Important Posts – E-Discovery LLC https://www.ediscoveryllc.com Mediation of E-Discovery Disputes Sat, 31 Oct 2026 18:49:59 +0000 en-US hourly 1 Motion to Permit Limited A.I. Review of “Confidential” Video Was Denied https://www.ediscoveryllc.com/motion-to-permit-limited-a-i-review-of-confidential-video-was-denied/ Sat, 31 Oct 2026 08:59:06 +0000 https://www.ediscoveryllc.com/?p=5724 […]]]> AI Protective Orders address how or if a receiving party can use artificial intelligence to analyze discovery materials produced by an opponent.  Some address all discovery materials.  Others are limited to confidential discovery materials.  Some specify permitted and prohibited platforms by name.  Others specify functionality.  A few do both.  Some are a page long; others are only a paragraph.  Most are entered by agreement of the parties.  However, there are a few cases in which a request for an AI Protective Order is contested.  This blog focuses on one of the contested cases and discusses several other related issues.

Plaintiff filed a Motion to Modify Protective Order to Permit Limited Artificial Intelligence (“AI”)-Assisted Review of Body-Worn Camera (“BWC”) evidence. The motion was denied in Baqleh v. Altamirano, 2026 WL 3257576 (N.D. Cal. Oct. 28, 2026).

After plaintiff sued for claims arising out of his arrest, “the Parties stipulated to a Protective Order (‘PO’) that mirrored the Court’s model protective order, which the Court granted.”

However, “Plaintiff now moves the Court to modify the PO to allow Plaintiff to use ChatGPT to review body-worn camera footage Defendant produced in discovery and marked as ‘CONFIDENTIAL.’” The court explained that:

Plaintiff seeks to use ChatGPT to create timestamps, prepare a chronology, transcribe the video, and identify events within the video for purposes of discovery, motion practice, settlement, and trial preparation, but says he will not prepare analysis to use as admissible evidence.

After Defendant replied, Plaintiff narrowed his request and “now seeks AI-assisted review through a ‘business or private AI service’ with ‘safeguards required by the Court.’”  In my words, Plaintiff’s modification proposed to use a secure, enterprise, or fiduciary grade A.I. provider instead of a public, consumer-grade system.  However, the Plaintiff asked the court to set those safeguards.

The Baqleh court wrote: “Some courts have recently permitted the use of AI-assisted review when parties have agreed to use AI during discovery to analyze large quantities of data….  In this instance however, Plaintiff is not seeking to use AI for that purpose. Instead, Plaintiff asks the Court to permit disclosure of confidential information to a third-party within the structure of the existing PO.”

The Baqleh court held that, balancing the need for discovery against the need for confidentiality, Plaintiff had not shown good cause.

The Plaintiff’s argument was:

Plaintiff believes the Court should permit him to use ChatGPT to review confidential evidence because he is pro se and does not have support “reviewing and organizing audiovisual discovery.”…

Defendant countered:

Defendant responds that Plaintiff does not need AI assistance because the BWC video evidence is only 30 minutes long, and Plaintiff does not contest this representation about the length of the footage…. Defendant argues that, aside from being unnecessary, there is substantial risk in using a public large-language model platform to process confidential information…. Defendant represents that uploading confidential evidence to ChatGPT risks disclosure of information such as non-party criminal history, parole status, probation status, drug use, police investigations, and driver’s license numbers, all of which Plaintiff acknowledges to be “sensitive information.”

Id. at *2.  The court explained its reasoning:

The Court finds that the balance of conflicting interests weighs in favor of Defendant’s position. It is not disputed that exposing confidential information to ChatGPT creates at least some risks with respect to how that information will be used and safeguarded…. While Plaintiff proposes several potential safeguards to prevent disclosure of the information Defendant identifies, he does not cite anything confirming the availability or reliability of those safeguards…. So, there would need to be some substantial offsetting benefit that outweighs these risks, but Plaintiff has not shown that specific harm will result if his motion is not granted. Reviewing a 30-minute video for the purposes Plaintiff says he would use ChatGPT for is not an overbearing task that disadvantages a pro se party. For example, to the extent Plaintiff wants to create a chronology of the video, he can easily do so by taking handwritten notes. The Court thus denies Plaintiff’s request to use ChatGPT to review the confidential BWC evidence.

Id. at *2 (emphasis added).

In a footnote, the court addressed and rejected Plaintiff’s offer to narrow the request:

Plaintiff’s Reply appears to narrow his request by asking for “AI-assisted review through a business or private AI service” where he can show that safeguards exist…. Plaintiff’s new formulation does not change the Court’s analysis, because the balance between (some) risks and (essentially no) benefits remains unchanged. Plaintiff also requests in the alternative for the Court to modify the PO using the framework outlined in Morgan v. V2X, Inc., 2026 WL 864223 (D. Colo. Mar. 30, 2026), to accommodate Plaintiff’s request to use AI to review the video footage…. The Court declines in its discretion to do so for the reasons it has already explained.

Baqleh may be a unicorn.  As the court pointed out, it was for a very limited purpose and not for review of voluminous ESI.  The video was “confidential” and short, unlike the voluminous discovery materials in many other cases. And, Plaintiff asked the court to do his “homework” for him by specifying the safeguards that the court would impose instead of offering a specific plan.

A.I. protective orders are one of the hottest recent topics in civil and criminal litigation. I have suggested that they are now the standard of care.

I have also echoed Craig Ball’s concerns that equity be considered and that they not become an obstacle to unrepresented or indigent litigants. See  A.I. Protective Orders Are Becoming Routine (Aug. 27, 2026), and  Another A.I. Protective Order – Part 10 (Aug. 21, 2026), both citing Craig Ball, The AI Protective Order Double Standard | Ball in your Court (Jul. 27, 2026).

Craig’s post also tackles the thorny issue of terminology.  He suggests specific terms that should be avoided in these protective orders. His list of five things that a properly-scoped order should contain is well worth reading. The AI Protective Order Double Standard | Ball in your Court.

For more information, please visit:

Restrictions on Use of Artificial Intelligence Tools. No party, non-party, or other person permitted to receive information designated “Confidential” [or “Confidential—Attorneys’ Eyes Only”] under Paragraph 6 of this Order may input, upload, or submit such information to any artificial intelligence tool, chatbot, generative AI system, large language model, or similar automated platform (“AI”) unless the AI provider is contractually prohibited from both:

(a) using the inputs, or information derived from them, for any purpose other than providing the specific service requested, including using them to train any model or to develop or improve any product or service; and

(b) disclosing the inputs, or any information derived from them to any third party, except where such disclosure is essential to delivering the requested service. Where disclosure to a third party is essential to delivering the requested service, that third party must be bound by obligations no less protective than those required by this paragraph.

Any party seeking to use AI under this paragraph must also confirm that the AI provider will, upon request, delete from its systems all information designated “Confidential.” The party must retain written documentation of the contractual protections required by this paragraph and provide that documentation to the Designating Party or the Court upon request….

Protected Material shall not be submitted to any “Open Generative AI Tool” (e.g., ChatGPT) or any substantially similar tool that is available to or accessible by the public, with or without a subscription or license. Providing Protected Material to any “Open Generative AI Tool” is considered a prohibited disclosure to a third party. The parties may use a closed Generative AI Tool that is private and secured, where there is a contractual agreement providing that the generative AI tool does not retain information beyond any use session, except to the extent retention is reasonably necessary for investigation of abuse violations, system functionality or internal data security operations, provided such information is not used for model training or product improvement purposes and remains subject to the provider’s confidentiality obligations. This provision is not intended to limit use of technology assisted review via qualified data storage persons (i.e., electronic discovery or litigation-support vendors). Any party receiving Protected Material must ensure that the contract, software license, terms of use, or any other agreement or terms for use of the AI tool complies with this Protective Order. [emphasis added].

]]> Contested Motion for AI Protective Order Was Granted https://www.ediscoveryllc.com/contested-motion-for-ai-protective-order-was-granted/ Mon, 26 Oct 2026 08:37:10 +0000 https://www.ediscoveryllc.com/?p=5703 […]]]> “This is an action by Fischman, proceeding pro se against Epic, asserting claims arising from the alleged role of Epic’s medical software in the death of her father. Fischman has served extensive discovery requests that seek, among other things, the software’s design, development, and implementation.”  Fischman v. Epic Systems, Corp. 2026 WL 3182476 (N.D. Tex. Oct. 23, 2026).

The parties differing positions were:

  • “Epic now moves for a protective order, contending that documents responsive to these requests contain confidential information and trade secrets.”
  • “In response to Epic’s motion, Fischman initially agreed to the entry of some form of a protective order. She now posits, however, that her ‘preferred course is for discovery to proceed … without a separate protective order.’”

The court wrote that it deemed Epic’s request to be opposed.  It held that Epic made the factual showing required by Fed.R.Civ.P. 26(c).

As to AI, Epic’s proposal had two features:

  • “It prevents the input of confidential and highly confidential information into an AI tool unless that tool is contractually prohibited from … storing or using the inputs to train or improve any model and prohibited from disclosing the inputs to any third party.” [cleaned up]; and,
  • It “shall not be construed to limit the use of artificial intelligence for legal research, drafting, organization, or other litigation tasks that do not entail the input of Protected Material.” [cleaned up].

The Fischman court wrote:

As to the provisions restricting the use of AI, Epic maintains that the entry of information into generic AI tools is “effectively irreversible” because these models “commonly retain the text and files that users submit and use those submissions to train and improve the … models.” …. Epic also cites ample authorities that show that courts across the country impose protective orders tailored to address the problems that modern AI tools pose. See, e.g., Morgan v. V2X, Inc., 2026 WL 864223, at *7 (D. Colo. Mar. 30, 2026) (imposing nearly the same restriction at issue here). Given Fischman’s acknowledged use of AI in this litigation, … the court finds that there is good cause to impose Epic’s proposed AI restrictions in the protective order.

The posts A Unique A.I. Protective Order (Oct. 9, 2026), and Three More A.I. Protective Orders (Oct 9, 2026), provide multiple additional examples

]]> Subpoenas to AI Service Providers Were Partially Approved, With Guardrails https://www.ediscoveryllc.com/subpoenas-to-ai-service-providers-were-partially-approved-with-guardrails/ Thu, 22 Oct 2026 16:14:57 +0000 https://www.ediscoveryllc.com/?p=5694 […]]]> In Fontainebleau, Plaintiff served four subpoenas on AI providers and requested A.I.  information in an RFP.  The court “directed the Parties to submit expedited supplemental briefing addressing whether the requested AI-related information falls within the scope of damages discovery authorized by the District Court, including its relationship, if any, to Defendant’s alleged malice and the statutory limitation on punitive damages under Florida law.”  Fontainebleau Florida Hotel, LLC v Shmuley Boteach Jacob, 2026 WL 3151630 (S.D. Fla. Oct.21, 2026).  It also applied the work product doctrine.

Defendant raised “blanket objections” to the AI subpoenas. “The Court emphasizes that a party’s communications with and use of AI does not automatically establish either discoverability or protection.”  Id. at *10.  After analyzing relevance, the court held that each piece of information should be analyzed under a traditional work product framework.  It wrote: “The Court must examine the purpose and circumstances of the particular communications rather than presume protection from the medium through which they were created.”

In sum: “The Court emphasizes that a party’s communications with and use of AI does not automatically establish either discoverability or protection. The analysis turns on the ordinary requirements of Rules 26(b)(1), 26(b)(3), and 26(b)(5).”

OVERVIEW

Plaintiff’s subpoenas sought all communications with AI: concerning the hotel, relating to any hotel employee, relating to the litigation, concerning Plaintiff’s counsel, relating to any Court Order in the litigation, or relating to the Judges.  The RFP sought: “All of Defendant’s AI Communications regarding Requests 1 through 7 that have, since their creation, been deleted.”

Much of the decision turns on whether the discovery sought information beyond the scope authorized by the court, which was limited to damages.  A related issue was whether this was liability discovery that was propounded after the discovery cutoff.

An important issue turned on work product under Heppner, Warner, and their progeny.

Defendant argues that the requests in the subpoenas and the Seventh RFP risk exposing materials protected by the work-product doctrine, including AI-assisted litigation preparation, legal research, draft analyses, and communications concerning defense strategy…. Defendant relies principally on Warner v. Gilbarco, Inc., 820 F. Supp. 3d 629 (E.D. Mich. 2026), for the proposition that litigation-related materials do not automatically lose work-product protection merely because a party used a third-party AI platform to prepare them.

SOME OF THE A.I. COMMUNICATIONS WERE RELEVANT

Plaintiff argued that the requested information was relevant to punitive damages.  As a matter of substantive, tort law, the court ruled that  “when a defendant’s state of mind bears on a legally cognizable issue, the mere fact that discovery probes the defendant’s thinking, motives, or preparation does not render it impermissible.” It then held:

The same principle applies to Defendant’s AI prompts and communications. To the extent Defendant used AI platforms to generate, revise, or refine statements concerning Plaintiff, its counsel, or its representatives, those communications may reveal not merely the final content of Defendant’s statements but also the purpose for which the statements were created, the instructions Defendant supplied, and the effects he sought to achieve. Such information may provide evidence of the intent underlying Defendant’s conduct that cannot necessarily be discerned from the resulting public statements alone. The Court does not determine that those communications will establish malice or specific intent to harm; it determines only that they bear a sufficient relationship to those disputed damages issues to warrant discovery.

The court wrote: “Here, Plaintiff alleges a continuing course of conduct directed toward the Hotel, its personnel, and its representatives, including conduct involving AI-generated content.”  It then added: “That ruling, of course, does not establish that every AI communication is discoverable, but it reinforces the relevance of evidence concerning Defendant’s alleged continuing conduct toward Plaintiff and persons associated with it.”  Id. at *5.

In short: “Defendant’s communications with AI platforms concerning the Hotel and the Hotel’s employees, owners, or representatives, including Plaintiff’s counsel, may reveal information relevant to Defendant’s alleged intent to harm Plaintiff, the nature and persistence of the challenged conduct, and the extent to which that conduct allegedly amplified Plaintiff’s injuries. The Court therefore concludes that Request Nos. 1 (seeking AI communications about the Hotel), 3 (seeking AI communications about Plaintiff’s counsel), and 4 (seeking AI communications about any employee, owner, or representative of the Hotel) in the subpoenas to the AI providers and the Seventh RFP fall within the scope of damages discovery authorized by the District Court. Accordingly, Defendant’s blanket scope objection to Request Nos. 1, 3, and 4 is OVERRULED.”

SOME OF THE A.I. COMMUNICATIONS WERE BEYOND THE SCOPE OF DISCOVERY

Discovery had been re-opened for a limited purpose. Id. at *8. “The Court agrees with Defendant’s general premise that an order reopening discovery for a limited purpose does not authorize unrestricted additional discovery. The District Court expressly limited the reopening to damages discovery, and that limitation must be enforced.”

“Although the Court finds that certain AI communications are discoverable, it agrees with Defendant that the requests in the subpoenas and Seventh RFP, as drafted, are broader than the damages discovery authorized by the District Court.”  Id. at *7.

For example, several were “exceptionally broad.” One “encompasses potentially every AI exchange on matters having no meaningful connection to the amount of damages or Defendant’s alleged intent to harm Plaintiff. Plaintiff has not established that every communication relating to this litigation bears on an issue remaining for the damages trial.”  Id.

As to another:

Turning to Request No. 5 in the subpoenas and the Seventh RFP, it seeks all AI communications concerning or relating to any Court Order entered during this litigation.... Such communications may include Defendant’s attempts to understand an Order, research its consequences, consider procedural options, or evaluate the litigation. The fact that Defendant may have expressed disagreement with, frustration about, or hostility toward a Court Order does not, without a sufficient connection to the alleged injury, make the communication relevant to Plaintiff’s damages. Although Plaintiff contends that Defendant’s reactions to Court Orders reflect the depth of his hostility and the reprehensibility of his conduct, that theory is too attenuated to justify discovery of every AI exchange concerning every Order entered in this case. The Court therefore finds that Request No. 5 exceeds the permissible scope of damages discovery.

Id. at *7.  “The Court likewise finds that Defendant’s private communications with AI platforms concerning judicial officers, without more, do not bear a sufficient relationship to the damages Plaintiff seeks for the alleged harm to the Hotel.”  Id.

THE WORK-PRODUCT DOCTRINE

DOES NOT CATEGORICALLY PROTECT DEFENDANT’S AI COMMUNICATIONS

The court wrote: “Defendant also contends that the requested AI communications may constitute protected work product because they include AI-assisted litigation preparation, research, draft analyses, and other materials concerning the defense of this action.”  Id. at *8.

It noted that “Defendant invokes work-product protection over a potentially diverse collection of AI communications without establishing that every communication within the permitted categories was created because of this litigation.”  Id.

Heppner & Warner

To set the stage, the Fontainebleau court wrote: “The Parties rely on two recent decisions addressing AI-assisted materials: United States v. Heppner, 820 F. Supp. 3d 292 (S.D.N.Y. 2026) and Warner v. Gilbarco, Inc., 820 F. Supp. 3d 629 (E.D. Mich. 2026).”

  • “The Court finds Heppner persuasive insofar as it recognizes that independently generated AI communications do not automatically become protected work product merely because counsel represents the user. At the same time, the Court does not read Heppner as establishing that all materials independently prepared by a represented party are categorically unprotected. This is because Rule 26(b)(3)(A) expressly extends to qualifying materials prepared ‘by or for another party or its representative.’ Fed. R. Civ. P. 26(b)(3)(A).”
  • “In Warner, the court recognized that AI-assisted litigation preparation may qualify for work-product protection and that using a third-party AI tool does not necessarily result in waiver. 820 F. Supp. 3d at 636–37. But Warner does not support Defendant’s blanket objection. The recognition that qualifying AI-assisted litigation preparation may remain protected notwithstanding the use of an AI platform does not establish that all communications with AI concerning an opposing party, its personnel, or its counsel were prepared in anticipation of litigation. The Court must examine the purpose and circumstances of the particular communications rather than presume protection from the medium through which they were created.”

The Fontainebleau Analysis

The Fontainebleau court wrote:

Neither decision establishes a categorical rule governing all AI communications. Rather, they underscore that the applicability of work-product protection depends on the nature, purpose, and circumstances of the particular materials. Here, communications independently generated by Defendant to create, revise, or disseminate content concerning the Hotel, its personnel, or Plaintiff’s counsel are not transformed into attorney work product merely because Defendant is a litigant or because the resulting content may bear on this action. To the extent such communications were created for the purpose of producing public-facing statements or pursuing Defendant’s independent activities, rather than preparing the defense of this litigation, they ordinarily would not satisfy Rule 26(b)(3)(A). The Court therefore rejects Defendant’s blanket work-product objection to Request Nos. 1, 3, 4, and 7. However, that conclusion does not foreclose a properly supported, document-specific claim of protection. If a particular communication was prepared because of this litigation and otherwise satisfies Rule 26(b)(3), Defendant may assert the protection in accordance with the Federal Rules. Defendant must, however, identify the withheld material and describe the nature of the asserted protection sufficiently to permit Plaintiff to assess the claim without revealing the protected information itself. See Fed. R. Civ. P. 26(b)(5)(A).

The Court emphasizes that a party’s communications with and use of AI does not automatically establish either discoverability or protection. The analysis turns on the ordinary requirements of Rules 26(b)(1), 26(b)(3), and 26(b)(5). Defendant may not withhold all communications within the permitted categories based on a generalized assertion that they concern litigation, and Plaintiff may not obtain genuinely protected litigation-preparation materials merely by describing them as AI communications.

Id. at *10 (emphasis added).

DEFENDANT WAS NOT COMPELLED TO CONSENT TO THE SUBPOENAS

The court also addressed an oral motion based on an alleged failure to preserve AI communications.  Id. at *10.  Plaintiff contended that third-party subpoenas were proper based on this assertion and Plaintiff asked for an order compelling Defendant to consent to the subpoenas.

The court denied the motion: “Although Plaintiff may pursue appropriately limited subpoenas, the Court declines to compel Defendant to execute consents or authorizations permitting the AI providers to disclose responsive communications. Rule 34 permits a party to request documents and electronically stored information within another party’s possession, custody, or control. See Fed. R. Civ. P. 34(a)(1). The Rule requires a responding party to produce responsive, nonprivileged materials within that party’s control, but it does not expressly require a party to execute an authorization allowing a third party to disclose records directly to an adversary.”

The court wrote that no rule “establishes that a litigant must execute a blanket authorization permitting a third-party service provider to disclose the litigant’s private records merely because an opposing party seeks relevant information. And Plaintiff cites to no authorities to support its request.”

The court stated two caveats:

  • “[T]he Court will not compel Defendant to execute prospective authorizations based solely on the possibility that such consent might facilitate Plaintiff’s third-party discovery.’
  • “The Court further emphasizes that permitting the issuance of subpoenas does not establish their enforceability. Rule 45 provides procedures for objections and motions to quash or modify subpoenas by nonparties. Accordingly, the Court’s ruling preserves Plaintiff’s ability to seek authorized third-party discovery consistent with this Order without prematurely adjudicating the obligations of nonparties that have not appeared before the Court.” [emphasis added].

PARTIES MUST DISCUSS LOGISTICS

Lastly, the court added:

But there is one additional wrinkle. The Court recognizes that Defendant may, consistent with this Order, assert that certain responsive AI communications within the possession of the AI providers are subject to work-product protection. And if Plaintiff serves the subpoenas on the AI providers and requires them to produce the document directly to Plaintiff’s counsel, Defendant will be deprived of the opportunity to determine whether there are any work-product protected documents within the production that should otherwise be withheld and identified on a privilege log. For that reason, the Parties are DIRECTED to meet and confer to discuss the logistics of how the AI providers, to the extent they have responsive materials, will be instructed to produce such materials. If the Parties cannot agree on a procedure, the Court will hear the Parties’ respective positions on this issue at the follow-up discovery hearing that is tentatively set for October 26, 2026 at 1 p.m.

Id. at *12 (emphasis added).

Resolving that wrinkle should not be a problem. There is plenty of guidance on that issue.  Maryland Appellate Decision Permitting Limited Forensic Search of Non-Parties’ Cell Phones (Mar. 24, 2022).

For more on Heppner, Warner, and their progeny, please visit:

]]> When a Redacted Document Was Used to Prepare a Witness, Only the Redacted Document Needs to Be Produced in Discovery https://www.ediscoveryllc.com/when-a-redacted-document-was-used-to-prepare-a-witness-only-the-redacted-document-needs-to-be-produced-in-discovery/ Tue, 20 Oct 2026 16:56:07 +0000 https://www.ediscoveryllc.com/?p=5676 […]]]> Doug Austin and Michael Berman

Plaintiff Mr. Corkery filed an emergency motion to amend or remove redactions to a document that he planned to use in an upcoming deposition. He also sought a privilege log for the redactions, “if any applies.”

Defendant Coinstar objected.  At issue was “whether Coinstar’s production of approximately 2,000 pages of redacted board meeting minutes was improper.”   Corkery v. Coinstar, LLC, 2026 WL 3001353 (D.R.I. Oct. 6, 2026).

The redaction of 2,000 out of 2,600 pages took place in a unique procedural posture.

Coinstar explains that these redacted documents were produced not in response to a discovery request but rather voluntarily out of fairness under Federal Rule of Evidence 612(b) because they were reviewed by a witness in preparation for his deposition testimony.

Id. at *2 (emphasis added).  Further, Coinstar added that the “witness reviewed the documents in redacted form while preparing for the deposition….”  Id. at *2. Coinstar added that no privilege log was required because it was not withholding the documents based on privilege. Id.

The Corkery court wrote:

Rule 612 of the Federal Rules of Evidence pertains to writings that are used to refresh the memory of a witness either while or before testifying. Rule 612(b) provides that “an adverse party is entitled to have the writing[.]” Here, the writing used to prepare the witness to testify was provided to the opposing party in the same redacted form that the witness reviewed it in preparation for his testimony. Mr. Corkery has not provided any authority, nor has the Court found any, indicating that an unredacted document must be provided when a witness prepared with a redacted document.

Id. at *3 (emphasis added).

As to the privilege log: “Here, the document was provided in the same format it was used to prepare a witness, with redactions. There is nothing before the Court to indicate that any of those redactions were for privilege, nor has Mr. Corkery provided any authority indicating that if documents provided pursuant to Rule 612(b) of the Federal Rules of Evidence are redacted for privilege, then a log must be produced.”

Corkery addresses a unicorn. Redactions for relevance are generally prohibited or disfavored. One exception, discussed in Episode 110: When are Relevancy Redactions Allowed in Responsive Documents? | Minerva26, is Kaiser Aluminum Warrick, LLC v. U.S Magnesium, LLC., No. 22-CV-3105 (JGK) (KHP) (S.D.N.Y. Feb. 27, 2023), where New York Magistrate Judge Katherine H. Parker, stated: “Although many courts do not permit relevancy redactions, this Court believes relevancy redactions can be appropriate in some cases.” Judge Parker did permit the defendant to retain redactions on “information irrelevant to this action”. However, she also wrote: “the parties shall seek the leave of the opposing party or the Court before redacting any future documents.”

Including the Kaiser Aluminum Warrick case, here are other cases involving redaction for relevancy:

This blog was initially posted on  Electronic Discovery Reference Model and  JD Supra.

]]> Examples of a Very Detailed A.I. Protective Order and Less Detailed Agreements and Orders https://www.ediscoveryllc.com/examples-of-a-very-detailed-a-i-protective-order-and-less-detailed-agreements-and-orders/ Fri, 16 Oct 2026 22:38:06 +0000 https://www.ediscoveryllc.com/?p=5656 […]]]> I have suggested that AI protective orders are becoming the standard of care.[1]  These orders limit the use or uses of AI by a recipient of discovery materials to review the materials produced by an opponent.  They clearly are important and have value in appropriate instances.

However, there is no standard form and the agreements vary greatly. Some cover only confidential information; others apply to all discovery material. Some list acceptable platforms by name; others describe functional requirements; and, some address named platforms and their configuration.  There has been discussion about the terms “consumer grade” AI v. “Enterprise” AI.  Thomsen-Reuters discusses Fiduciary-Grade AI.  The terms “open” and “closed” AI have been used–and criticized.

All protective orders are not created equal and they are not always fair.  I have echoed Craig Ball’s novel post on the need for equity in limiting an opponent’s right to use AI.  Another A.I. Protective Order – Part 10 quoting Craig Ball, The AI Protective Order Double Standard | Ball in your Court (Jul. 27, 2026).

One of the most detailed AI protective orders that I have seen was stipulated in Corent Technology, Inc. v. Amazon Web Services, Inc., 2026 WL 2924482 (C.D. Cal. Sep. 28, 2026).  It deals with both the use of AI on source code and more generally.  At the end of this post, I cite a number of less detailed examples.

As to source code, the Corent Technology order states:

Neither party shall use any software tools or applications on the Source Code Computer powered by or leveraging artificial intelligence or machine learning that automatically scan, analyze, or provide feedback on source code. The Producing Party shall accommodate reasonable requests by the Receiving Party that other mutually agreeable commercially available software tools for viewing and searching Source Code be installed on the Source Code Computer, provided that such other software tools are reasonably available to the Producing Party, non-destructive to the source code, and reasonably necessary for the Receiving Party to perform its review of the Source Code consistent with all the protections herein. The Receiving Party must provide the Producing Party with the licensed software tools at least three (3) business days in advance of the date upon which the Receiving Party wishes to have the additional software tools available for use on the Source Code Computer.

As to A.I. in general, the Corent Technology order states:

USE OF AI TOOLS

No party, and no person or entity authorized to receive materials under this Order, may input, upload, submit, or otherwise provide any Discovery Material—whether or not designated as Protected Information—into any modern artificial intelligence platform, including any generative, analytical, or large language model-based tool (“AI Tool”), unless all requirements of Section 12.1 and Section 12.2 are satisfied.

12.1 A party or person subject to this Order may use an AI Tool with Discovery Material only if the AI Tool is provided on an enterprise or commercial tier subscription and the Receiving Party has not configured the AI Tool to:

a) retain or store inputs (including prompts, documents, and metadata) for use beyond the session necessary to deliver the requested service, except as provided in Section 12.7.2, For clarity, this requirement does not prohibit the use of AI Tools that maintain a history of a user’s prompts and/or sessions, provided that any prompts containing Protected Information shall not be disclosed to any person not authorized to receive Protected Information under this Order;

b) use, incorporate, or reference inputs to train, fine-tune, retrain, or otherwise improve any AI model or system, whether that of the vendor or any third party, except as provided in Section 12.7.2, For clarity, this requirement does not prohibit the use of AI tools that incorporate or reference a user’s past prompts to optimize responses to subsequent prompts by the same user, provided that such AI tools (1) do not use, reference, or incorporate prompts or other inputs received from one user to respond to prompts from any other user; and (2) otherwise comply with all of the requirements in Section 12;

c) disclose inputs to any third party, except where strictly essential to facilitate delivery of the service; any such third party must be bound by confidentiality obligations no less protective than those required by this Order;

d) allow any human employee, contractor, or reviewer of the AI vendor to read, access, or review input content for any purpose other than immediate technical support explicitly requested by the party; and

e) make inputs or retained outputs accessible to any person not authorized to receive Protected Information under this Order.

12.2 In addition to the requirements of Section 12.1, a party or person may use an AI Tool with Discovery Material only if the AI vendor is contractually obligated to afford the Receiving Party the right to remove or delete all Discovery Material and AI-generated outputs from the vendor’s systems upon request, and in any event no later than sixty (60) days following the entry of a final non-appealable judgment or order, or the complete settlement of all claims asserted against all parties in this action (the “Conclusion of This Action”). The Receiving Party shall: (a) within fourteen (14) days following the Conclusion of This Action, take all reasonable steps necessary to ensure deletion of all Protected Material uploaded and/or output by AI tools during this action. The Receiving Party may meet this requirement by submitting a deletion request to the vendor if the vendor accepts deletion requests during the term of its vendor agreement with the Receiving Party; or, if the AI vendor does not accept such requests, by manually deleting all uploaded information over the AI tool’s user interface; and (b) within sixty (60) days following the Conclusion of This Action, provide the Producing Party with written certification and documentation demonstrating that the Receiving Party took all reasonable steps necessary to delete all Discovery Material uploaded to AI tools using the procedures and means provided by the vendor(s), as applicable.

12.3 Any party intending to use an AI Tool with Discovery Material must, prior to such use:

a) retain written documentation (including the vendor’s data processing agreement, terms of service, or equivalent contractual instrument) demonstrating that the vendor’s commitments satisfy Sections 12.1 and 12.2; and

b) make such documentation available to the Producing Party within five (5) business days of a written request.

12.4 A party that intends to use an AI Tool with Discovery Material produced by the other party must provide written notice to the Producing Party at least seven (7) business days before first use, except with respect to AI Tools pre-approved under Section 12.6. Such notice shall identify:

a) the name, version, and operator of the AI Tool;

b) the categories of Discovery Material the Receiving Party intends to process; and

c) a representation that Sections 12.1 and 12.2 are satisfied, accompanied by the documentation required by Section 12.3(a).

The Producing Party shall have five (5) business days to object in writing, stating with particularity the basis for objection. If a timely objection is served, the parties shall meet and confer within three (3) business days. If the dispute is not resolved, the Receiving Party may not proceed with such AI Tool use until the Court resolves the dispute pursuant to Local Rule 37-1 et seq. and Section 3 of Judge McCormick’s Procedures.

12.5 Notwithstanding any other provision of this Section 12, no party may use any AI Tool, regardless of vendor contractual protections, to review, analyze, summarize, or otherwise process any Discovery Material designated as HIGHLY CONFIDENTIAL—SOURCE CODE. This prohibition is absolute and applies whether such Source Code material is accessed on the Source Code Computer, in printed form, in electronic excerpts, or in any other form or medium.

12.6 The parties agree that the following AI Tools, when used at the enterprise or commercial tier identified below and subject to the configuration requirements in this subsection, satisfy the requirements of Sections 12.1 and 12.2 based on each vendor’s published enterprise terms of service, data processing addendum, and security certifications as of the Effective Date of this Order. Use of any AI Tool pre-approved under this Section 12.6 shall not require the pre-use notice otherwise required under Section 12.4.

a) Microsoft Copilot, including Microsoft 365 Copilot and Copilot Cowork, when deployed within a Microsoft 365 Enterprise or Business tenant governed by Microsoft’s enterprise Data Protection Addendum (and not the free or consumer-facing Copilot product available at copilot.microsoft.com), satisfies the requirements of Sections 12.1 and 12.2, provided that: (i) the Receiving Party’s tenant is configured to disable any consumer connectors or web-grounding features that would route Discovery Material outside the enterprise data-processing boundary; and (ii) the Receiving Party has not enabled any consumer-tier or non-Enterprise Copilot product for use with Discovery Material.

b) Lexis+ AI (LexisNexis), at the firm’s standard commercial or enterprise subscription tier, satisfies the requirements of Sections 12.1 and 12.2, provided that: (i) the Receiving Party’s account is configured under LexisNexis’s then-current commercial terms and its published zero-training and zero-retention commitments applicable to user-submitted content; and (ii) the Receiving Party has not enabled any consumer or free-tier LexisNexis AI product for use with Discovery Material.

c) Westlaw AI and CoCounsel (Thomson Reuters), at the firm’s standard commercial or enterprise subscription tier, satisfy the requirements of Sections 12.1 and 12.2, provided that: (i) the Receiving Party’s account is governed by Thomson Reuters’s then-current commercial terms, including its non-retention API arrangements with underlying model providers and its contractual prohibition on third-party model providers retaining or training on customer data; and (ii) the Receiving Party has not enabled any consumer or free-tier Westlaw AI or CoCounsel product for use with Discovery Material.

d) Patlytics (patlytics.ai), at the firm’s standard commercial or enterprise subscription tier, satisfies the requirements of Sections 12.1 and 12.2, provided that: (i) the Receiving Party’s account is configured under Patlytics’s then-current customer terms, including its published non-training commitment and its non-retention agreements with underlying third-party AI services; and (ii) the Receiving Party has not enabled any consumer or free-tier Patlytics product for use with Discovery Material.

e) Perplexity Enterprise (Pro or Max tier) satisfies the requirements of Sections 12.1 and 12.2, based on Perplexity’s Enterprise Terms of Service, Data Processing Addendum, and SOC 2 Type II attestation as of the Effective Date of this Order, provided that: (i) the Receiving Party’s Enterprise instance has Zero Data Retention enabled or, alternatively, has configured the shortest available data retention period; and (ii) the Receiving Party has not enabled any consumer-tier or non-Enterprise Perplexity product for use with Discovery Material.

f) Claude (Anthropic), at the firm’s standard commercial or enterprise subscription tier, satisfies the requirements of Sections 12.1 and 12.2, provided that: (i) the Receiving Party’s subscription is configured to disable any consumer connectors or web-grounding features that would route Discovery Material outside the enterprise data-processing boundary; and (ii) the Receiving Party has not enabled any consumer-tier or non-Enterprise Claude product for use with Discovery Material.

g) Harvey (harvey.ai) satisfies the requirements of Sections 12.1 and 12.2, provided that: (i) the Receiving Party’s subscription is configured to disable any consumer connectors or web-grounding features that would route Discovery Material outside the enterprise data-processing boundary; and (ii) the Receiving Party has not enabled any consumer-tier or non-Enterprise Harvey product for use with Discovery Material.

h) Legora aOS (Legora) satisfies the requirements of Sections 12.1 and 12.2, provided that: (i) the Receiving Party’s subscription is configured to disable any consumer connectors or web-grounding features that would route Discovery Material outside the enterprise data-processing boundary; and (ii) the Receiving Party has not enabled any consumer-tier or non-Enterprise Legora product for use with Discovery Material.

i) Crane Patents satisfies the requirements of Sections 12.1 and 12.2, provided that: (i) the Receiving Party’s subscription is configured to disable any consumer connectors or web-grounding features that would route Discovery Material outside the enterprise data-processing boundary; and (ii) the Receiving Party has not enabled any consumer-tier or non-Enterprise Crane Patents product for use with Discovery Material.

12.7.1 No party or authorized recipient of Discovery Material shall use any Discovery Material, or any AI-generated output derived therefrom, within any AI Tool workspace, project, index, session, or other AI-Tool-hosted environment that is accessed, maintained, or used for any matter other than this action. Each party shall configure its use of AI Tools such that Discovery Material and any AI-generated outputs derived therefrom are logically segregated from unrelated matters, and shall not carry over, migrate, or otherwise make available such Discovery Material or derived outputs to any workspace, project, index, or session unrelated to this action.

12.7.2 For the avoidance of doubt, nothing in this Section 12 prohibits an AI Tool from retaining Discovery Material or AI-generated outputs derived therefrom within a single matter-specific workspace, project, index, or session dedicated to this action for the purpose of enabling follow-up queries, retrieval-augmented generation, context-preserved analysis, or similar within-matter functionality, provided that (i) such retention is confined to the matter-specific environment and does not train, fine-tune, or otherwise update the underlying model or any model used by any other party, matter, or user; and (ii) the matter-specific environment remains subject to the deletion obligations set forth in Section 12.2. [emphasis added]

For more information on, and examples of, less detailed A.I. protective orders, a discussion of standard of care, and a discussion of the possible equity impacts of such orders, please visit:

USE OF ARTIFICIAL INTELLIGENCE

Should a Party elect to utilize artificial intelligence (“AI”) to assist in reviewing, analyzing, or summarizing discovery material produced or exchanged in this litigation, the Party shall take all necessary steps to ensure that the application, service, or analytical software is fully containerized. “Fully containerized,” as used in this Order, means an AI tool that does not retain the substance of a prompt or documents reviewed, share the substance of a prompt or documents reviewed for training of large language models, or use the substance of a prompt or documents reviewed in any other matter or inquiry other than this litigation. For the avoidance of doubt, this restriction expressly applies to the use of non-containerized advanced large language models (such as publicly available large language models), generative AI tools, and other advanced AI systems, including, but not limited to, OpenAI, Gemini, LLAMA, MidJourney, and Stable Diffusion, but this provision does not limit the use of services leveraging the technology underlying these generative AI tools in a fully containerized environment, including, but not limited to, Copilot Harvey, Relativity aiR, DISCO Cecelia, Everlaw AI, Syllo, Alchemy, Lexis+ AI, Epiq AIDA, Lighthouse AI, Westlaw AI, ChatGPT Enterprise, Microsoft 365 Copilot Enterprise, Claude Enterprise, NotebookLM Pro and Ultra, and Gemini Pro and Ultra. Before submitting any Protected Material to a fully containerized AI tool, a receiving party shall ensure that it (or its vendor) can delete all such Protected Material from the AI tool at the conclusion of this matter, including any derivative information stored within the tool. The obligations and restrictions of this paragraph apply even where the Protected Material has been anonymized.

  • S.A.  v. Lockhart, 2026 WL 3003727 (S.D.N.Y.  Oct. 6, 2026)(by consent):

Restrictions on Use of Artificial Intelligence Tools to Process Disclosure Materials.

a. “AI tool” means any automated system that uses machine learning, or similar techniques to process inputs and generate outputs, recommendations, or predictions, including but not limited to large language models, generative AI services, or AI-assisted software tools, whether cloud-based or otherwise.

b.  The defense team will not submit Disclosure Materials to a publicly accessible version of an AI tool (e.g., ChatGPT, Claude, Grok, Gemini, Copilot, Perplexity, Midjourney, Canva). The defense team will not submit Disclosure Materials to an AI tool for which the AI tool provider or vendor (e.g., OpenAI, Anthropic, Microsoft, Google) uses submitted data for its own model training, allows submitted data to be publicly accessed, discloses submitted data to third parties, or, except as specified below, retains submitted data. Publicly accessible AI tools pose unique risks to the security and integrity of Disclosure Materials, given the practical inability to claw back or delete data once it has been incorporated into a model.

c.  The defense may submit Disclosure Materials to an AI tool (i) that maintains the confidentiality of Disclosure Materials; (ii) that will not expose any Disclosure Materials to third parties not authorized under this Protective Order to receive such materials; (iii) that will not retain or use any Disclosure Materials for model training; and (iv) all Disclosure Materials will be removed from the AI tool(s) at the conclusion of this case.

d.  The Government will not submit any materials disclosed by the defense pursuant to Federal Rules of Criminal Procedure 26.2, 32.1 or other defense disclosure obligations (“Defense Materials”) to a publicly accessible version of an AI tool. The Government will not submit Defense Materials to an AI tool for which the AI tool provider or vendor uses submitted data for its own model training, allows submitted data to be publicly accessed, discloses submitted data to third parties, or, except as specified below, retains submitted data.

e.  The Government may submit Defense Materials to an AI tool (i) that maintains the confidentiality of Defense Materials; (ii) that will not expose any Defense Materials to third parties not authorized under this Protective Order to receive such materials; (iii) that will not retain or use any Defense Sealed Materials for model training; and (iv) all Defense Materials will be removed from the AI tool(s) at the conclusion of this case.

  • S.A. v. Santos, 2026 WL 3003653 (Oct. 6, 2026)(same, by consent);
  • S.A. v. Robinson, 2026 WL 3003677 (Oct. 6, 2026)(same).
  • S.A. v. Peplow, 2026 WL 3026474 (E.D. Cal. Oct. 8, 2026)(stipulation):

Artificial Intelligence Tools

  1. No person or entity authorized to have access to protected material under the terms of the requested order shall input, transmit, upload, submit, quote, feed, process, generate output from, or otherwise expose any protected material received pursuant to the requested order to any artificial intelligence (“AI”) tool absent the prior written consent of the government or an order of the Court following a properly noticed motion to which the government has an opportunity to respond, unless that AI tool—a) is an enterprise-grade platform that counsel for the defendant has licensed; b) is subject to a binding written agreement that (i) requires the provider to keep all user-supplied data strictly confidential, and (ii) expressly prohibits the provider from using such data for training, fine-tuning, product improvement, or any purpose other than providing the contracted-for services; and c) employs technical and organizational security measures reasonably designed to prevent any unauthorized access, disclosure, or use of protected material.

  2. The obligations and restrictions of this section apply even where the data or the protected material has been anonymized. “AI tool” means any automated system that uses statistical modeling, machine learning, or similar techniques to process inputs and generate outputs, recommendations, or predictions, including but not limited to large language models, generative AI services, or AI-assisted software tools, whether cloud-based or otherwise.

  3. By participating in this agreement, counsel for the defendant certifies: (i) they have implemented reasonable measures to ensure, if using any AI tool, it will maintain the protected discovery’s confidentiality as required by this Agreement and (ii) they will ensure all protected discovery is deleted from the tool once this case concludes.

  4. Under no circumstances will any member of the defense team submit protected material to a publicly accessible AI system that retains and uses submitted data to train models. Such AI tools pose unique risks to the security and integrity of protected material, given the practical inability to claw back or delete data once it has been incorporated into a model.

  • S.A. v. Sarris, 2026 WL 2937892, at *2 (S.D.N.Y.  Sep 30, 2026)(by consent).
  • S.A. v. DeJesus, 2026 WL 3034296 (S.D.N.Y. Oct. 9, 2026)(by consent).

Note: In a recent decision, the District of Maryland wrote: “In addition, Campbell must ensure that any use of generative artificial intelligence does not violate the stipulated order governing confidential materials. See ECF 21. TidalHealth expresses concern that Campbell violated the February 27, 2026 Protective Order’s prohibition on submitting discovery material to public generative artificial intelligence tools. ECF 41-1 at 11; see ECF 21 at ¶ 8 (imposing limitations on use of generative artificial intelligence). Campbell denies doing so. ECF 45 at 3. If future filings suggest improper submission of discovery materials to generative artificial intelligence tools, the Court will be more inclined to issue a show cause order and consider sanctions.” Campbell v. TidalHealth, Inc., 2026 WL 2162749, at *10  (D. Md. Jul. 28, 2026)(Austin, J.)(emphasis added).  This case was discussed in a prior post.

UPDATE:  STT Webos, Inc. v. ByteDance Ltd., 2026 WL 2138557, *13 (N.D. Cal. Jul. 24, 2026)(emphasis added):

Artificial Intelligence Use. The Parties agree that any documents, electronically stored information (“ESI”), or other materials produced or exchanged in the course of this Action, including but not limited to those designated as “CONFIDENTIAL,” “HIGHLY CONFIDENTIAL – ATTORNEYS’ EYES ONLY,” or “HIGHLY CONFIDENTIAL – SOURCE CODE” under this Order, shall not be used, in whole or in part, for prompts (inputs), completions (outputs), or embeddings that are used to train, improve, re-train, or otherwise develop any artificial intelligence, foundation model, or large language model (collectively “AI Models”). This prohibition applies regardless of whether the AI Models are proprietary, open-source, or provided by a third party, and regardless of whether the use is for commercial, research, or any other purpose. The Parties further agree to take reasonable steps to ensure that their employees, counsel, agents, consultants, contractors, and any other persons or entities who may have access to such materials comply with this prohibition. Any violation of this provision may be addressed by the Court through appropriate remedies, including but not limited to sanctions.

____

[1] Three More A.I. Protective Orders (Oct. 9, 2026).

]]> Defendant Claimed Video of Abuse Was a “Deepfake”; Victim Was Asleep and Could Not Attest to Abuse https://www.ediscoveryllc.com/defendant-claimed-video-of-abuse-was-a-deepfake-victim-was-asleep-and-could-not-attest-to-abuse/ Wed, 14 Oct 2026 17:29:28 +0000 https://www.ediscoveryllc.com/?p=5650 […]]]> After an alleged “deepfake” video of a sexual assault was admitted over objection, the Defendant appealed his conviction.  The Supreme Court of Iowa wrote: “In this sexual abuse case, we are asked to decide whether the district court erred in admitting a video. The video graphically shows the defendant abusing the victim. However, there was no chain-of-custody evidence connecting the video to the defendant. The victim was asleep at the time and could not personally attest to the abuse shown in the video. And the defendant claimed ‘deepfake.’”  State of Iowa v. Amyda, __ N.W.2d __, 2026 WL 3030750 (Sup. Ct. Iowa Oct. 9, 2026).

A conclusory and unsupported “deepfake” assertion did not overcome circumstantial authentication evidence to the contrary.

The appellate court affirmed the conviction.  While the court’s decision is explicit, I have chosen to omit some facts with a notation [… Intimate Facts Omitted…].

The video came to the victim’s attention about a month after the defendant had been in the victim’s bedroom. The video appears to be taken by the perpetrator of the abuse at close range using a cellphone. It shows the victim sleeping, but she testified in detail that she recognized her bedroom, her bed and furnishings, the clothing she had been wearing, her legs [… Intimate Facts Omitted…], and the defendant’s tattooed hand—all as they had been that evening. Collectively, these circumstances were enough to admit the video. And the defendant’s mere speculation that the video was a deepfake was not enough to bar its admission.

Id. at *1 (emphasis added).

Skipping over some facts, the victim, Clara, got together with her friend’s uncle, the Defendant Amyda.  She knew him and  believed she could trust him.  They drank a bottle of rum, but when he said that he wanted to have sex with her, she refused.  “Clara, fully clothed but having consumed a lot of alcohol, fell asleep.” Id.  When she awoke the next day, “everything seemed fine.”  Defendant Amyda was still there, but left.

About a month later, Clara’s sister forwarded a video to her. It showed her being abused. Id. at *1.  When Clara tried to contact him, he blocked her.  She pressed charges. The police tried to interview Amyda—he agreed to meet but repeatedly failed to appear.

At that point, Officer Hansen decided to have Amyda arrested. Yet Officer Hansen never obtained a search warrant for Amyda’s electronic devices. Officer Hansen explained that Amyda did not have his cellphone on his person at the time of his arrest. Officer Hansen also was unable to retrace the digital footprint of the video.

Id. at *2 (emphasis added).

The court wrote: “When Amyda was arrested, his left hand had a different tattoo from the one shown in the video. But a tattoo artist who examined a booking photo of the hand testified at trial. According to his testimony, the tattoo appeared to have been freshly applied within the last eight to twelve days.”  Id. at *2.  For another case in which a suspect tried to change his appearance before trial, see Suspect’s Haircut = Destruction or Concealment of Evidence Leading to “Missing Evidence” Jury Instruction (Aug. 13, 2022).

Defendant Amyda filed a pretrial motion in limine to bar the video.  “Amyda argued that the video was not properly authenticated, was vulnerable to ‘deepfake’ technology, and furthermore was not admissible under the best-evidence rule because it was a copy and not the original. The State resisted.”  Id.

The district court held a pretrial evidentiary hearing. There Clara testified that she recognized the room depicted in the video as her bedroom. She identified her bed, her blanket, her pillowcases, and the top that she had been wearing on May 5. [… Intimate Facts Omitted…] Overall, she confirmed that her appearance in the video matched how she would have looked on the night of May 5. She also recognized the hand appearing in the video as Amyda’s based on the tattoo, scars, and other features. At the conclusion of the hearing, the district court issued a detailed ruling denying Amyda’s motion in limine.

On August 13, 2024, the case went to trial. When the State sought to introduce the video into evidence, Amyda renewed his objections. They were again overruled. On August 16, the jury returned a verdict finding Amyda guilty.

Id. at *2.

The intermediate appellate court affirmed, holding that there was sufficient circumstantial evidence to authenticate the video.  It pointed to “a showing with the victim’s detailed testimony that she recognized her body, her partial clothing, the room, and Amyda’s hand—all consistent with the only time that he was in her room in early May 2023.”  Id. at *2.  It rejected the “deepfake” challenge, writing:

So too was the [trial] court within its discretion to reject Amyda’s best-evidence challenge. Assuming the rule applies to this video and that the video is not an original, it was still admissible as a duplicate because Amyda’s speculative claims that the video was a “deepfake” without any evidentiary support did not raise a genuine question about the video’s authenticity.

Id. at *3 (emphasis added).

The Supreme Court of Iowa agreed. A video may be authenticated by its “appearance, contents, substance, internal patterns, or other distinctive characteristics of the item, taken together with all the circumstances.”  Id. at *3. The court noted that a video may be authenticated “even though no one could directly confirm its accuracy or testify as to how it was taken….  The harder cases are those in which direct testimony is not available and the proponent must rely on circumstances to lay a foundation for admission….  This, of course, was one of the ‘harder cases.’”  Id. at *3 (cleaned up).  In short, circumstantial evidence may be sufficient to authenticate a video.

Here, that test was met:

The video appears to be taken with a cellphone held in the right hand of the person whose left hand is visible committing sexual abuse. Clara had communicated with Amyda by messaging him that evening; presumably, he would have brought his cellphone to their encounter. Also, Clara fell asleep in her bedroom with Amyda after they both had been drinking; the video shows her face and closed eyes as she is sleeping soundly. Further, Clara identified Amyda’s tattooed hand, her bedroom, her bed and pillows, her clothing, and her legs [… Intimate Facts Omitted…] in the video. Indeed, the only thing in the video that Clara couldn’t personally verify was the abuse itself. Moreover, the video surfaced a little over a month after May 5, and when Clara tried to communicate with Amyda about it, he blocked her.[1]

Rejecting Defendant’s argument that this was legally insufficient, the court wrote:  “But the circumstantial evidence here went beyond just the distinctive characteristics of items in a video. Furthermore, the distinctive characteristics here ranged from the private (Clara’s bedroom) to the shockingly private [… Intimate Facts Omitted…].  What third-party fabricator would have access to all that?”  Id. at *5.

The Iowa court cited Mooney v. State, 487 Md. 701, 321 A.3d 91, 107 (2024), which I discussed in Can a witness authenticate a video if the video contains images that the witness did not see? (Aug. 16, 2024).

It also cited other precedent that: “Defendant’s allusions to the possibility the videos could be altered are not backed by any evidence in the record nor does Defendant make any specific claim to such effect.” Id. at *6.  However, it explained: “We are not oblivious to the risks of video evidence. Like the New York Court of Appeals, we recognize that ‘most fabricated videos ‘leverage’ real details from real photos and videos of real places and people, then alter the pieces the person wishes to alter to create a realistic, but manipulated, video.’ … But here, by contrast, there are no identifiable circumstances to suggest that the video might be a deepfake.”  Id. at *6 (emphasis added).

The video was a version that had been sent to the victim.  The court also rejected a “best evidence” challenge:

Much of the closing argument made by Amyda’s counsel urged the jury to conclude that the video was a deepfake. Yet his advocacy rested on popular culture rather than any of the evidence in the case. The closing argument began, “Is it fake? Is it real? Ladies and gentlemen, we live in the age of artificial intelligence.” It continued, “In this state of artificial intelligence, people can manipulate a video to look like whatever they want to look — they want it to look like.” Amyda’s counsel proceeded to refer to various news stories about celebrities and other people who had been the subject of deepfake videos, including deepfake pornography.

 But apart from trying to cash in on the so-called “liar’s dividend,”[2] Amyda’s counsel had little to work with. The prosecutor countered as follows in her rebuttal argument:

So when [defense counsel] is wa[ ]ving around articles and talking about deepfake porn, that is not evidence, ladies and gentlemen. That is not evidence. What is evidence is what you heard from that stand.

Fake video, imagine someone to make a fake video in this matter would have to know, as [Clara] described, the intrinsic details of her body.

Unfortunately, we had to talk a lot about private matters, right? They had her [… Intimate Facts Omitted…]  on her legs that she identified. [… Intimate Facts Omitted…]  They had stubble on her legs. They had her hair positioned as such.

They had her exact clothing that she would have been wearing the night she was with the defendant, and even more so, [… Intimate Facts Omitted…]. They would have to know all those intrinsic details to make up this video. Who would know that?

The “best evidence” objection was overruled.

Interestingly, the Iowa court held that its rules operated similarly to proposed Fed.R.Evid. 901(c).

This rule would provide that “a party challenging the authenticity of an item of evidence on the ground that it has been fabricated, in whole or in part, by generative artificial intelligence … must present evidence sufficient to support a finding of fabrication to warrant an inquiry by the court.” … At that point the proponent would have to demonstrate to the court that the item “is more likely than not authentic.”

Id. at *7-8 (emphasis added).  “[T]o block its admission under rule 5.1003, Amyda had the initial burden to raise a genuine question about the authenticity of the original. He failed to do so.”  Id. at *8.

For a deep evidentiary dive into debunking a “deepfake” audio, see A Deep Dive Into the Pikesville, MD, Racist/Antisemitic Deepfake Audio (Jul. 21, 2026).

For another case rejecting a “deepfake” assertion, see Evidence Sufficient to Demonstrate that Audio Recording Was Not a Deepfake (Mar. 23, 2026).

The National Center on State Courts has published guidance: Deepfakes – National Center for State Courts (Mar. 6, 2026).

See generally:

This blog was initially posted on  Electronic Discovery Reference Model.

____

[1] For a decision admitting “unfriending” evidence, see “Unfriending” Evidence – Maryland Follows “Reasonable Juror” Standard in Authentication of Social Media (May 12, 2020).

[2] “The term ‘liar’s dividend’ refers to the possibility that public awareness of deepfakes will make it easier to raise doubts about legitimate evidence.”  The court cited Bobby Chesney & Danielle Citron, “Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security,” 107 Cal. L. Rev. 1753, 1785–86 (2019) (“[A] skeptical public will be primed to doubt the authenticity of real audio and video evidence. This skepticism can be invoked just as well against authentic as against adulterated content.”).

]]> Three More A.I. Protective Orders https://www.ediscoveryllc.com/three-more-a-i-protective-orders/ Fri, 09 Oct 2026 08:30:52 +0000 https://www.ediscoveryllc.com/?p=5617 […]]]> I have written a series of posts suggesting that AI protective orders are becoming the standard of care.  See A.I. Protective Order by Consent in Google Case (Sep. 13, 2026).

The jointly requested A.I. protective order in Hoffman v. Univ. of Nebraska Medical Ctr., 2026 WL 2969621 (D. Neb. Oct. 2, 2026), states:

Artificial Intelligence. Qualified Recipients shall not, without the Producing Party’s express written consent, upload, submit, or otherwise provide Confidential Discovery Material (or any information extracted therefrom) or any deposition testimony taken in this matter to any open-source artificial intelligence platform or service including, but not limited to, platforms or services that use or assemble uploaded information to train or improve machine-learning models. This prohibition applies regardless of whether the artificial intelligence platform or service represents it will maintain the confidentiality of information submitted by the user. The parties and their counsel are not prohibited from uploading, submitting, or otherwise providing Confidential Discovery Material (or information extracted therefrom) or any deposition testimony taken in this matter to a closed-source artificial intelligence platform or service, including Harvey, Westlaw, LexisNexis, Servient’s AI Canvas, or any other closed-source artificial intelligence platform or service agreed to by the parties.

In ANR Pipeline Co., LLC v. B&B Metal Processing Co., Inc., 2026 WL 2970050 (E.D. Wisc. Oct. 2, 2026), the court wrote:

On September 22, 2026, the parties filed a joint motion for the entry of a protective order, with a stipulated proposed order…. The parties request that the Court enter such an order to “prevent public disclosure of nonpublic confidential technical, commercial, or business information that may be exchanged in discovery.” … Federal Rule of Civil Procedure Rule 26(c) allows for an order “to protect a party or person from annoyance, embarrassment, oppression, or undue burden or expense” including “requiring that a trade secret or other confidential research, development, or commercial information … be revealed only in a specified way.” Fed. R. Civ. P. 26(c)(1)(G); see also Civ. L.R. 26(e).

Protective orders are an exception to the general rule that pretrial discovery must occur in the public eye…. Litigation must be conducted in public to the maximum extent consistent with respecting trade secrets … and other facts that should be held in confidence….

Nonetheless, the Court can enter a protective order if the parties have shown good cause and that the order is narrowly tailored to serve that cause…. The Court can find that even broad, blanket orders are narrowly tailored and permissible when it finds that two factors are satisfied: (1) that the parties will act in good faith in designating the portions of the record that should be subject to the protective order; and (2) that the order explicitly allows the parties to the case and other interested members of the public to challenge the sealing of documents….

The Court finds that the parties have requested the protective order in this action in good faith. The parties report that this case will entail the disclosure of confidential technical information about the parties…. Thus, the Court is satisfied that there exists a sufficient basis for the requested protective order. [cleaned up].

However, the court modified the parties’ request to establish parameters for the use of A.I. as it pertains to confidential documents.  Id. at *1.  The order states:

A receiving party may use AI tools in connection with discovery materials produced in this action only if the AI tool is operated in a secure environment and is subject to enforceable contractual terms that (a) prohibit the provider from retaining, using, disclosing, selling, or using the materials to train, improve, or develop any AI system; (b) prohibit the materials from being made publicly available or accessible to any other customer or third party; and (c) require commercially reasonable administrative, technical, and physical safeguards to protect the confidentiality of the materials. Discovery materials, whether designated CONFIDENTIAL or ATTORNEYS’ EYES ONLY, shall not be entered into any publicly available or consumer-based AI platform. Any use of AI tools must remain consistent with this Order and with the requirement that discovery materials be used solely for purposes of this litigation.

Id. at *5.

A protective order was also entered by consent in U.S.A. v. Amanuel, 2026 WL 2964210 (S.D.N.Y. Oct. 1, 2026):

Restrictions on Use of Artificial Intelligence Tools to Process Disclosure Materials.

a.  “AI tool” means any automated system that uses machine learning, or similar techniques to process inputs and generate outputs, recommendations, or predictions, including but not limited to large language models, generative AI services, or AI-assisted software tools, whether cloud-based or otherwise.

b.  The defense team will not submit Disclosure Materials to a publicly accessible version of an AI tool (e.g., ChatGPT, Claude, Grok, Gemini, Copilot, Perplexity, Midjourney, Canva). The defense team will not submit Disclosure Materials to an AI tool for which the AI tool provider or vendor (e.g., OpenAI, Anthropic, Microsoft, Google) uses submitted data for its own model training, allows submitted data to be publicly accessed, discloses submitted data to third parties, or, except as specified below, retains submitted data. Publicly accessible AI tools pose unique risks to the security and integrity of Disclosure Materials, given the practical inability to claw back or delete data once it has been incorporated into a model.

c. The defense may submit Disclosure Materials to an AI tool (i) that maintains the confidentiality of Disclosure Materials; (ii) that will not expose any Disclosure Materials to third parties not authorized under this Protective Order to receive such materials; (iii) that will not retain or use any Disclosure Materials for model training; and (iv) all Disclosure Materials will be removed from the AI tool(s) at the conclusion of this case.

Some prior posts are:

This blog was initially posted on  Electronic Discovery Reference Model.

UPDATE:

White v. Chime Financial, Inc., 2026 WL 3000904 (W.D. Wash. Oct. 6, 2026)(stipulated):

USE OF ARTIFICIAL INTELLIGENCE
Should a Party elect to utilize artificial intelligence (“AI”) to assist in reviewing, analyzing, or summarizing discovery material produced or exchanged in this litigation, the Party shall take all necessary steps to ensure that the application, service, or analytical software is fully containerized. “Fully containerized,” as used in this Order, means an AI tool that does not retain the substance of a prompt or documents reviewed, share the substance of a prompt or documents reviewed for training of large language models, or use the substance of a prompt or documents reviewed in any other matter or inquiry other than this litigation. For the avoidance of doubt, this restriction expressly applies to the use of non-containerized advanced large language models (such as publicly available large language models), generative AI tools, and other advanced AI systems, including, but not limited to, OpenAI, Gemini, LLAMA, MidJourney, and Stable Diffusion, but this provision does not limit the use of services leveraging the technology underlying these generative AI tools in a fully containerized environment, including, but not limited to, Copilot Harvey, Relativity aiR, DISCO Cecelia, Everlaw AI, Syllo, Alchemy, Lexis+ AI, Epiq AIDA, Lighthouse AI, Westlaw AI, ChatGPT Enterprise, Microsoft 365 Copilot Enterprise, Claude Enterprise, NotebookLM Pro and Ultra, and Gemini Pro and Ultra. Before submitting any Protected Material to a fully containerized AI tool, a receiving party shall ensure that it (or its vendor) can delete all such Protected Material from the AI tool at the conclusion of this matter, including any derivative information stored within the tool. The obligations and restrictions of this paragraph apply even where the Protected Material has been anonymized.

U.S.A.  v. Lockhart, 2026 WL 3003727 (S.D.N.Y.  Oct. 6, 2026)(by consent):

Restrictions on Use of Artificial Intelligence Tools to Process Disclosure Materials.

a. “AI tool” means any automated system that uses machine learning, or similar techniques to process inputs and generate outputs, recommendations, or predictions, including but not limited to large language models, generative AI services, or AI-assisted software tools, whether cloud-based or otherwise.

b. The defense team will not submit Disclosure Materials to a publicly accessible version of an AI tool (e.g., ChatGPT, Claude, Grok, Gemini, Copilot, Perplexity, Midjourney, Canva). The defense team will not submit Disclosure Materials to an AI tool for which the AI tool provider or vendor (e.g., OpenAI, Anthropic, Microsoft, Google) uses submitted data for its own model training, allows submitted data to be publicly accessed, discloses submitted data to third parties, or, except as specified below, retains submitted data. Publicly accessible AI tools pose unique risks to the security and integrity of Disclosure Materials, given the practical inability to claw back or delete data once it has been incorporated into a model.

*2 c. The defense may submit Disclosure Materials to an AI tool (i) that maintains the confidentiality of Disclosure Materials; (ii) that will not expose any Disclosure Materials to third parties not authorized under this Protective Order to receive such materials; (iii) that will not retain or use any Disclosure Materials for model training; and (iv) all Disclosure Materials will be removed from the AI tool(s) at the conclusion of this case.

d. The Government will not submit any materials disclosed by the defense pursuant to Federal Rules of Criminal Procedure 26.2, 32.1 or other defense disclosure obligations (“Defense Materials”) to a publicly accessible version of an AI tool. The Government will not submit Defense Materials to an AI tool for which the AI tool provider or vendor uses submitted data for its own model training, allows submitted data to be publicly accessed, discloses submitted data to third parties, or, except as specified below, retains submitted data.

e. The Government may submit Defense Materials to an AI tool (i) that maintains the confidentiality of Defense Materials; (ii) that will not expose any Defense Materials to third parties not authorized under this Protective Order to receive such materials; (iii) that will not retain or use any Defense Sealed Materials for model training; and (iv) all Defense Materials will be removed from the AI tool(s) at the conclusion of this case.

U.S.A. v. Santos, 2026 WL 3003653 (Oct. 6, 2026)(same, by consent); U.S.A. v. Robinson, 2026 WL 3003677 (Oct. 6, 2026)(same).

U.S.A. v. Peplow, 2026 WL 3026474 (E.D. Cal. Oct. 8, 2026)(stipulation):

Artificial Intelligence Tools
15. No person or entity authorized to have access to protected material under the terms of the requested order shall input, transmit, upload, submit, quote, feed, process, generate output from, or otherwise expose any protected material received pursuant to the requested order to any artificial intelligence (“AI”) tool absent the prior written consent of the government or an order of the Court following a properly noticed motion to which the government has an opportunity to respond, unless that AI tool—

a) is an enterprise-grade platform that counsel for the defendant has licensed;
b) is subject to a binding written agreement that (i) requires the provider to keep all user-supplied data strictly confidential, and (ii) expressly prohibits the provider from using such data for training, fine-tuning, product improvement, or any purpose other than providing the contracted-for services; and
c) employs technical and organizational security measures reasonably designed to prevent any unauthorized access, disclosure, or use of protected material.

16. The obligations and restrictions of this section apply even where the data or the protected material has been anonymized. “AI tool” means any automated system that uses statistical modeling, machine learning, or similar techniques to process inputs and generate outputs, recommendations, or predictions, including but not limited to large language models, generative AI services, or AI-assisted software tools, whether cloud-based or otherwise.

17. By participating in this agreement, counsel for the defendant certifies: (i) they have implemented reasonable measures to ensure, if using any AI tool, it will maintain the protected discovery’s confidentiality as required by this Agreement and (ii) they will ensure all protected discovery is deleted from the tool once this case concludes.

18. Under no circumstances will any member of the defense team submit protected material to a publicly accessible AI system that retains and uses submitted data to train models. Such AI tools pose unique risks to the security and integrity of protected material, given the practical inability to claw back or delete data once it has been incorporated into a model.

U.S.A. v. Sarris, 2026 WL 2937892, at *2 (S.D.N.Y.  Sep 30, 2026)(by consent).

U.S.A. v. DeJesus, 2026 WL 3034296 (S.D.N.Y. Oct. 9, 2026)(by consent).

]]> Should There Be an “A.I. Exception” to Liberal Construction of Pro Se Pleadings? https://www.ediscoveryllc.com/should-there-be-an-a-i-exception-to-liberal-construction-of-pro-se-pleadings/ Sat, 26 Sep 2026 08:57:12 +0000 https://www.ediscoveryllc.com/?p=5596 […]]]> Pro se pleadings are entitled to liberal construction.  One court has repeatedly suggested, but not held, that the liberal construction rule should be revisited in the age of A.I.  Hickam v. Chesterfield Co. School Bd., 2026 WL 2861611 (E.D. Va. Sep. 23, 2026)(Young, J.).  Litigants opposing a pro se Complaint that appears to have been prepared with GenAI, may wish to consider Judge Young’s suggestion.

To the same effect, see Almudhafer v. Spotsylvania Cnty. Pub. Schs., 2026 WL 49654, at n. 2 (E.D. Va. Jan. 7, 2026)(Young, J.); Almudhafer v. Spotsylvania Volunteer Rescue Squad, 2026 WL 49652, at n. 3 (E.D. Va. Jan. 7, 2026)(Young, J.).  Another decision states:

The Court respectfully proposes that the time may be near for an exception to the Erickson liberal-construction rule, where a pro se individual relies on AI to draft pleadings and thus blurs the line between what is a good faith pro se assertion of an actionable claim and what is a computer-generated morass that only serves to waste court time and resources.

Mitchel v. Stellantis Fin. Servs., Inc., 2025 WL 2676569, at n. 6 (E.D. Va. Sept. 18, 2025)(Young, J.).

In Hickam, among other things, the unrepresented plaintiff made a “Motion for Leave to File Second Amended Complaint…, which is accompanied by a 121-page proposed Second Amended Complaint.”

The court wrote: “Based on the Court’s inherent power to manage its docket, and mindful of the requirements set forth in the Federal Rules of Civil Procedure to which all parties, including those proceeding pro se, are bound, the Court will sua sponte dismiss without prejudice Plaintiff’s First Amended Complaint for failure to comply with Federal Rule of Civil Procedure 8. Furthermore, the Court will deny without prejudice Plaintiff’s Motion for Leave to File Second Amended Complaint and direct Plaintiff, should he desire to continue this matter, to submit a proposed Second Amended Complaint that complies with Federal Rule of Civil Procedure 8.”

In part, the court explained that “to determine the viability of Plaintiff’s claims, the Court would have to engage in the improper task of sifting through 416 paragraphs to piece together a cogent timeline of events in order for the Court to assess how each of Plaintiff’s claims relates to each of the Defendants.”

The Hickam court recognized that a pro se complaint is “to be liberally construed” and “held to less stringent standards than formal pleadings drafted by lawyers,” citing Erickson v. Pardus, 551 U.S. 89, 94 (2007) (internal quotation marks omitted).

In a footnote, the court added:

The Court notes that Erickson’s liberal-construction requirement for pro se pleadings predates the advent of artificial intelligence (“AI”) drafting tools, which are being used with increasing prevalence by pro se parties in particular.…  Such tools both increase pro se litigants’ ability to draft voluminous pleadings and increase the odds that such pleadings contain fractured assertions and “hallucinated” legal propositions, which courts must wade through in construing claims…. While its analysis does not ultimately turn on this point, the Court notes that several of Plaintiff’s filings appear to carry the hallmarks of AI drafting assistance beyond that to which Plaintiff certifies. [emphasis added].

In the Spotsylvania Volunteer Rescue Squad decision, the court added: “[I]t respectfully proposes that the time may be near for an exception to the Erickson liberal-construction rule, where a pro se individual relies on AI to draft pleadings and thus blurs the line between what is a good faith pro se assertion of an actionable claim and what is a computer generated morass that only serves to waste court time and resources.”  2026 WL 49652, at n. 3.

The liberal construction rule has been recently and frequently cited. For example: “When a party is pro se, the court will construe her submissions ‘liberally to raise the strongest arguments that they suggest.’” Shah v. Bhagat, 2026 WL 2859875, at *3 (S.D.N.Y. Sept. 23, 2026)’

It will be interesting to see if an A.I. exception to the rule of liberal construction begins to emerge.

This blog was initially posted on  Electronic Discovery Reference Model.

]]> Another AI Protective Order https://www.ediscoveryllc.com/another-ai-protective-order/ Wed, 16 Sep 2026 13:46:12 +0000 https://www.ediscoveryllc.com/?p=5549 […]]]> The following Stipulated Protective Order was entered in Rapacon v. Abbott  Laboratories, 2026 WL 2722543 (N.D. Cal. Sep. 15, 2026):

  1. USE OF GENERATIVE AI TOOLS

“CONFIDENTIAL” Information or Items shall not be submitted to any Open Generative AI tool (e.g., Chat GPT) or any substantially similar tool that is available to the public. Providing “CONFIDENTIAL” Information or Items to an Open Generative AI tool is considered disclosure to a third party. For purposes of this Protective Order, “Open Generative AI tool” means an artificial intelligence system that may be accessed by the public, or that allows for the underlying code and materials to be accessed by the public.

The Parties may only submit “CONFIDENTIAL” Information or Items to a Permitted Generative AI Tool. For purposes of this Protective Order, a “Permitted Generative AI Tool” includes the use of enterprise AI tools that: (i) restrict access to authorized users; (ii) do not use customer data to train, refine, fine-tune, or improve publicly available models; (iii) provide reasonable retention and deletion controls; and (iv) maintain commercially reasonable security and confidentiality protections. For the avoidance of doubt, enterprise legal and productivity platforms that satisfy the foregoing requirements, including Thomson Reuters CoCounsel, Microsoft 365 Copilot, and similar enterprise tools, are permitted.

For more information on A.I. Protective Orders, please visit:

UPDATE:  ZL Technologies, Inc. v. KLDiscovery Ontrack, LLC, 2026 WL 2731631, at *9 (S.D.N.Y. Sep. 16, 2026):

Use of Artificial Intelligence Tools

(a) Scope: To protect the confidentiality and security of all information disclosed or produced in this proceeding (“Discovery Information”), any Party receiving Discovery Information from another Party (“Receiving Party”) shall comply with the following requirements governing the use of artificial intelligence (“AI”) tools with such information.

(b) Permitted AI Use: A Receiving Party may use AI tools to process Discovery Information solely for purposes of this proceeding, provided that the AI tool and its use collectively satisfy all of the following requirements:

i.  Restricted, Enterprise Environment: The AI tool is used pursuant to a written enterprise or professional agreement and is deployed within a secure, access-controlled environment.

ii.  No Model Training or External Use: Discovery Information is not used to train, fine-tune, improve, develop, benchmark, or otherwise contribute to any AI model or service accessible outside the Receiving Party’s environment, except to the limited extent necessary to provide the requested functionality for this proceeding.

iii. Data Retention and Deletion: Within 30 days after final disposition of this action, Discovery Information retained by the AI tool shall be destroyed or caused to be destroyed to the extent reasonably possible.

iv.  Security Safeguards: The AI tool is subject to technical and contractual safeguards reasonably designed to protect Discovery Information against unauthorized access, use, or disclosure, including encryption in transit and at rest and appropriate access controls.

(c) No Backdoor Access: Nothing in Section 23 permits a Party, Agent, Expert, or other person that is not able to access Confidential Discovery Information or Highly Confidential – Attorneys’ Eyes Only Information under this Order to use AI tools to query, analyze, search or otherwise process such information.

(d) Responsibility and Attribution: Each Receiving Party is responsible for compliance with this Order by its counsel, experts, consultants, vendors, or other agents using AI tools on its behalf.

(e) Attorney Work Product: Use of an AI tool shall not waive or diminish any applicable privilege or work-product protection, and prompts, instructions, workflows, chat histories, and outputs shall be treated the same as analogous attorney work product created without the use of AI tools.

UPDATE: U.S. v. Perez-Feliz, 2026 WL 2754823 (S.D.N.Y. Sep. 17, 2026)(consent order):

Restrictions on Use of Artificial Intelligence Tools to Process Disclosure Materials.

a. “AI tool” means any automated system that uses machine learning, or similar techniques to process inputs and generate outputs, recommendations, or predictions, including but not limited to large language models, generative AI services, or AI-assisted software tools, whether cloud-based or otherwise.

b. The defense team will not submit Disclosure Materials to a publicly accessible version of an AI tool (e.g., ChatGPT, Claude, Grok, Gemini, Copilot, Perplexity, Midjourney, Canva). The defense team will not submit Disclosure Materials to an AI tool for which the AI tool provider or vendor (e.g., OpenAI, Anthropic, Microsoft, Google) uses submitted data for its own model training, allows submitted data to be publicly accessed, discloses submitted data to third parties, or, except as specified below, retains submitted data.

c. The defense may submit Disclosure Materials to an AI tool (i) that maintains the confidentiality of Disclosure Materials; (ii) that will not expose any Disclosure Materials to third parties not authorized under this Protective Order to receive such materials; (iii) that will not retain or use any Disclosure Materials for model training; and (iv) all Disclosure Materials will be removed from the AI tool(s) at the conclusion of this case.

UPDATE: U.S. v. Montes, 2026 WL 2824059, at *1 (S.D.N.Y. Sep. 21, 2026)(by consent); U.S. v. Hawkins 2026 WL 2825014, at *2 (S.D.N.Y. Sep. 21, 20260(same).

a. “AI tool” means any automated system that uses machine learning, or similar techniques to process inputs and generate outputs, recommendations, or predictions, including but not limited to large language models, generative AI services, or AI-assisted software tools, whether cloud-based or otherwise.

b. The defense team will not submit Disclosure Materials to a publicly accessible version of an AI tool (e.g., ChatGPT, Claude, Grok, Gemini, Copilot, Perplexity, Midjourney, Canva). The defense team will not submit Disclosure Materials to an AI tool for which the AI tool provider or vendor (e.g., OpenAI, Anthropic, Microsoft, Google) uses submitted data for its own model training, allows submitted data to be publicly accessed, discloses submitted data to third parties, or, except as specified below, retains submitted data. Publicly accessible AI tools pose unique risks to the security and integrity of Disclosure Materials, given the practical inability to claw back or delete data once it has been incorporated into a model.

c. The defense may submit Disclosure Materials to an AI tool (i) that maintains the confidentiality of Disclosure Materials; (ii) that will not expose any Disclosure Materials to third parties not authorized under this Protective Order to receive such materials; (iii) that will not retain or use any Disclosure Materials for model training; and (iv) all Disclosure Materials will be removed from the AI tool(s) at the conclusion of this case.

d. No person or entity authorized to have access to Disclosure Materials under the terms of this Order shall input, transmit, upload, process, generate output from, or otherwise expose any Disclosure Materials received pursuant to this Protective Order to any AI tool without prior written notice to, and corresponding receipt of acknowledgement from, the Government.

UPDATE: U.S.A. v Amanuel, 2026 WL 2972618 (S.D.N.Y. Oct. 2, 2026)(by consent).

Restrictions on Use of Artificial Intelligence Tools to Process Disclosure Materials.

a. “AI tool” means any automated system that uses machine learning, or similar techniques to process inputs and generate outputs, recommendations, or predictions, including but not limited to large language models, generative AI services, or AI-assisted software tools, whether cloud-based or otherwise.

b. The defense team will not submit Disclosure Materials to a publicly accessible version of an AI tool (e.g., ChatGPT, Claude, Grok, Gemini, Copilot, Perplexity, Midjourney, Canva). The defense team will not submit Disclosure Materials to an AI tool for which the AI tool provider or vendor (e.g., OpenAI, Anthropic, Microsoft, Google) uses submitted data for its own model training, allows submitted data to be publicly accessed, discloses submitted data to third parties, or, except as specified below, retains submitted data. Publicly accessible AI tools pose unique risks to the security and integrity of Disclosure Materials, given the practical inability to claw back or delete data once it has been incorporated into a model.

c. The defense may submit Disclosure Materials to an AI tool (i) that maintains the confidentiality of Disclosure Materials; (ii) that will not expose any Disclosure Materials to third parties not authorized under this Protective Order to receive such materials; (iii) that will not retain or use any Disclosure Materials for model training; and (iv) all Disclosure Materials will be removed from the AI tool(s) at the conclusion of this case.

d. The Government will not submit Defense Materials to a publicly accessible version of an AI tool (e.g., ChatGPT, Claude, Grok, Gemini, Copilot, Perplexity, Midjourney, Canva). The Government will not submit Defense Materials to an AI tool for which the AI tool provider or vendor (e.g., OpenAI, Anthropic, Microsoft, Google) uses submitted data for its own model training, allows submitted data to be publicly accessed, discloses submitted data to third parties, or, except as specified below, retains submitted data.

e. The Government may submit Defense Materials to an AI tool (i) that maintains the confidentiality of Defense Materials; (ii) that will not expose any Defense Materials to third parties not authorized under this Protective Order to receive such materials; (iii) that will not retain or use any Defense Sealed Materials for model training; and (iv) all Defense Materials will be removed from the AI tool(s) at the conclusion of this case.

 

 

 

 

 

 

]]> Proposal to Use Unilaterally-Designed A.I. for Search Was Rejected https://www.ediscoveryllc.com/proposal-to-use-unilaterally-designed-a-i-for-search-was-rejected/ Tue, 15 Sep 2026 19:25:55 +0000 https://www.ediscoveryllc.com/?p=5546 […]]]> The court in Goldstein v. RTX Corp., 2026 WL 2690139 (D. Del. Sep. 14, 2026), rejected a proposal to use a unilaterally-designed AI tool for a search.

The case is a putative collective and class action alleging a “nationwide pattern, practice, and/or policy of discriminating against older workers … in their hiring processes.”  At issue was the manner of collective notice procedures.

The parties raise four primary disputes regarding the proposed collective notice procedures: (1) whether notice should be authorized for applicants of the non-RTX Defendants, (2) how to determine which job postings qualify as “Recent Graduate Positions,” (3) how to identify putative collective members, and (4) whether a website to return consent forms should be publicly available or password protected.

Id. at *2.

This post focuses on No. 2. The court wrote:

Next, the parties dispute how to determine which job postings qualify as “Recent Graduate Positions” reserved for “those who have graduated in the prior year or two (or who will soon graduate), or applicants with college degrees who have less than 12 or 24 months of work experience.” … Plaintiffs propose negotiating search terms with Defendants that target experience and recent graduate limitations…. In contrast, Defendants suggest filtering positions in their Workday database based on several data fields and then training an artificial intelligence model to identify job descriptions meeting Plaintiffs’ definition of “Recent Graduate Positions.”

Id. at *3 (emphasis added).

The court adopted Defendants’ proposal to employ the Workday database fields.  However, addressing the A.I. proposal, the court wrote:

Plaintiffs’ proposal to negotiate appropriate search terms targeting experience and recent graduate limitations is ADOPTED-IN-PART, and Defendants’ competing proposal to train an artificial intelligence model is REJECTED. Defendants contend that the Workday database does not have a field that can cleanly isolate Recent Graduate Positions, and the requirements for these positions are phrased differently, posing the risk of an underinclusive search….  Plaintiffs note that they “requested the sample of job postings Defendants reviewed in reaching that conclusion[,]” but “Defendants have not produced any sample.” … A collaborative process to agree on search terms is preferable to the unilateral training of an artificial intelligence model, which is no less likely to raise issues of under-inclusiveness. However, Plaintiffs’ proposal is REJECTED to the extent that they seek to compel Defendants to produce a 30% sample of Recent Graduate job postings across the relevant time period. Plaintiffs provide no authority in support of this proposal, which is effectively a motion to compel discovery. Therefore, IT IS ORDERED that the parties shall meet and confer on search terms to employ in the search process within fourteen (14) days of this Memorandum Order.

Id. at *4 (emphasis added). The court ordered that:

Defendants’ proposals to: (i) employ the Workday database fields identified in their supplemental submission to identify which positions quality as “Recent Graduate Positions,” and (ii) exclude from the realm of Recent Graduate Positions any positions open to recent master’s graduates or “degree equivalency” positions open to either recent college graduates or those with equivalent work experience are ADOPTED. Plaintiffs’ proposal to negotiate appropriate search terms targeting experience and recent graduate limitations to further identify Recent Graduate Positions is ADOPTED-IN-PART, and Defendants’ competing proposal to train an artificial intelligence model is REJECTED. Accordingly, IT IS ORDERED that the parties shall meet and confer to negotiate appropriate search terms within fourteen (14) days of this Memorandum Order.

Id. at *5 (emphasis added).

]]>