AI for Psychologists in 2026: Evaluation Reports, Scores and Your Signature

AI can draft the narrative of a psychological evaluation, turn a stack of outside records into a chronology, write progress notes and reword findings for parents. It can't choose a battery, weigh conflicting scores, or carry your responsibility for the interpretation. Every number in a draft still has to be checked against the protocol.

The report is where most of the opportunity sits for a psychologist, and also most of the risk. A letter to a pediatrician with a clumsy sentence costs you little. A report with a wrong percentile can follow a child through school placement or a client into a courtroom. Below, the work is split by task, with the tools built for it and the rules that decide what you can paste where.

Where a psychologist's hours go

In psychological assessment, the written product is often the service. O*NET's summary of clinical and counseling psychologists lists session notes and treatment plans, client reports, written competence evaluations for court, reviews of medical, social and family histories, and supervising interns and staff.

Put that against a real week and the backlog tends to look like this:

• The evaluation report. Background from the intake, a records review, score tables, integration across measures, diagnostic reasoning, recommendations. In one 2025 preprint, summarized below, human authors averaged 2.5 hours on a report built from mock data alone.
• The feedback session and its plain-language version. The same findings, rewritten for a parent, a teacher or the client.
• Outside records. School files, earlier evaluations and medical charts that arrive as a thick, unordered packet.
• Payer paperwork. The referral question and the case for testing. Medicare's billing article for testing says the record must show why the tests were performed.
• Weekly therapy documentation, if you also see clients for treatment.
• Everything around the clinic. Referral sources, attorneys asking for dates, supervisees, continuing education, board and insurer mail.

Use is rising quickly. In the American Psychological Association's 2025 Practitioner Pulse Survey, 56% of responding psychologists had used AI at work in the past year, up from 29% in 2024. Among those users, 52% drafted emails and other materials, 32% summarized clinical notes or articles, and 22% used it for note-taking, while only 8% used it to help with diagnosis. The survey had 1,742 respondents.

Which AI tools handle which part of the work

Dedicated assessment-report tools now exist alongside therapy-note tools, and they solve different problems. Prices and terms below come from each vendor's page as published. They are descriptions of what the vendors offer, not test results or rankings.

Access is the catch with the publisher tools. Pearson's assistant is only available through its Digital Assessment Library, and a March 2026 webinar listing on that page still called it a pilot. Pearson says it supports non-Pearson tests too, and its stated goal of cutting report time by more than 50% is described as an objective, not a measured result.

How good are AI-written reports?

One direct comparison came out mixed, and it tested something you shouldn't do: AI reports with no clinician editing. In a January 2025 preprint by Lockwood and colleagues, as summarized by NovoPsych, 249 licensed psychologists blindly rated human and GPT-4 reports written from mock child cases. Humans scored better on summaries, AI scored better on recommendations, and readability was level. AI took about 91.5 seconds per report against 2.5 hours.

The figure worth remembering is the signature. Only 49.6% of raters were willing to sign the human report and 43.1% the AI one. No clinician edited the AI output, the model is now old, and NovoPsych sells its own AI scribe, so read the summary with that in mind. What it supports is a fast first draft that still needs you. It says nothing about a finished report you've rewritten.

Invented numbers are the failure to plan for. BastionGPT's own guide, written by a psychologist who uses the product, gives the example of a draft claiming a 99th percentile score that wasn't in the data.

What AI can't do, and what the rules say

The APA Ethics Code already answers most of the AI question for psychologists, because it was written with automated scoring and interpretation in mind. Standard 9.09(c) keeps responsibility for the application, interpretation and use of instruments with you, whether you score a test yourself or rely on an automated service. Treat an AI narrative with the same care.

Test security: keep items and manuals out

Standard 9.11 asks you to protect test materials, meaning manuals, instruments, protocols and test questions or stimuli, consistent with law and your contracts. Test data, such as scores and client responses, is treated differently, and the HIPAA section below covers it. In practice:

• Don't paste item content, stimuli or manual text into a general AI tool. A signed privacy agreement covers client data and does nothing for the publisher's copyright or your license terms.
• Read the publisher's terms first. Pearson's product pages reserve rights over text and data mining and AI training. Ask the publisher before sending its score reports to an outside tool.
• Know which side a score report falls on. Interpretive reports often mix your client's results with the publisher's proprietary wording, which is why PAR built a step to strip that text.

HIPAA: scores aren't psychotherapy notes

HIPAA's extra protection for psychotherapy notes is narrower than many people assume. The definition covers notes about session conversation kept separate from the record, and it excludes results of clinical tests along with summaries of diagnosis, treatment plan, symptoms and progress. An identifiable score table is ordinary protected health information, so the tool handling it needs a business associate arrangement. No signed agreement means no identifiable client data, and that rules out a consumer chatbot account. BastionGPT says a signed agreement comes with every plan. Upheal's pricing page calls it HIPAA compliant without mentioning one, so ask.

Consent, and the states getting specific

The APA's guidance on AI in practice says psychologists should tell clients the purpose, application, benefits and risks of the AI tools involved, and critically evaluate what those tools produce. Some states have gone further. Illinois's 2025 law allows AI for administrative and supplementary support, but bars it from making independent therapeutic decisions and requires specific informed consent before AI is used on recorded or transcribed therapy sessions. Penalties run up to $10,000 per violation. Check your own board.

Forensic work, clinical judgment and billing

• Court evaluations. Standard 9.01 ties your opinions, including forensic testimony, to information and techniques that support them. Before AI touches a forensic report, make sure you could defend each sentence under cross-examination as your own reasoning.
• Client-facing chatbots. A 2026 systematic review in Clinical Psychology & Psychotherapy found preliminary and inconsistent evidence, with many studies lacking control groups, and concluded that human oversight has to stay central.
• Time-based billing. Medicare's article says a per-hour code needs at least 31 minutes. If a draft shortens your report work, bill the time you actually spent, not the time reports used to take.

Merlin's place in a psychology practice

Merlin belongs on the non-clinical side of your practice, and it doesn't belong anywhere near a chart. This Mac assistant uses Gmail and Google Calendar to keep track of people and what you've promised them. It has no assessment tools, no EHR and no clinical documentation features, and it doesn't answer phones. Merlin's privacy policy makes no HIPAA commitment and doesn't mention a business associate agreement. Transcribing a meeting also uploads the audio to a cloud service. So keep client names, sessions, scores and reports out of it.

That leaves the relationship work a referral-based practice depends on, as long as the email and calls carry no clinical content:

• Referral relationships. Which pediatrician or school contact you owe a reply or has gone quiet, and a draft reply you approve before it goes out. This fits mail with no client in it: check-ins, introductions and scheduling with the referrer.
• Non-client calls. Supervision logistics, vendor demos and practice meetings, recorded on your Mac with speaker-labeled transcripts on paid plans, and a brief before the next one from Prep me in Meetings.
• Commitments. A to-do list of what you agreed to send people, plus routines (jobs that repeat on a timer), such as a weekly list of referrers you owe a reply.

Three precautions matter here. Connect a Google account that doesn't hold client information, since Merlin reads the mail and calendar you connect. A referral that names a child or client is client information, so have those sent to your secure system and not to this account. If you run telehealth on your Mac, turn off the call and calendar recording prompts under Settings → Preferences → Meetings. And remember that routines run unattended, so don't give one permission to send anything without your review. To weigh it against other recorders for those calls, see seven AI note takers compared.

Your first five steps with AI in an assessment practice

1. Write down your tools and what each one touches. Note any client data going into a consumer account and stop that first.
2. Get the paperwork. From any tool you'd use, ask for a business associate contract to sign and a written retention and training policy, then read your test publishers' terms on AI.
3. Update your consent form. Name the tool, what it captures and how a client can decline, then check your state board's position.
4. Trial one section on a closed case. De-identify an old evaluation and have AI draft the background history only. Compare it with what you wrote, line by line.
5. Add a sign-off step. Before any report leaves, check scores against protocols, and keep the raw data as the source of truth. To send the finished report securely, see how to encrypt email on a Mac.