Where Does “Using AI” Begin? Authorship and Automated Tools in Psychological Evaluation Reports

Psychologists are being asked, by referral sources, attorneys, and sometimes clients, whether AI was used to write a report. The question sounds simple. In practice it isn’t, because software already touches almost every evaluation report long before anyone opens a chatbot.

Consider the tools a typical evaluation passes through. Which of these count as “using AI to create a document”?

• An AI tool helps develop an outline.
• Dictation software transcribes the psychologist’s spoken draft.
• A word processor checks spelling.
• A word processor checks grammar.
• An OCR converter turns scanned records into text.
• A program extracts specific fields, such as medications or diagnoses, from a PDF.
• A program restructures text into a particular report format.
• A program “cleans up” language the writer finds awkward, and the writer brings it back into the document.
• A program turns data fields into narrative.

There is no official bright line yet. This post works through where a defensible one can be drawn, and why the answer depends less on which software you use than on what it does to the content and what you do with its output.

Function, not label

Tool labels don’t settle the question. The “grammar checker” built into a modern word processor may be a generative model rewriting whole sentences. “OCR” may run on a neural network. A more useful approach sorts tools by what they do to the content, using two questions:

1. Did the tool introduce words, facts, or inferences the psychologist didn’t originate?
2. Could it have changed the meaning in a way the psychologist wouldn’t catch without checking a source?

Applied to the list above, the tools fall roughly into three groups.

Clerical tools. Traditional spell check, rule-based grammar check, and OCR either flag errors for the writer to accept or reproduce an existing document that can be checked against its source. Their errors are clerical, not authorial.

Assisted transformation. Dictation, field extraction, reformatting, and language cleanup all work on content the psychologist originated, but each can alter it. Speech recognition drops a “not” or confuses “hypo-” and “hyper-.” An LLM extracting medications can miss one or invent a plausible one. Language cleanup carries a subtler risk: drift in certainty. “Suggests” becomes “indicates.” “Consistent with” becomes “meets criteria for.” In an evaluation, those words carry real weight.

Generative drafting. Converting data fields into narrative is AI-written text, even when every fact came from the psychologist, because the model chooses emphasis, transitions, and framing. An AI-built outline belongs here too. In an evaluation, the outline decides which questions get addressed and in what order, which is part of the clinical reasoning.

We’ve been here before: the PAI interpretive report

That framework raises an obvious objection. A publisher’s interpretive report for the PAI generates narrative the psychologist didn’t write. Is it AI?

It isn’t. It’s computer-based test interpretation (CBTI): fixed, validated decision rules mapped to prewritten statements. The same profile always produces the same text, and every statement traces back to specific scale elevations. But it does produce interpretive content the psychologist didn’t author, which means authorship and “AI” are two separate questions.

The profession has handled the authorship half for decades. APA Ethics Code Standard 9.09(c) holds that psychologists keep responsibility for how assessment instruments are applied, interpreted, and used, whether they interpret tests themselves or rely on automated services. Accepted practice is to treat CBTI statements as hypotheses, integrate them with history, interview, collateral data, and other measures, and attribute any quoted text. Interpretive-report language pasted verbatim into a forensic report is a familiar target on cross-examination.

CBTI is the best existing model for AI-generated content. The obligations are the same. What differs is how much the tool can be trusted before checking. CBTI output is constrained by validated rules and traceable to scores. Generative output is neither, so every claim needs verification, not just a judgment about relevance.

“I told it not to interpret”

What if the LLM is instructed not to interpret or guess? That instruction moves the tool from interpreter to transformer, which is real progress. But it’s a request, not a guarantee. Nothing in the model enforces it, and lowering the temperature makes output repeatable, not validated. The only way to confirm compliance on a given document is to check the output against its source.

How much the instruction helps depends on the task:

• Field extraction constrains well, because compliance can be checked mechanically. If the tool must return the exact source text for every item it extracts, software can reject anything that doesn’t match, and a hallucinated medication can’t get through.
• Reformatting constrains fairly well. A diff can confirm nothing was added or reworded.
• Language cleanup is harder, because rewording is the task. A diff shows what changed, but someone still has to judge whether the meaning shifted.
• Data-to-narrative can’t be made interpretation-free, even with perfect compliance. Order, emphasis, and connecting words like “however” and “despite” are interpretive choices. The model can avoid adding facts. It can’t avoid framing them.

In testimony, “every extracted item is automatically checked against the source document” carries far more weight than “I told it not to guess.”

Doesn’t spell check infer too?

It does. Spell check guesses that “recieve” meant “receive,” and it can introduce meaningful errors, like “elicit” corrected to “illicit.” So “inference versus no inference” isn’t the dividing line. Nearly every tool infers something.

A more useful distinction is what the inference is about. Spell check infers what the writer meant to type, and the writer is the authority on that. Interpretive tools, including CBTI, make inferences about the client. Those are clinical claims, and that’s where authorship matters.

What about the review process? It’s tempting to say spell check is safe because each change is approved individually. But accepting a wrong correction with “change all” spreads an error through a document just as surely. And if an LLM’s edits are shown as tracked changes, each accepted or rejected one at a time, the mechanics are the same as grammar check.

What remains different is how much one change can do. A spelling suggestion is limited to a few candidate words. A single LLM edit can turn “symptoms may be consistent with” into “symptoms indicate,” and it reads fluently enough to look right. That’s a difference of degree, not kind. It calls for more careful review of each change, not a different category of tool.

Following that argument through leads to the central point: the tool doesn’t make you the author or not the author. Your review does. Whatever a psychologist individually examines and adopts through professional judgment is the psychologist’s own, whether it came from spell check, an LLM, or a colleague’s edits. That’s the same standard that already governs CBTI.

Disclose without disclaiming

If authorship comes from review, why disclose AI use at all? Some courts, agencies, and referral sources now expect it regardless. But a disclosure statement raises its own worry: doesn’t it weaken every statement in the report, since the reader can’t tell which parts the psychologist fully owns?

Only if it’s written as a disclaimer. A disclosure says how the work was produced. A disclaimer limits responsibility for it. “Portions of this report were generated with AI assistance and may contain errors” tells the reader some sentences are less yours than others, without saying which, and every statement becomes suspect.

The fix is to own everything explicitly:

Drafting tools, including AI-assisted transcription and language editing, were used in preparing this report. All content was reviewed by the undersigned, and every statement, finding, and conclusion reflects my independent professional judgment, for which I take full responsibility.

That leaves the reader nothing to sort out. It actually strengthens the report, because it’s an affirmative attestation covering every sentence. Reports already work this way for other help: statements that tests were scored by publisher software, that a psychometrist administered the battery, or that an interpreter assisted don’t undermine the conclusions, because the signing psychologist clearly adopts all of it.

Two practical points follow. Don’t tag individual passages as AI-assisted, which creates exactly the selective-ownership problem. And weigh the alternative: AI use that surfaces later through a subpoena or deposition looks like concealment, which costs far more credibility than a clear attestation up front.

What the standards and Ohio law say

No current rule draws the line for us, but several existing standards point the same direction.

APA Ethics Code. Several standards bear directly on AI-assisted reports, even though none mention AI:

• 2.01, Boundaries of Competence: using a tool competently includes understanding how it can fail.
• 2.04, Bases for Scientific and Professional Judgments: conclusions must rest on established knowledge, not on unvalidated model output.
• 4.01, Maintaining Confidentiality: sending client data to a tool is a disclosure, which is why self-hosting or a HIPAA business associate agreement matters.
• 5.01, Avoidance of False or Deceptive Statements: relevant both to unverified AI errors and to how AI use is described.
• 9.01, Bases for Assessments: opinions must be grounded in information sufficient to substantiate them.
• 9.09, Test Scoring and Interpretation Services: the psychologist retains responsibility however interpretation is produced.

APA’s AI guidance. In June 2025, APA published Ethical Guidance for AI in the Professional Practice of Health Service Psychology. Its consistent theme is that AI should support professional judgment, not replace it, with emphasis on transparency, data privacy, validation of tools, and human oversight.

Ohio rules of professional conduct. Ohio Administrative Code 4732-17-01, promulgated under Revised Code 4732.17, sets the Board of Psychology’s rules of professional conduct. A violation constitutes unprofessional conduct. The rule directs that APA and ASPPB ethics codes be used to resolve ambiguities, with Ohio’s rules prevailing in any conflict. That means APA’s standards and guidance on AI are likely to shape how the Board reads general duties. The rule also treats a psychologist as negligent when conduct clearly falls below accepted practice standards, and it requires the record to include a copy of any reports prepared.

Pending Ohio legislation. House Bill 525 of the 136th General Assembly would enact Revised Code 4743.21 on AI in therapy services. It would cover psychologists among other licensed therapy professionals. As introduced, it allows AI for administrative and supplementary support, but prohibits using it for independent diagnostic decisions, direct therapeutic communication, or treatment recommendations without professional review. As of this writing, the bill remains pending in the House Health Committee, where it was referred on October 22, 2025. It’s aimed at therapy rather than evaluation, but its core principle, that AI may assist while the licensed professional decides, matches the framework above.

Ohio courts. In April 2026, the Ohio Board of Professional Conduct issued an AI ethics guide for lawyers and judges. It notes that Ohio courts vary widely, with some requiring disclosure of AI use in filings and others prohibiting generative AI in filings. One Ohio lawyer was sanctioned in State v. Coleman, 2026-Ohio-965, after hallucinated facts and quotations appeared in court filings. Those rules govern attorneys, not psychologists, but a forensic report becomes part of what an attorney files. Expect the same scrutiny to reach evaluators.

Ohio Psychological Association. OPA’s Communication and Technology Committee has also published guidelines for AI use in psychology, a useful state-specific resource.

The bottom line

All of these tools sit on a continuum, so any line is a reasoned one rather than a natural one. A defensible version:

• A tool raises authorship concerns when it makes inferences about the client, or makes changes the psychologist can’t individually see and approve.
• Review effort should scale with how much a single accepted change could alter clinical meaning.
• The psychologist is the author of whatever they examine and adopt through professional judgment, regardless of its source.
• Disclosure should describe the process and affirm ownership of the whole report, never disclaim parts of it.

What holds up under scrutiny isn’t avoiding AI. It’s being able to say exactly what each tool did, how its output was verified, and why every sentence in the report is your own.