Six Questions Forensic Evaluators Should Ask Before Letting AI Near a Report

Most forensic evaluators I talk to are not afraid of AI.

We're cautious about it, which is different.

Our reports end up in front of judges, opposing counsel, and sometimes the person we evaluated. Every sentence has to survive someone whose job is to find the weak one.

So when a new tool promises to save us time, the first question isn't "how much faster?" It's "what new vulnerability does this create?"

That question deserves a straight answer. Below are the six concerns I hear most often from colleagues, why each one is legitimate, a question you can ask of any AI tool before you upload a single page, and how we've tried to answer it in building ForensicShield.

Where we haven't fully solved something, I'll say so.

1. "Where does the report go once I upload it?"

This is the first concern almost every time, and it should be.

A forensic report can contain diagnoses, substance use history, trauma history, criminal allegations, and information about third parties who never consented to anything. Court-ordered evaluations add another layer: the examinee often didn't choose us, and didn't choose our software.

Many general-purpose AI tools were not built with that in mind. Their terms can allow inputs to be retained, reviewed, or used to improve the model.

The question to ask any tool: Is my data used to train or improve your models, how long is it kept, and will you sign a Business Associate Agreement?

How ForensicShield handles it: ForensicShield was built as a HIPAA-aligned platform from the start. We sign a Business Associate Agreement. We do not use anything uploaded to the platform to train models. Reports are encrypted at every stage.

De-identification is available as an optional feature for evaluators who want that extra layer. It isn't required to use the platform.

The honest tradeoff: no cloud tool removes your own obligations. Your jurisdiction, your contract with the court or agency, and your licensing board may set rules that go beyond HIPAA. Check those first.

2. "What if it makes something up?"

This one is no longer hypothetical.

Courts have sanctioned attorneys for filing briefs with citations that didn't exist. In Kohls v. Ellison (D. Minn. 2025), an expert's declaration was excluded after it was found to contain AI-generated citations to studies that were never published. The court's concern wasn't only the bad citations. It was what they did to the expert's credibility on everything else.

For us, the risk runs deeper than citations. A tool that drafts content can invent a collateral source, misstate a test score, or describe an interview statement the examinee never made. Once that sentence is in a signed report, it's ours.

The question to ask any tool: Does it create new content in my report? And if it cites anything, how does it verify that the source exists and says what the tool claims?

How ForensicShield handles it: ForensicShield doesn't write reports. It reads one you've already completed and flags where it may be vulnerable. Nothing it produces goes into your report unless you decide to put it there.

Its feedback does cite case law, because that's how it explains why a passage may be vulnerable. Every citation is checked multiple times, across multiple sources, for two things: that the case exists, and that it holds what the feedback says it holds.

If a citation can't be verified on every pass, it's marked "not verified." It is never presented as settled.

The honest tradeoff: verification lowers the risk. It doesn't erase it. Before any case goes into something you sign or testify about, read it yourself. That's true of a citation from any tool, or any colleague.

3. "Will it start forming opinions for me?"

Our opinions are the product of our training, our data, and our reasoning. The Specialty Guidelines for Forensic Psychology ask us to be able to explain the basis for every one of them.

A tool that suggests a diagnosis, a risk level, or a competency conclusion puts something between us and that explanation. Even if we disagree with it, it has entered the room.

The question to ask any tool: Does it ever offer a diagnostic, risk, or psycholegal conclusion, even as a "suggestion"?

How ForensicShield handles it: It doesn't. ForensicShield looks at how your opinion is supported and expressed, not whether it's correct. It might note that a conclusion isn't clearly tied to the data you cite, or that a phrase overstates certainty. Whether the opinion itself holds is your call, and only yours.

The honest tradeoff: that means ForensicShield can't tell you whether you got the answer right. Peer consultation still matters for that. We built it to sit alongside consultation, not replace it.

4. "Could I explain it on the stand?"

The December 2023 amendment to Federal Rule of Evidence 702 made explicit that the proponent of expert testimony must show, by a preponderance of the evidence, that the opinion reflects a reliable application of reliable methods. Many states look to the federal rule.

If an AI tool shaped any part of how we reached or wrote an opinion, a sharp cross-examiner can ask us to explain how that tool works. "I'm not sure" is not a good answer.

The question to ask any tool: If I'm asked under oath what this tool did to my report, can I answer in two sentences?

How ForensicShield handles it: The answer is meant to be simple. "I wrote and signed the report. Before submitting it, I used a quality-assurance tool that flagged passages for my review. I decided what, if anything, to change."

That's the same structure as asking a colleague to read a draft. The methodology that produced the opinion is still yours, start to finish.

Every flag also shows its reasoning and the case law that reasoning rests on. So if you're asked why you changed a passage, or why you didn't, you can explain it in your own words.

5. "Do I have to disclose it, and is it discoverable?"

This is the concern with the least settled answer.

Courts, bar associations, and licensing boards are still working out what AI use requires experts to disclose. Some judges now have standing orders about AI in filings. Expectations differ by jurisdiction and are changing quickly.

I'm not going to tell you what your obligations are. That's a conversation for retaining counsel, and it's worth having before a case starts, not during a deposition.

The question to ask any tool: If I'm required to disclose that I used this tool, would I be comfortable describing exactly what it did?

How ForensicShield handles it: Because ForensicShield reviews rather than writes, disclosure is usually a short, unremarkable sentence.

ForensicShield also keeps a record of what was flagged, and stores iterations of your report so you can compare how it changed. For quality control, that history is useful.

The honest tradeoff: that same history may be discoverable. Know what exists, and talk it through with retaining counsel before you're asked about it.

6. "Will it make us worse at this?"

This is the quietest concern, and maybe the most important one.

Automation bias is well documented: when a system checks our work, we tend to check less ourselves. Early-career evaluators who lean on a tool may never build the instincts it was supposed to support.

The question to ask any tool: Does it explain why something is a problem, or just tell me to fix it?

How ForensicShield handles it: Every piece of feedback explains why: the reasoning behind the flag and the case law it rests on. The goal is for you to recognize the pattern yourself next time, before any tool sees the report.

That also makes it useful in training and supervision, where understanding the why matters more than the fix.

The honest tradeoff: no tool can stop someone from clicking "accept" without thinking. That part is on us.

What ForensicShield doesn't do

We'd rather you hear this from us.

•         It doesn't write, draft, or rewrite reports.

•         It doesn't offer diagnoses, risk levels, or psycholegal opinions.

•         It can't evaluate what it can't see: raw test data, interview recordings, or collateral records that aren't in the report.

•         It can't guarantee a report will survive challenge. It can only help you find where it's most likely to be challenged.

•         It doesn't replace peer consultation, supervision, or your own judgment.

The question underneath all six

Every one of these concerns comes back to the same thing: our names are on the report.

Any tool we bring into that process should make us more able to stand behind our work, not less.

That's the standard we built ForensicShield to meet. It answers one question: if this report were challenged tomorrow, where would it be vulnerable?

If you're weighing AI tools for your own practice, I'd encourage you to put all six questions above to every vendor, including us.

Which of these concerns weighs on you most? Or is there one I missed?

www.forensicshield.net