Courts have begun excluding expert testimony and striking expert declarations when experts use generative AI tools (“gen AI”) without adequate verification or disclosure. In Kohls v. Ellison, No. 24-CV-3754, 2025 WL 66514 (D. Minn. Jan. 10, 2025), a federal court excluded an expert’s declaration in its entirety because the expert used GPT-4o and failed to verify AI-generated citations—some of which were fabricated. In Matter of Weber as Trustee of Michael S. Weber Trust, 85 Misc.3d 727, 220 N.Y.S.3d 620 (N.Y. Sur. Ct., Saratoga Cnty. 2024), the Surrogate’s Court found a damages expert’s gen AI-assisted calculations unreliable under the Frye standard and imposed an affirmative duty on counsel to disclose AI use before such evidence is introduced. Together, these decisions underscore why expert engagement agreements should expressly address gen AI use—and why practitioners who fail to do so may risk exclusion of critical testimony, sanctions, and malpractice exposure.

The Legal Framework

Two recent decisions illustrate the courts’ developing approach to expert use of gen AI tools.

Kohls v. Ellison involved a Stanford professor and AI researcher who submitted a declaration citing two nonexistent academic articles and misattributing a third. The expert used OpenAI’s GPT-4o to identify supporting literature. Although he routinely validated citations with reference software in his academic work, he failed to do so before submitting the sworn declaration. The court excluded the declaration in its entirety, holding that fabricated AI-generated sources “shatter” expert credibility under Federal Rule of Evidence 702 and Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993). The court emphasized that intent was irrelevant: the declaration was excluded even though the errors were “innocent.” It further observed that the expert’s reliance on unverified gen AI output represented an “abdication of independent judgment,” which is the very quality that makes expert testimony admissible. The court suggested that Rule 11 “may now require attorneys to ask their witnesses whether they have used AI in drafting their declarations and what they have done to verify any AI-generated content.”

Matter of Weber provides the more detailed roadmap for AI-evidence challenges. Charles Ranson, a damages expert in a probate and trust-accounting dispute, used Microsoft Copilot to “cross-check” his calculations. At his deposition, Ranson could not recall the prompt he used, identify the sources on which Copilot relied, or explain how Copilot works or arrives at a given output. The Surrogate’s Court of Saratoga County, New York, presided over by Surrogate Jonathan G. Schopf, independently ran Ranson’s query on three separate court-system computers and received three different numerical results—none of which matched Ranson’s figure. Applying the Frye standard, the court found the record “devoid of any evidence as to the reliability of Microsoft Copilot,” rejected Ranson’s assertion that AI-assisted analysis was “generally accepted,” and held that courts “cannot blindly accept as accurate, calculations which are performed by artificial intelligence.” The court characterized its holding as potentially an issue of first impression in Surrogate’s Court practice and held that, before AI-generated evidence is introduced, counsel has an affirmative duty to disclose the use of AI and the evidence should be subject to a Frye hearing.

Together, these decisions underscore four practical points: (1) unverified gen AI outputs can be fatal to expert admissibility, regardless of intent; (2) experts must be able to explain their methodology independently of gen AI tools; (3) attorneys may have an affirmative duty to inquire about and disclose an expert’s gen AI use; and (4) AI-generated content will face the same (and potentially heightened) reliability scrutiny as any other expert methodology.

Best Practices for AI Provisions in Expert Engagement Agreements

Drawing on Kohls and Weber, together with emerging ABA guidance and court rules, practitioners should consider including the following provisions in expert-retention agreements:

1. Require written disclosure of gen AI use. The engagement agreement should require the expert to disclose in writing, before finalizing any work product, whether and to what extent the expert used a gen AI tool. The disclosure should identify the specific tool and version, the tasks for which it was used, and the steps taken to verify its outputs.

2. Require independent verification of every gen AI output. Experts must independently verify all gen AI-generated content—including citations, calculations, and factual assertions—through reliable, independent means such as primary sources, authoritative databases, or the expert’s own calculations. Verification using the same gen AI tool is insufficient.

3. Preserve the expert’s independent methodology and judgment. Gen AI tools must supplement—not replace—the expert’s independent professional judgment. The expert must author all opinions and conclusions and be able to explain the methodology without relying on the gen AI tool’s output.

4. Prohibit inputting confidential or privileged material into public gen AI tools. The agreement should prohibit the expert from submitting any confidential information, privileged materials, attorney work product, or information subject to a protective order into any gen AI tool unless the tool is a closed system that meets specified contractual and technical safeguards and the use is otherwise permitted by the governing protective order.

5. Build in compliance with evolving local rules and standing orders. Include current and forward-looking compliance obligations requiring the expert to comply with any court order, local rule, or standing order governing gen AI use during the engagement, including any requirements adopted after the agreement is signed. Standing orders increasingly require disclosure with each filing and identification of the specific gen AI programs used.

6. Protect privilege and work product through due diligence and operational safeguards. In practice, attorneys rarely can negotiate bespoke vendor terms with gen AI providers. The focus instead should be on vetting and selecting closed-system or enterprise-tier gen AI tools whose publicly available terms of service and privacy policies prohibit using user inputs for training and disclosing them to third parties. Operational safeguards are equally important: require the expert to avoid entering privileged communications or sensitive case strategy into any gen AI tool, use internal or enterprise deployments when available, and anonymize inputs to the maximum extent practicable. In high-stakes or document-intensive matters where the added procedural burden is justified, counsel may also consider seeking an FRE 502(d) clawback order as an additional safeguard against inadvertent waiver, though this step is not warranted as a routine measure for every engagement.

7. Require recordkeeping and a full audit trail. The expert should maintain records of all prompts, inputs, tool versions, and outputs associated with gen AI use. This audit trail is critical to defending the expert’s methodology in Daubert or Frye challenges and to demonstrating compliance with court disclosure requirements.

8. Make informed consent specific, not boilerplate. Consistent with ABA guidance, the agreement should specify why the tool is being used, what information will be input, and the benefits, risks, and limitations of the tool. The ABA Task Force on Law & AI has cautioned that “merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use [AI] is not sufficient.”

9. Attach meaningful consequences for breach. The agreement should give the retaining party clear remedies for noncompliance, including the right to require correction at the expert’s expense, withdraw reliance on affected work product, adjust fees, or terminate the engagement.

Proposed Contract Language

The following model provisions build on the confidentiality, subpoena-notice, protective-order, and document-return provisions found in most expert retainer agreements.

Disclosure Obligations. The following provision requires the expert to provide detailed written disclosure of gen AI use before finalizing any work product:

The Expert shall disclose in writing to the Retaining Party, before the finalization of any Gen AI-Assisted Work Product, whether and to what extent any Gen AI Tool was used in forming opinions, conducting analysis, performing calculations, identifying or verifying sources, or preparing any Work Product in connection with the Engagement. Such disclosure shall include, at a minimum: (a) the identity and version of each Gen AI Tool used; (b) a description of the specific tasks for which the Gen AI Tool was employed; (c) the nature and extent of the Gen AI Tool’s contribution to the Work Product; (d) the steps taken by the Expert to independently verify any Gen AI-generated output; and (e) any limitations, caveats, or known risks associated with the Gen AI Tool’s output as relevant to the opinions expressed.

Supplemental Use Only / Prohibited Uses. The following provision ensures AI tools remain supplemental to the expert’s independent judgment and identifies categories of work requiring prior authorization:

The Expert may use Gen AI Tools only as supplemental aids in the performance of the Engagement. Gen AI Tools shall not substitute for the Expert’s independent methodology, professional judgment, or core analytical work. The Expert must be the author of all opinions and conclusions expressed in any Work Product. The Expert shall not use any Gen AI Tool for the following purposes without prior written authorization from the Retaining Party: (a) forming ultimate opinions or conclusions on any issue within the scope of the Engagement; (b) generating the substantive analytical framework or methodology for any analysis or opinion; or (c) producing content that the Expert cannot independently explain, reproduce in material substance, or validate through the Expert’s own expertise.

Confidentiality and Closed-System Requirements. The following provisions are designed to restrict the use of gen AI tools to closed systems and prohibit the input of confidential or privileged materials into public platforms:

The Expert shall not input, upload, submit, or otherwise make available any Confidential Information, attorney work product, or information subject to any protective order into any Gen AI Tool that is not a Closed System meeting the requirements of this Agreement. Before inputting any information related to the Matter into any Gen AI Tool, the Expert shall anonymize or de-identify such information to the maximum extent practicable. Unless expressly authorized in writing by the Retaining Party, the Expert shall not use any consumer-facing, publicly available Gen AI Tool (including free-tier or personal-account versions of any Gen AI platform) for any purpose in connection with the Engagement.

A Closed System means a gen AI tool or deployment that operates within a secure, access-controlled environment under contractual terms that generally: (a) prohibit the retention, storage, or use of user inputs for model training or improvement; (b) prohibit sharing of user inputs with third parties; (c) keep inputs within a defined security perimeter; and (d) provide adequate confidentiality and data security protections consistent with the obligations of this Engagement.

Verification Obligations. The following provision imposes an independent verification duty that directly addresses the failures identified in Kohls and Weber:

The Expert shall independently verify the accuracy, completeness, and reliability of all content generated or assisted by any Gen AI Tool before incorporating such content into any Work Product. Verification shall be performed through reliable, independent means—i.e., primary sources, authoritative databases, original documents, the Expert’s own calculations, or other methods independent of the Gen AI Tool that generated the content. The Expert shall not rely solely on the same Gen AI Tool or a different query to the same Gen AI Tool to verify Gen AI-generated output.

Remedies for Breach. The following provision gives the retaining party meaningful remedies for noncompliance:

In the event the Expert breaches any provision of this Agreement governing the use of AI Tools, the Retaining Party shall have the following remedies: (a) correction and resubmission—require the Expert to correct, revise, or resubmit any affected Work Product at the Expert’s sole expense; (b) withdrawal of reliance—decline to submit, file, or otherwise rely upon any affected Work Product; (c) termination—terminate the Engagement immediately upon written notice; (d) fee adjustment—withhold, reduce, or seek reimbursement of fees associated with affected Work Product; (e) indemnification—the Expert shall indemnify and hold harmless the Retaining Party from costs, sanctions, or adverse consequences arising from breach.

What You Should Be Doing Now

Given the rapid development of this area, practitioners should take the following steps now:

Update your standard expert engagement agreement. Every expert-retention agreement should include AI-specific provisions addressing disclosure, verification, confidentiality, permitted uses, recordkeeping, and remedies for breach. Do not wait for a court order or adverse ruling to address these issues.

Ask the question at the outset. Before retaining an expert, ask specifically whether and how the expert intends to use gen AI tools, and document the response. As the Kohls court observed, Rule 11 may now require attorneys to ask witnesses about gen AI use in preparing declarations.

Verify before finalization. Before the report is finalized and submitted, confirm with the expert whether gen AI was in fact used, which tools and versions were employed, what tasks they supported, and whether every output was independently verified.

Require closed-system AI tools for any work involving confidential materials. Experts working with materials subject to protective orders, trade secrets, source code, or other sensitive information should be contractually prohibited from using consumer-facing gen AI platforms. Require the use of enterprise-tier tools with, to the extent possible, contractual protections against data retention and training.

Maintain an audit trail. Require experts to preserve prompts, inputs, outputs, and tool versions. This documentation is essential to defending the expert’s methodology in Daubert or Frye challenges and to complying with emerging court disclosure requirements.

Monitor evolving court rules and standing orders. Courts are rapidly adopting standing orders and local rules requiring gen AI disclosure. Engagement agreements should include current and forward-looking compliance obligations, and counsel should regularly check applicable local rules and orders for new AI-related requirements.

Treat these provisions as a floor, not a ceiling. The model provisions above are designed for broad applicability. In cases involving heightened confidentiality concerns—such as IP litigation involving trade secrets or source code—counsel should consider additional restrictions, including express prohibitions on gen AI use for specific categories of work or mandatory pre-approval for any AI tool use.

Educate your experts. Many experts may not appreciate the legal consequences of gen AI use in litigation contexts. The engagement letter itself serves an educational function, but practitioners should also discuss gen AI expectations at the outset of every retention and confirm the expert’s understanding in writing. 

We will continue to monitor evolving case law and court guidance regarding expert use of generative AI tools. Please contact us with questions about how these decisions may affect your expert-retention practices or litigation strategy.