AI adoption among psychologists doubled between 2024 and 2025. The APA's December 2025 Practitioner Pulse Survey found 56% of 1,742 psychologists had used AI at least once, up from 29% the year before (APA, December 2025). The concerns rose just as fast: 67% worried about data breaches, 64% about unanticipated social harms, 63% about biased outputs. That tension is not a malfunction. It's your professional judgment working correctly.
---
What the ethics codes actually say
The profession has been faster to adopt AI than to write rules for it, but the guidance exists and it's clearer than the vendor marketing suggests. In June 2025, the APA adopted its first formal ethical guidance on AI in health service psychology — the first ever (APA, June 2025). The core principle is direct: psychologists "remain responsible for final decisions and must not blindly rely on AI-generated recommendations." AI should augment, not replace, human decision-making. Human oversight remains essential.
That language tracks Ethics Code Standard 2.01, Boundaries of Competence. You are responsible for understanding what a tool does before you use it with clients. If a documentation assistant produces a note and you sign it without reading it, that's yours now. The tool didn't sign it.
The NASW Code of Ethics (2021) takes the same position from a different angle: all ethical standards apply equally whether a service is delivered in person or through technology (NASW, 2021). Standard 1.07 governs privacy and confidentiality in electronic communications; Standard 1.05 requires you to assess whether a client can actually use technology before you build it into their care. The Code doesn't create a special AI track. It says the standard is the standard, everywhere.
Worth naming directly: as of June 2026, NASW has not published AI-specific ethics guidance. A Code of Ethics Revision Workgroup is working on recommendations. That gap is real, and it's one reason clinicians feel like they're navigating this alone. They are, somewhat. But the underlying principles — competence, confidentiality, client welfare, human accountability — aren't waiting.
The ONC HTI-1 Final Rule (January 2024) added a federal layer: certified health IT systems now have to disclose how their AI and predictive tools are trained, and the rule assumes clinicians will evaluate those outputs (HHS/ONC, January 2024). That's not a bureaucratic footnote. It means the federal government is operating on the assumption that a human professional is reviewing what the AI generates. If you're not, you're not meeting the assumed baseline.
---
What you can automate
The short version: automate anything that currently eats time without requiring clinical judgment. The longer version involves understanding where your professional liability actually lives.
Scheduling and reminders. Automated appointment booking and reminder systems don't make clinical decisions. They handle logistics. A reminder that goes out incorrectly is annoying; it doesn't harm a client the way a missed safety assessment does. This is the clearest case for automation in private practice. The time you recover here goes back into sessions.
Documentation drafts. AI-generated session notes are now the most common clinical application, and also the most misunderstood one. The output of an AI documentation tool is a draft, not a note. You read it, correct it, and sign it. If the draft captures something wrong — a symptom it hallucinated from a prior session, a word the client never said — and you sign it, that's your clinical record. The APA guidance is explicit that you remain responsible for final decisions. Documentation drafts are automatable; documentation sign-off is not.
Administrative tasks. Eligibility verification, billing code lookups, intake form routing, release-of-information tracking — these are paperwork tasks that happen around clinical work, not inside it. Most carry low clinical risk if automated imperfectly. They're also where a huge fraction of administrative burden accumulates. The NASW 2017 Technology Standards identified managing client information as one of four defined domains for technology use in practice (NASW, 2017). Administrative information management belongs in that domain.
Psychoeducation delivery. Sending a client a resource — a worksheet, a reading, a mood-tracking prompt — between sessions is not a clinical act. Selecting what to send them is. The delivery mechanism is automatable; the clinical judgment about what's appropriate for this client right now is yours.
[INTERNAL-LINK: therapist-documentation-burnout → supporting article on documentation time burden and burnout]
---
What you keep human
This is the list that doesn't have a workaround, and shouldn't.
Clinical judgment. Diagnosis, treatment planning, and formulation require you to synthesize information in ways that are context-dependent, relationship-dependent, and ethically grounded in your knowledge of a specific person. An AI tool can surface patterns across a dataset. It cannot know what your client didn't say in a session last month and why that matters today. The APA guidance is unambiguous on this: AI should not replace human decision-making in clinical contexts.
The therapeutic relationship. The alliance between clinician and client is itself a treatment mechanism. Research on what makes therapy effective consistently points here, not to the specific modality. No documentation assistant, chatbot, or scheduling tool is a substitute for it, and nothing in the ethics codes suggests otherwise. This isn't a romantic idea about therapy — it's a clinical claim with a substantial evidence base.
Crisis assessment. When a client presents with suicidal ideation, a plan, access to means, or any acute safety concern, that assessment is yours. Fully, immediately, and without delegation. An AI tool that flags risk based on a questionnaire score can be a useful data point. It is not an assessment. The decision about how to respond to a client in crisis sits with the licensed clinician, period.
Informed consent for AI use. If you're using an AI tool with any client data, they need to know. This falls under Standard 1.07 of the NASW Code and is reinforced by the APA guidance on transparency. Automating the consent process itself is possible; the decision about whether to use AI in a client's care is not something you outsource.
[INTERNAL-LINK: what-therapists-want-in-an-emr → article on evaluating clinical tools against ethics standards]
---
The question isn't if — it's which
A 2025-2026 survey of 1,179 social workers found that roughly two-thirds were already using AI (UT Austin Moritz Center/NASW, 2025-2026). The "should I use AI" conversation has mostly already happened inside the profession, whether the profession meant to have it or not. The question clinicians are actually asking is which uses are defensible, and how to tell the difference.
Here's a practical framework that holds up against the guidance above.
Ask: does clinical quality depend on a human doing this? If yes, keep it human. If no, it's a candidate for automation — with clinician review at the output before anything reaches a client or record.
Ask: whose license is on the line if this goes wrong? The answer is always yours. Which means you need to understand the tool, know what it's doing with client data, and be able to explain your use of it. The ONC HTI-1 rule assumes you'll evaluate AI outputs. The APA says you must not blindly rely on them. That's a consistent standard across both regulatory and professional ethics frameworks.
Ask: was this tool built for clinical practice, or adapted for it? The general-purpose AI tools that have filtered into clinical workflows were not designed with HIPAA-eligible data handling, licensing board ethics, or therapeutic context in mind. The NASW survey found 64.2% of social workers want guidance on AI's impact on bias and safety, and 62.3% want ethical guidelines (UT Austin Moritz Center/NASW, 2025-2026). Part of what they're asking for is tools built to the standard, not retrofitted to it.
That distinction matters. A tool built by someone who has done the paperwork, understands the ethics frameworks, and designed the system around clinician accountability is a different product than a general AI assistant someone is using for session notes because it's available. The difference shows up in how data is handled, where human review is required, and whether the product was designed to keep you in the loop or to make it easy to step out of it.
[INTERNAL-LINK: prior-auth-reform-2026-behavioral-health → prior article on systemic burden; shows the same pattern of systems not built for clinicians]
VibeCheck was built by a clinician doing this work. The design starts from the assumption that the clinician is in the loop, because the ethics frameworks — APA, NASW, ONC — require exactly that.
---
FAQ
Is it ethical to use AI for therapy session notes?
Yes, with a critical condition: the AI produces a draft and the clinician reviews, corrects, and signs it. The APA's June 2025 guidance is clear that psychologists remain responsible for final decisions and must not blindly rely on AI-generated content. An AI-written note you sign without reading is your clinical record, not the tool's.
Do I need to tell my clients I'm using AI?
Yes. NASW's Code of Ethics Standard 1.07 governs confidentiality in electronic communications, and the APA guidance emphasizes transparency. If any client data is being processed by an AI tool, informed consent covers it. What that disclosure looks like will depend on your state licensing board and the specific tool — but the obligation to disclose is not optional.
Does using AI in my practice violate HIPAA?
Using AI tools with client data requires a Business Associate Agreement (BAA) with the vendor. A tool without a BAA is not appropriate for protected health information, regardless of how the vendor markets its security. The ONC HTI-1 rule (January 2024) also requires certified health IT systems to disclose how AI tools are trained. "HIPAA-eligible" means the tool can sign a BAA and has the infrastructure to support compliant use — that's the bar to check before adopting any AI tool in practice.
What should I do before adopting an AI tool in my private practice?
Ask three things: Does the vendor offer a BAA? Can they explain how the tool was trained and what it does with my client data? Was it designed for clinical practice or adapted from a general-purpose AI? Then check your state licensing board for any specific guidance on AI in practice. The NASW survey found 54.3% of social workers want privacy protections built into the tools they use — meaning you're not unusual for asking these questions before signing up.
---