AI for Testing Psychologists: Report Drafts Without Pasting Raw Scores

A testing psychologist can use AI to draft the prose of an assessment report, but only from a short de-identified summary that you write yourself. Never paste raw scores, protocols, item content, test record forms, names or other identifiers into a chatbot. The clinical conclusions, recommendations and signature stay entirely with you.

The efficient version of this is not "feed the data in and get a report back". It is "write a few lines of your own findings, ask for polished paragraphs around them, then rewrite whatever does not match the evidence". That keeps the sensitive material on your own systems and gives you most of the time saving, because the slow part of report writing is usually phrasing, not thinking.

Follow me on Instagram@sagnikteaches

The four things that must never be pasted

Treat this as a hard list, taped near the monitor.

Connect on LinkedInSagnik Bhattacharya

1. Raw scores and score tables. Standard scores, percentile ranks, subtest scaled scores and index profiles can count as test data under professional rules, and test publishers commonly restrict how they are shared. Even if a number looks harmless, a profile of scores can identify a person and reveal a test's structure.
2. Protocols, record forms and item content. Test items, stimulus material, manuals and scoring criteria are the publisher's copyrighted, security-sensitive material, and a record form with the client's answers written on it is test data as well. Their security matters because a leaked item can invalidate the test for future clients.
3. Identifiers. Name, date of birth, address, school, employer, referring professional, case number, and photographs. Also the combinations that identify in practice: a rare diagnosis plus a small town, a distinctive family history, an unusual occupation.
4. Anything verbatim from the clinical interview. Quotes, collateral reports and records belong to the file, and they often contain third parties who never consented to anything.

Professional codes address several of these points. The APA ethics code has Standard 4.01 on maintaining confidentiality; Standard 9.04 on releasing test data, which defines test data as raw and scaled scores, client responses to test questions or stimuli, and the psychologist's notes and recordings of what the client said and did during the examination; and Standard 9.11 on maintaining test security, which covers manuals, instruments, protocols and test questions and asks psychologists to protect them consistent with law and contractual obligations. Read the current text and your own licensing board's guidance for the exact wording, because rules differ by place and change. The APA also publishes "Ethical guidance for AI in the professional practice of health service psychology" (created June 2025, last updated December 2025), which covers transparency and informed consent, data privacy and security, accuracy and human oversight; check it for anything newer than this page. Your test publishers' licence terms are a separate source: look in the manual or on the publisher's website for what they say about copying, sharing and software.

Subscribe on YouTube@sagnikteaches

What you can give the AI instead: a de-identified summary

The substitute for raw data is a summary written in your own words, at the level of conclusions rather than numbers. Aim for something a colleague could read without being able to name the client or reconstruct the test. A good template, filled in by hand each time:

Case code: [use a code, not initials; key stored elsewhere]
Age band: [e.g. late adolescent / mid-thirties]
Referral question: [one line, general]
Domains assessed: [e.g. general cognitive, attention, reading]
Findings, in words only (no numbers):
- Verbal reasoning: [e.g. broadly average, relative strength]
- Working memory: [e.g. low average, variable]
- Attention: [e.g. clinically significant difficulty on the
  self-report, consistent across two raters]
- Reading accuracy and fluency: [e.g. below expectation]
Behavioural observations: [two or three generic lines]
Clinician's formulation: [your own two sentences]
Recommendations I have already decided on: [list]

Notice what is absent: no numbers, no names, no setting, no quotes. The descriptors are yours, chosen from the full data and your judgement. That is the safeguard: the model never had the data, so it cannot repeat it.

A worked report-section example

This is illustrative and uses invented descriptors, not data from any real case. The clinician pastes the summary above, with the findings filled in, and the prompt below.

You are helping a psychologist draft one section of an assessment
report. Use only the summary I give you. Do not add findings,
numbers, causes or diagnoses. Do not name any test. Write in the
third person, past tense, plain professional English, 150 to 200
words, for a reader who is not a clinician. Where my summary is
silent, write "[clinician to add]". Avoid certainty beyond what the
summary says; use "was consistent with" and "suggests", not "proves".

Section: Cognitive profile.
Summary: [paste de-identified summary]

A sample output (illustrative) might read: "[Client] showed reasoning in the average range on verbal tasks, which was a relative strength. Working memory was lower than other areas and varied across tasks, suggesting that holding several instructions in mind at once may be harder than understanding each one. Observations during testing were consistent with this pattern. [Clinician to add: effect on daily functioning.]"

What a careful clinician then does. First, replace "[Client]" with the real name on their own computer, after the AI step. Second, check every claim against the full data: "average range" and "lower than other areas" should match the profile, and "varied across tasks" should be true. Third, delete any sentence that implies a cause. The draft says "suggesting that holding several instructions in mind may be harder": if the clinician's formulation does not support that, it goes. Fourth, add the effect on daily functioning from the interview, which the AI correctly left blank. The final text is the clinician's. The model supplied sentence shapes and saved ten minutes of typing.

Where AI is useful and where it is not

A second worked section: attention findings from several sources

Results narratives get harder when sources disagree, and this is where a clinician's judgement shows. Say a young person was assessed for attention difficulties. In the clinician's own summary, the parent ratings point to clinically significant difficulty, the teacher ratings to mild difficulty, and the self-report to moderate difficulty. Observations during testing were mixed, with good focus on one-to-one tasks early in the session and drift later.

The clinician writes that in four short lines in the de-identified summary, adds the sentence of formulation they have already settled on ("the pattern suggests difficulties that depend on context and demand"), and asks for 150 words that present the agreement and disagreement between raters fairly, without resolving it. The draft that comes back usually does one of two things wrong. It either smooths the disagreement into a single confident statement ("attention was impaired across settings"), or it invents an explanation ("the teacher may have underestimated the difficulty because of classroom structure"). Both go. The clinician restores the disagreement in plain words and keeps the formulation line as the only interpretation.

The useful habit is to prompt for the shape you want: "present each source separately in one sentence, then one sentence on where they agree, and one on where they differ. Do not explain the difference." That instruction turns the model into a formatter, which is what it is safest as. It also makes checking easy, because each sentence maps to one line of your summary.

Building a phrase bank from your own past reports

The best way to get a draft that sounds like you is not a better prompt but a better sample. Over a few weeks, collect ten to fifteen paragraphs from your own signed reports that you like, from different cases, and strip them of every identifier and number by hand. Keep them in a document that is not tied to any case. Paste three of them into the prompt as examples of your style: "Write in the voice and structure of these examples. Do not copy any facts from them."

Two cautions. Examples must be genuinely de-identified: a paragraph about a rare presentation can identify a client even when the name is gone, so choose common ones or rewrite until the case is unrecognisable. And never reuse examples from an employer's or a colleague's reports without permission. A well-chosen sample set cuts editing time more than any other single change, because the model matches your sentence length, hedging habits and the order in which you present information.

A checking routine for every draft

An AI draft fails in predictable ways, so a fixed routine beats a general read-through. Use these five passes, in order, on every section:

1. Facts against file. Every claim about performance, history or behaviour matched to the full record. Mark each one that you cannot find.
2. Numbers and test names. There should be none. If the draft contains a figure, find out where it came from and delete it.
3. Causes and labels. Search for "because", "due to", "caused by", "typical of", "consistent with" and any diagnostic term. Keep only the ones your formulation supports.
4. Certainty. Look for "clearly", "definitely", "always", "never" and "proves". Replace them with the hedging your evidence justifies.
5. Identity. Put the real name and details back in on your own system, then read the whole report once as the client would. Check that nothing reads as generic filler.

An illustrative review shows the pattern: in a results paragraph, pass 1 found "good verbal skills across the board". The summary had said "average, relative strength", not "good across the board". Pass 3 found "likely due to anxiety", which the clinician had not concluded. Both were removed in under a minute. That is the pattern: the draft is mostly fine, and the errors are small, confident and easy to miss when reading fast.

Hours saved on a five-hour report, roughly

Suppose a report takes five hours to write, of which two are the results and background narrative, one is recommendations, one is formatting and tables, and one is review. AI from a de-identified summary mainly helps the narrative and the plain-language letter. If it cuts the narrative from two hours to about 75 minutes including the summary writing and the five-pass check, and the letter from 40 minutes to 15, you save about 70 minutes a report (45 on the narrative, 25 on the letter). Across 40 reports a year, that is a little under 47 hours. The saving is real but modest, and it depends on the discipline above. A clinician who skips the summary step to save time loses most of the benefit and takes on all of the risk.

If something has already been pasted

It happens. If a staff member or you pasted scores, a protocol or an identifiable passage into a chatbot, act the same day. Stop using that chat. Check the account's history and delete the conversation, noting that deletion may not remove copies the vendor already holds, depending on the plan and its retention terms. Record what was pasted, when, which account and plan. Then read your obligations: your board, your publisher's licence terms, your insurer and any data-protection duty to inform a client may all apply. Ask your professional body or insurer for advice rather than guessing, and add the incident to your log. Tighten the rule that failed, usually by moving the summary template into the one place everyone starts from.

Plain-language summaries: the safest high-value use

A second good use is turning a finished section into a version a family can read. The clinician pastes a de-identified paragraph they wrote and asks for a rewrite at a lower reading level, with no change in meaning. The check is a side-by-side comparison for softened or hardened claims. A paragraph saying "difficulty was clinically significant" can become "a little trouble", which changes the meaning. A paragraph saying "may be harder" can become "will be harder", which overstates it.

A quick before-and-after (illustrative). Before: "Performance on tasks requiring sustained attention was below expectation and was consistent with reported difficulties at school." After, plain version: "On tasks that needed steady concentration, the results were lower than expected, which fits what the school has described." The edit that matters: "described" is right because the school's report was a reported source, not something observed. Models sometimes change "reported" to "confirmed", and that one word can change what a report asserts.

Using the same pattern for feedback and school letters

Many testing psychologists also write a feedback letter for the family and a short summary for a school or employer. These are shorter, more repetitive and more clearly suited to a de-identified draft. Write the three or four points that the recipient needs: what was found in plain words, what helps, what the recipient should do next and who to contact. The assistant turns those points into a letter at a stated reading level.

The check is who will read it and what they are entitled to see. A school letter should hold only what the family has agreed to share, so strip anything the recipient does not need before drafting, not after. Then read the finished letter as the recipient, looking for a sentence that could be taken as a diagnosis you did not intend to give, or a recommendation that commits the school to something it has not agreed.

Consent, disclosure and the paperwork

Tell clients before assessment starts, in plain words, that you may use an AI writing tool for parts of the report, what it will and will not see, and that you write and sign the conclusions. Put it in your consent form and privacy notice, and give a way to opt out. The APA's AI guidance treats this as part of informed consent: say what the tool is for, how it is used, and its potential benefits and risks, in language the client can follow. Our consent wording tutorial has adaptable language, and AI scribes for psychologists covers the separate issue of recording sessions, which raises its own consent questions.

Keep a short log for your own file: the date, the tool and plan, what you gave it (de-identified summary, no scores) and what you changed. It takes thirty seconds and answers a board, an insurer or a client who asks. Check your professional liability insurer's position on AI, since policies differ.

Choosing a tool you can defend

For de-identified drafting, a general assistant on a business or team plan is adequate. As of November 2026, ChatGPT Business and Claude Team say they do not train on your content by default, and consumer plans let you switch training off (ChatGPT's switch, "Improve the model for everyone", is on by default for Free, Plus and Pro). Check the exact label yourself, as names change. Do not use a free consumer account with a shared login. Individual plans such as ChatGPT Plus and Claude Pro list at $20 a month; team plans list at $25 a seat monthly or $20 billed annually, with a two-seat minimum. If you would like AI built into your practice software instead, the practice software comparison sets out what is available.

A tool that handles identifiable client information is a different category. It needs a written agreement covering confidentiality, a clear retention and deletion policy, and a vendor who will answer security questions in writing; use the vendor security questions. Even then, test material and raw data should stay out unless the publisher expressly permits otherwise.

Three slips that turn up in assessment reports

Three realistic ones. The first is the "just one table" slip: a clinician pastes a small score table to ask for a plain-language explanation. It looks harmless, but it is exactly what the list forbids. The fix is to write the explanation of what each domain means in general terms, and add the numbers in your own document. The second is a plausible but unsupported sentence: the draft says "this pattern is typical of anxiety". Nothing in the summary said so. Always scan for causes, labels and prognoses the summary did not contain. The third is template drift: after a year of reusing prompts, reports start to sound the same, and clients notice. Vary the structure for each case, and keep your own voice in the formulation, which the AI should never write.

To check the workflow works, time five reports against your previous average, count the sentences you rewrote and note any statement you removed as unsupported. If you are rewriting more than half the draft, the summary is too thin or the prompt too loose. If you are removing nothing at all, you may be reading too quickly.

If you would like help setting up the template, the consent wording and the handling rules for your practice, that is what my AI implementation consultation covers.

Questions testing psychologists ask about AI reports

Is a bracketed or coded client name enough to de-identify?

Rarely on its own. A report can identify someone through a rare combination of age, school, occupation, diagnosis, family detail or an unusual history, even with the name removed. De-identify by generalising those details too, and keep the key that links a case to a person outside any chatbot.

Can I paste test scores if my plan says it does not train on my data?

That setting addresses one risk. It does not settle test publisher licence terms, your duty to protect test security, your client's consent or your retention rules. Raw scores and protocols can be test data, and your publisher and board may restrict where they go. Keep them out.

Should I tell clients I use AI to help draft reports?

Yes, in plain words, in your consent paperwork and privacy notice, and ahead of the assessment. Say what you use it for, what it never sees and that you write and sign the conclusions. Check your own board's current guidance on disclosure and consent.

Can AI suggest a diagnosis or recommendations?

It can propose generic wording for an option you have already chosen, but it should not generate diagnostic conclusions or case-specific recommendations. Those come from your clinical judgement, the full data and your training. Anything it suggests is a prompt to compare with your own reasoning, not a source.

Further reads

• Can Psychologists Use AI Beyond Session Notes? Seven Admin Jobs — Seven admin jobs that never touch client data.
• AI Consulting Report Writing: From Findings to a Client-Ready Draft — Structure for long reports from findings, outside clinical work.
• Claude vs ChatGPT for Business Writing, Proposals and Reports — Which assistant suits long, careful writing.
• How to Roll Out an AI Scribe Without Losing Patient Trust — Winning client trust when AI enters the practice.
• Is an AI Scribe Worth It for a Small Private Clinic? — Whether a scribe pays for itself in a small practice.
• How Tutoring Centres Use AI to Write Parent Progress Reports — Parent-facing progress reports, a lighter-stakes cousin.
• Medical Writing AI Software: What Small Agencies Can Use Safely — A medical writing agency owner's guide to AI software: tool tiers by plan, a never-paste list, NDA clauses and the human checks on every reference.
• Medicare Codes for Billing Psychology in Australia: AI Claim Checks — A pre-submission claim check for Australian psychologists: match MBS item numbers to times, referrals and session counts, flag mismatches and never upcode.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the APA Ethical Principles of Psychologists and Code of Conduct (Standards 4.01, 9.04 and 9.11); the APA's Ethical guidance for AI in the professional practice of health service psychology (archived copy of September 2026); the ChatGPT and Claude pricing pages and OpenAI's data controls help article (checked 4 November 2026); and test publishers' licence terms, which readers must check for their own tests.