Your Prompts Are Discoverable: What 2026's AI Rulings Mean for Forensic Evaluators

Most of the conversation about AI in the courtroom has been aimed at attorneys. That conversation has now reached the expert's chair and it arrived through discovery rather than through a new rule.

If you perform forensic evaluations, the developments of the past four months change something concrete about how you work. Not because a new standard was adopted. Because two courts decided that how an expert used AI is part of the expert's methodology and methodology has always been fair game.

Here is what actually happened, what did not happen, and what it means for the report currently sitting on your desk.

What did not happen: Rule 707 stalled

Proposed Federal Rule of Evidence 707 has been widely described as the coming framework for AI evidence. It is worth being precise about its status, because a fair amount of commentary has overstated it.

The proposed rule reads:

When machine-generated evidence is offered without an expert witness and would be subject to Rule 702 if testified to by a witness, the court may admit the evidence only if it satisfies the requirements of Rule 702(a)–(d). This rule does not apply to the output of simple scientific instruments.

The timeline:

• May 2025 — The Advisory Committee on Evidence Rules voted 8–1 to publish the rule for comment. The Department of Justice dissented.

• June 2025 — The Standing Committee on Rules of Practice and Procedure approved publication.

• August 2025 – February 2026 — Public comment period.

• May 2026 — The Advisory Committee reviewed more than seventy comments, noted greater overall concerns than anticipated, and chose not to advance the rule.

• June 2026 — The Standing Committee decided not to recommend action on Rule 707 at this time.

• October 2026 — A mini-conference on the rule is scheduled for the Advisory Committee's next meeting.

There is no Judicial Conference approval, no transmittal to the Supreme Court, and no effective date. The objections were substantive: the American Association for Justice argued the rule was overbroad enough to sweep in routine material such as geolocation data and surveillance footage; others noted it does nothing about deepfakes, because it only applies when the proponent concedes the evidence is machine-generated; and several commenters raised the resource asymmetry created by requiring technical rebuttal experts.

The more important point for forensic evaluators is a scope issue that has gone largely unremarked. Rule 707, even if adopted tomorrow, applies only to machine-generated evidence offered without an expert witness. Anything you use in forming or supporting an opinion has never been within its reach. Your work stays where it has always been: under Rule 702 and the applicable Daubert or Frye framework.

That is not a lower bar. The December 2023 amendment to Rule 702 tightened it considerably. The proponent must now demonstrate to the court by a preponderance of the evidence that the reliability requirements are met, and the opinion must reflect a reliable application of the methodology to the facts of the case. That amendment is in force. It is doing the work that Rule 707 has not yet been adopted to do.

What did happen: two rulings that reached the expert

An expert's AI prompts were ordered produced

In Conservation Law Foundation, Inc. v. Shell Oil Co., a magistrate judge in the District of Connecticut ruled in May 2026 that an expert's generative AI prompts were discoverable.

The expert had used a generative AI tool to filter voluminous document productions while preparing her report. She disclosed the AI use. She did not initially produce the prompts. The court held that an expert witness's methodology is fair ground for discovery, and that the AI-assisted filtering process qualified as methodology. The court further found that the parties' existing protective stipulation did not shield the prompts, because its language was not explicit about AI materials.

The order was narrow on its face — it reached the prompts, not the full universe of inputs and outputs. The reasoning is not narrow. If prompt design is methodology, there is no obvious principle that stops at prompts.

The order is not final. The Conservation Law Foundation filed a Rule 72(a) objection to the district judge, an objection within the district court, not an appeal, and as of early June 2026 that objection remained pending with enforcement stayed. No public ruling had issued as of this writing. The question is unresolved. It is also now being asked in depositions wherever AI use is suspected.

And in another matter, the logs became the exhibit

In litigation arising from the 2020 Watson Grinding explosion in Houston, reporting in August 2026 described an engineering expert whose report had been generated substantially by a consumer AI chatbot. Opposing counsel suspected as much and sought the underlying chat activity in discovery. Roughly 350 pages were produced.

What surfaced was not a fabricated citation. It was the prompt history, including instructions directing the tool to produce a report defending the retaining party's standard of care and to show that it bore no fault. The expert acknowledged at deposition that the prompts had been framed to favor the client, and plaintiffs later called him at trial to testify about the AI use.

Two points of precision, because they matter. No court sanctioned the expert or excluded his testimony; the matter was tried in Texas state court, so there was no federal Rule 702 ruling. And the jury returned roughly $61.5 million, apportioning thirty percent responsibility to the retaining party. The consequence here was not a sanction. It was that the expert's own file became the other side's best exhibit.

This is the failure mode worth sitting with. The exposure was not inaccuracy. It was documented outcome-directed reasoning, a written record of the expert working backward from a conclusion, produced in the expert's own words, admissible, and impossible to explain away. Five years ago that record did not exist. Today, for anyone using a chatbot in report preparation, it does.

The line courts are actually drawing

Read together, the expert-side decisions of the past three years describe a reasonably coherent boundary.

Where AI generated the analysis, the testimony did not survive. In Kohls v. Ellison (D. Minn. 2025), an expert declaration containing AI-fabricated citations was excluded outright; the court observed it could not accept false statements, innocent or not, in an expert's declaration. In In re Celsius Network LLC (Bankr. S.D.N.Y. 2023), a 172-page report generated by AI in roughly 72 hours, signed by the expert but not written by him, was excluded in its entirety. In Matter of Weber (N.Y. Surr. 2024), an expert who used Microsoft Copilot for damages calculations but could not recall his prompts or explain how the tool worked had his opinions rejected as unreliable, and the court went further, holding that AI-assisted evidence warrants a Frye hearing and that counsel has a duty to disclose AI use. That last holding is the one forensic readers should note.

Where AI checked work the expert had already done, the testimony was admitted. In Ferlito v. Harbor Freight Tools (E.D.N.Y. 2025), the expert used a chatbot only to confirm conclusions he had reached through decades of experience. The court admitted the testimony and drew the distinction explicitly.

The organizing principle is not whether AI touched the file. It is whether the expert's reasoning was delegated or merely assisted. Courts are asking who did the thinking. An expert who can articulate the analysis independently of the tool, explain what the tool did and did not contribute, and demonstrate that the conclusion was his or her own is in materially different shape than one who cannot.

Worth noting for context: the broader wave of AI sanctions has fallen almost entirely on attorneys filing fabricated citations, not on experts. Counts vary enormously depending on methodology. One deliberately narrow tracker, limited to U.S. matters where a licensed attorney was actually sanctioned for fabricated legal authority, verified nineteen through June 2026, with consequences beyond monetary fines in eleven, including three suspensions. A broader academic database using looser inclusion criteria lists more than 1,300 U.S. entries and roughly 1,900 worldwide. Either way the trend line is up.

Expert-side exposure runs through a different channel entirely: exclusion, impeachment, and credibility. Those consequences do not appear on a sanctions tracker. They appear in the transcript.

Where the profession stands

The professional guidance has not caught up to the case law, and it is worth being candid about that gap.

The APA's Ethical Guidance for AI in the Professional Practice of Health Service Psychology (June 2025) established the general architecture: AI should augment rather than replace human decision-making, final responsibility rests with the practitioner, and disclosure obligations scale with the significance of the use — a distinction the guidance draws between subtle applications such as predictive text and substantial ones that shape professional conclusions. The companion Responsible Use of AI in Assessment materials address testing specifically.

Two caveats belong with that. The guidance is aspirational rather than enforceable under the Ethics Code, and it is written for health service psychology. In forensic work there is no patient to consent; the disclosure duty runs to the retaining party and, ultimately, to the court. The architecture transfers. The mechanics do not.

The American Board of Professional Psychology's forensic specialty commentary is more pointed. It observes that no research, case law, or official position currently establishes AI as generally accepted within forensic psychology, which is itself a Daubert problem, compounded by the fact that most commercial algorithms are proprietary and therefore not testable, not peer-reviewed, and without a published error rate. It also documents concrete failures, including a chatbot misinterpreting an Inventory of Legal Knowledge score, and notes poor performance on violence risk assessment and psycholegal questions.

A framework published in August 2026 in the Journal of Forensic Psychology Research and Practice offers the most operationally useful structure to date. Its core principle is non-delegation of cognition, the expert retains complete analytical ownership. It distinguishes substantive uses, which materially influence the opinion and warrant proactive disclosure in the report itself, from peripheral administrative uses, which do not. It calls for zero-trust verification of every AI output against primary sources, demonstrated technical competence sufficient to explain a tool's probabilistic limitations under examination, and enterprise-grade data security that preserves privilege.

The Specialty Guidelines for Forensic Psychology, adopted in 2011, published in 2013, and extended by APA through December 31, 2026, contain nothing on AI, and no revision addressing it has been published. In that vacuum, several state associations have issued their own guidance. For now, the practicing evaluator is operating ahead of the standard.

What this means in practice

Five things follow reasonably directly from the above.

Assume the record is discoverable. Whatever tool you use, work as though the interaction history could be produced. That is not yet settled law, and the pending objection in Shell may narrow it. But designing your workflow around the assumption costs little and protects a great deal.

Never prompt toward a conclusion. The Watson Grinding record is instructive precisely because the problem was visible in the prompts themselves. Consider whether any instruction you give a tool would read defensibly if displayed on a screen during cross.

Know which side of the line your use falls on. Generation of analysis and confirmation of analysis are being treated very differently. Consider documenting, contemporaneously, that the reasoning was yours.

Be able to explain the tool. Under Daubert, an expert who cannot describe what a tool does, what its limitations are, and how its output was validated has a methodology problem regardless of whether the output was correct.

Consider whether disclosure is warranted. The emerging distinction, substantive versus peripheral, is a useful place to start, and retaining counsel is increasingly asking the question in writing. Whether and how to disclose in a particular matter is a question for the retaining attorney.

Where ForensicShield fits

These developments are the reason ForensicShield was built the way it was.

ForensicShield does not write reports and does not generate clinical opinions. It reviews a completed report and identifies where its defensibility could be strengthened, construct adequacy, language that may be exploited, and the gaps opposing counsel is most likely to target. The evaluator retains full authorship and full responsibility for every conclusion, which is the posture the case law is converging on and the one the professional guidance already requires.

If you would like to see how structured pre-testimony review works on your own material, you can request access at forensicshield.net.

ForensicShield is a report review tool. It does not provide clinical diagnoses, treatment recommendations, or clinical opinions. This article is a summary of publicly reported legal and professional developments for informational purposes and is not legal advice. Questions about discovery obligations, disclosure, or admissibility in a specific matter should be directed to counsel.

Sources: National Law Review and Nelson Mullins on the status of proposed FRE 707; Mayer Brown on Conservation Law Foundation, Inc. v. Shell Oil Co.; Drug & Device Law on expert witnesses' use of AI and on court AI disclosure orders; contemporaneous reporting on the Watson Grinding litigation; American Board of Professional Psychology, Implications for Artificial Intelligence in Forensic Psychological Practice and Board Certification; American Psychological Association, Ethical Guidance for AI in the Professional Practice of Health Service Psychology (2025) and Responsible Use of AI in Assessment; Rilen, Journal of Forensic Psychology Research and Practice (2026).