects the clinician's language and way of thinking rather than imposing a generic template.

Continuity. Whether the record helps clarify what changed, what continues, and what may need to be revisited.

Until this review is complete, the text should be treated as a draft.

Uses that require particular caution

Some practices create avoidable risk:

• Pasting identifiable information into a consumer tool without suitable terms for clinical data.
• Accepting a draft as the final record without reading it.
• Asking the system to diagnose from one session.
• Using an automated score as the sole assessment of risk.
• Providing the whole clinical history when only part is needed.
• Retaining transcripts or audio without a defined purpose.
• Choosing a tool without understanding its retention, training, and access terms.

Our comparison of AI clinical note tools can help structure an evaluation, but the final decision needs to reflect the real context of the practice.

A practical checklist before using AI notes

Before bringing a tool into clinical work, it is worth checking:

• Do I know what information enters the system and why?
• Do I understand where it is processed and how long it is retained?
• Is it clear whether the data is used for model training?
• Can I control access and remove information when appropriate?
• Does the draft distinguish facts, attribution, and hypotheses?
• Is there human review before the record is saved?
• Does the tool support my therapeutic orientation or impose a generic format?
• Can I explain clearly to the client how their information is handled?

AI-assisted notes can support documentation when the workflow is designed to protect data and keep clinical judgment at the centre. They do not replace professional responsibility. They should make the draft more reviewable, not the decision more automatic.

Clara prepares clinical records for you to review through your own therapeutic orientation, keeping your judgment as the reference point for the final document.

Try Clara free →