What the AI Disclosure Laws Actually Require, and Why a Verification Step Isn't What They're Worried About

The question I hear most often now isn't whether to use AI tools in forensic work.

It's a quieter worry. If I use one, do I have to disclose it, and to whom, and will that disclosure become the thing opposing counsel uses against me?

It's a fair worry. The legal landscape around AI disclosure has filled in fast, it's uneven from state to state, and most of what's written about it is aimed at corporate compliance officers rather than a solo evaluator trying to write a defensible report. So it's worth walking through what these laws actually say, and, just as important, what they don't reach at all.

One caveat before I start. I'm a forensic psychologist, not an attorney, and nothing here is legal advice. Disclosure obligations turn on your jurisdiction, your license, and the specific facts of what you're doing. Run your own situation past your own counsel. What I can offer is the map, and a way to think about where any given tool sits on it.

There is no federal AI disclosure statute. There is a patchwork.

Start with the part that surprises people. As of now, there is no federal statute, and no adopted amendment to the Federal Rules of Civil Procedure or the Federal Rules of Evidence, that requires you to disclose your use of AI.

There is a proposed rule, and it's worth knowing about so you aren't caught flat-footed. The Advisory Committee's draft Federal Rule of Evidence 707 went out for public comment through early 2026. It addresses something other than disclosure. It would hold machine-generated evidence offered without a human expert to the same reliability standard as expert testimony under Rule 702. In other words, it's about whether a machine's output is admissible, not about whether you have to announce a tool you used to prepare a report. And for now, it remains only a proposal.

What actually governs the ground today is a growing set of judge-specific standing orders, district by district and sometimes courtroom by courtroom.

The pattern across them holds even when the wording differs. A judge in the Southern District of New York requires parties to disclose when a generative AI tool was used to prepare a filing and to certify that a human independently reviewed and verified its accuracy. A judge in the Northern District of Texas requires attorneys to certify that any language drafted by generative AI was checked for accuracy against print reporters or traditional legal databases by a human being. A few judges go further and prohibit AI in briefing outright. The most comprehensive orders reach past briefs to evidence itself.

Two things are worth noticing about all of this.

First, most of these orders are aimed at attorney filings, not at expert reports directly. Whether a given order reaches your report is not always clear, and that ambiguity is exactly the kind of thing to raise with retaining counsel before you're on the stand rather than after.

Second, look at what the orders actually demand. Not "don't use these tools." They demand disclosure, and a certification that a human verified the output. The courts are not policing the existence of AI in your workflow. They're policing unverified AI presented as if it were checked. That distinction carries most of the weight here, and I'll come back to it.

The state laws regulate AI at the point it touches a person

The state statutes are where most of the anxiety lives, because there are a lot of them and they don't look alike. But there's a structural logic underneath the patchwork, and once you see it, most of the worry drains out.

These laws regulate AI at the moment it interacts with, communicates to, or makes a decision about a person.

Utah is the clearest example for our field. The Utah AI Policy Act, amended in 2025, requires professionals in licensed occupations, psychology among them, to give a prominent disclosure at the start of what it calls a high-risk AI interaction. That means one where AI is collecting sensitive personal information, or giving personalized mental-health, medical, legal, or financial advice that could influence a significant decision. The trigger is an AI system interacting with the person directly. The disclosure has to come at the start of that interaction, and violations carry penalties.

California's AB 3030, in effect since the start of 2025, points the same direction from the healthcare side. It requires providers to disclose when generative AI is used to produce a patient communication about clinical information, and to tell the patient how to reach a human. The part that rarely makes the headline is the exemption. The law does not require those disclosures for any communication a licensed provider has reviewed and approved. The statute treats a human clinician in the loop as the thing that resolves the concern.

Colorado and Texas are building variations on the same frame: consumer-facing AI, consequential decisions, direct interaction. Both moved again this year. Texas's Responsible Artificial Intelligence Governance Act took effect at the start of 2026, and Colorado's AI Act has already been rewritten, with the revised version now pointing to a 2027 compliance date after enforcement of the original stalled in federal-court litigation. Meanwhile, federal efforts to preempt this state-by-state patchwork have so far failed, which means the patchwork is what governs. The details differ and they keep changing, which is its own reason to check current law in your state rather than trust any article's snapshot, including this one.

But hold the common thread. Every one of these statutes is worried about AI that reaches the evaluee. A chatbot dispensing advice. An automated message to a patient. A system making or steering a decision about someone's care, credit, or liberty. That is a different activity from a licensed expert using software to pressure-test a report they have already written, where the tool never contacts the evaluee, never generates the opinion, and produces nothing that leaves the clinician's own review.

What our own field says, and, more tellingly, what it doesn't

The statutes are only half the picture. The other half is what our professional bodies have said, and here the guidance is newer, thinner, and more honest about its own gaps.

On the psychology side, the American Psychological Association issued ethical guidance on AI in the professional practice of health service psychology in June 2025. Its spine is what you'd expect and what most of us already believe. AI should augment rather than replace human judgment. The psychologist critically evaluates any AI output. The psychologist retains responsibility for the accuracy of the final work.

On disclosure, the guidance scales the obligation to how substantial the AI's role is. Predictive text while drafting a note sits at one end, AI shaping a treatment decision at the other. The American Psychiatric Association reached similar ground earlier, in a 2024 position statement holding that patients should be informed when clinical decisions are being driven by AI, that AI belongs in an augmentative role, and, notably, that AI fabricating citations and medical information is a specific hazard to guard against.

Both of those, though, are written for the clinical relationship. Neither was built for the courtroom, and neither tells a forensic evaluator what to do about the report headed for a judge.

The most directly forensic treatment I've seen came in 2024, in an article published through the American Board of Professional Psychology by four board-certified forensic psychologists. It's worth being precise about what that is. It's authored commentary, not binding board policy. But it says the quiet part out loud. The field, the authors note, "has yet to offer suggestions for addressing AI in informed consent, report writing, or expert testimony." They urge transparency about AI use and its limitations. They warn that these tools can be "blatantly wrong" in ways an inexperienced user won't catch. And they flag the admissibility problem underneath all of it. Proprietary, black-box algorithms resist the testability, error rates, and peer review that Daubert asks for, and nothing yet establishes AI use as "generally accepted" in forensic psychology.

The Specialty Guidelines for Forensic Psychology predate generative AI entirely. They were adopted in 2011 and published in 2013, so they name none of this. But their spine still governs. Be transparent about your methods and the data behind your opinion. Document your process. Don't overstate. That lens applies cleanly to a tool whether or not the tool existed when the guidance was written.

So the honest summary is this. No psychology or psychiatry body has yet published a forensic-specific rule telling you when or how to disclose AI use in a report or on the stand. That absence is not a permission slip.

Read the other way, it's a reason to hold to the most conservative, defensible posture available. Be transparent about a tool you personally stand behind, and stay vigilant that AI is never characterized as the source of your opinion, because that characterization is exactly where the "generally accepted" admissibility problem bites.

Where a verification step actually sits

This is the part I was asked to address directly, and I want to do it carefully rather than conveniently.

I built ForensicShield as a quality-assurance layer. It reviews a finished report the way a tough colleague would before you file it, looking for the unverified citation, the certainty language that draws Rule 702 challenges, the methodological gap that reads clean until someone pulls the source. It does not write the report. It does not form or express a clinical opinion. It does not interact with the evaluee. And in our architecture, the report content is processed within a single secure environment and is not exposed to a third-party AI provider.

I'm not going to tell you that using it, or any tool, relieves you of a disclosure obligation. That's a legal conclusion, and it depends on facts I can't see from here. That determination belongs to your attorney.

What I'll say is narrower, and I think more useful. The disclosure statutes on the books are built to govern AI that touches the person receiving the service. A back-office verification step that checks your own finished work, with you as the licensed human reviewing and standing behind every word, is a different category of activity. And where these laws do carve exemptions, the licensed human in the loop is repeatedly the thing that satisfies them. Whether that reasoning controls in your jurisdiction is a question for counsel. But it is not the scenario these laws were written to catch.

The disclosure worth making

Here's what I'd actually lose sleep over, and it isn't a statute.

In the forensic context, the disclosure that matters most is the one to the court, and the risk there isn't disclosing too much. It's the appearance of concealment. An expert who can say plainly, "I used a tool to check my finished report for reliability, and I independently verified every finding and citation myself," is describing diligence. An expert who used something, didn't mention it, and gets it drawn out on cross is handing the other side a credibility story that has nothing to do with whether the underlying work was sound.

That framing also keeps you clear of the real trap. What you disclose is never "AI wrote my report." It's that you checked your own work and remain its author. The moment a tool is described as the source of the opinion rather than a check on it, you've created a methodology problem that no disclosure statute required and no disclosure can fix.

So the anxiety, I think, is aimed slightly off target. The laws are narrower than they feel, and the honest, protective move in our world, transparency about a verification step you personally stand behind, was already the right one before any of them passed.

When you picture disclosing a tool you used to check a report, what's the fear underneath it? The statute, or the cross-examination?

References

American Psychiatric Association. (2024). Position statement on the role of augmented intelligence in clinical practice and research. https://www.psychiatry.org/getattachment/a05f1fa4-2016-422c-bc53-5960c47890bb/Position-Statement-Role-of-AI.pdf

American Psychological Association. (2013). Specialty guidelines for forensic psychology. American Psychologist, 68(1), 7–19. https://doi.org/10.1037/a0029889

American Psychological Association. (2025). Ethical guidance for AI in the professional practice of health service psychology. https://www.apa.org/topics/artificial-intelligence-machine-learning/ethical-guidance-ai-professional-practice

Hodges, H. J., Armstrong, N. E., Formon, D. L., & Silber, B. J. (2024). Implications for artificial intelligence in forensic psychological practice and board certification. On Board with Professional Psychology, 2(2).

American Board of Professional Psychology. https://abpp.org/newsletter-post/implications-for-artificial-intelligence-in-forensic-psychological-practice-and-board-certification/

Legal authorities

Health Care Services: Artificial Intelligence, Assemb. B. 3030, 2023–2024 Reg. Sess. (Cal. 2024).

Colorado Artificial Intelligence Act, S.B. 24-205, 2024 Reg. Sess. (Colo. 2024) (amended by S.B. 26-189, 2026).

Fed. R. Evid. 702.

Fed. R. Evid. 707 (proposed draft published for public comment Aug. 2025).

Standing Order for Civil Cases Before Judge Vernon S. Broderick (Use of Generative Artificial Intelligence), U.S. District Court for the Southern District of New York.

Mandatory Certification Regarding Generative Artificial Intelligence, Judge Brantley Starr, U.S. District Court for the Northern District of Texas.

Texas Responsible Artificial Intelligence Governance Act, H.B. 149, 89th Leg., R.S. (Tex. 2025).

Utah Artificial Intelligence Policy Act, S.B. 149, 2024 Gen. Sess. (Utah 2024) (amended by S.B. 226, S.B. 332 & H.B. 452, 2025 Gen. Sess.).