The Evolving Landscape of AI in Defensive Document Discovery

For years, technology-assisted review (“TAR”) was state of the art in AI-assisted discovery. That is no longer true. Generative AI tools have moved from novelty to regular use in the defensive review and production of documents in discovery.  

For in-house legal departments managing complex litigation, understanding these tools, their limitations, and the disclosure and professional responsibility obligations they trigger is no longer optional. 

Technology-Assisted Review 

The first mainstream use of AI to facilitate defensive document discovery was TAR, sometimes referred to as “computer-assisted review” or “predictive coding.” The technology running behind the scenes in a TAR review is a form of machine learning where example documents are provided (both responsive and nonresponsive) and algorithms are used to identify similar documents and predict the likelihood of responsiveness of those documents. The first iteration of TAR involves simple passive learning or simple active learning, in which a human reviewer codes a sample set of documents (a “seed” set) for responsiveness that the software then uses to code the remaining documents. The advantage of using a simple passive learning approach is once the seed set has been reviewed, the model can make predictions about the remaining, unreviewed corpus. The second iteration of TAR involves continuous active learning (TAR 2.0). This iteration starts with a small seed set of documents to train the model. After this initial training, the software promotes additional documents for human review that it identifies as most likely to be responsive to the top of the workflow. As human reviewers evaluate and code those documents as responsive or nonresponsive, the model is continuously refined in real-time. The advantage of a continuous active learning approach is the ability to continue training and refining the model, but it does require more human review.  

TAR, in either iteration, is now a well-known, court-sanctioned method for conducting defensive document discovery. See Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125, 127 (S.D.N.Y. 2015) (“[I]t is now black letter law that where the producing party wants to utilize TAR for document review, courts will permit it.”). However, because TAR requires examples to compare with the review corpus, it may struggle to categorize documents that are unique or complex. 

Generative AI 

Generative AI closely mimics how human reviewers approach a defensive document review project, both in terms of the input into the software and the output. As to the input, instead of training the software by coding individual documents and allowing the software to learn through that coding (as is done with both iterations of TAR), the software is trained in much the same way as human reviewers are trained. Written prompts are used to instruct the software how to make responsiveness determinations. These written prompts can take a variety of forms, but closely resemble the instructions that would be given to a human reviewer in a document review protocol. The ability to mimic the human review workflow makes the entire training process more efficient. In our experience, the prompts can largely be pulled from a traditional document review protocol that is routinely prepared for instructing the human reviewers overseeing the review project or performing second level or quality-control review before documents are produced. In terms of the output, like TAR, generative AI provides a recommendation as to whether a document should be coded responsive or non-responsive. However, unlike TAR, generative AI provides the logic and reasoning, in narrative form, behind the software’s responsiveness determinations. 

Success in using generative AI relies heavily on how well the software is trained, which is dependent on how well the prompts are drafted. Initial prompts must be clear and precise. The process of generating the final set of prompts that are applied to the document universe can be iterative. Prompts can be continuously revised until the desired results are achieved, and results can be tested and validated using standard validation metrics. However, it is important to note that generative AI review is not an iterative process once the prompts have been run over the corpus. Thus, it is critical to test, revise, and validate the prompts before they are run against the review set. The narrative explanations that are provided make the entire process quicker and more efficient—if a non-responsive document is miscoded as responsive, or vice versa, the narrative reasoning for the responsiveness determination allows a human reviewer to identify precisely why the software miscoded the document, which makes it easy to revise the prompts. The narrative reasoning provided also helps with the validation process and facilitates subsequent human review. 

In addition to the superior output in terms of receiving narrative explanations, generative AI review is also superior to TAR because it does not have the limitations of an exemplar-based review application and it allows the subject matter expert to provide the instructions directly to the model, knowing that the instructions will be followed consistently and quickly and that the results can be validated. However, we recommend using continuous active learning workflows for human reviewers to make a responsiveness determination where the generative AI review process determined that responsiveness was borderline or a close call. 

Beyond simply identifying a document as responsive or non-responsive, generative AI can be used to identify privileged documents that may need to be redacted or withheld and can facilitate the generation of descriptions of the privileged material for input into a privilege log. It can also be used to identify personally identifiable information that may need to be redacted, such as bank account or social security numbers. Further, it can categorize documents by issue and identify key documents that may require early attention. The early identification of noteworthy documents is crucial in developing case strategy, in assessing the strength and weaknesses of a case or theory, and in preparing for the taking and defending of depositions. The use of generative AI has the capability to transform document review from a purely defensive exercise into a strategic one. And, significantly, not only can generative AI perform these functions quickly, but the use of generative AI can result in substantial cost savings in document discovery, which is one of the most expensive phases in the life of a case. Although running generative AI over a set of documents does typically generate costs, those costs are a fraction of the fees typically incurred to human review those documents. 

Disclosure Requirements for Use of Generative AI 

The level of disclosure required concerning the use of AI in defensive document discovery is constantly evolving. Although there is a body of case law concerning the level of disclosure required when utilizing TAR, there is no set standard. One principal underlying most decisions, however, is that transparency is crucial. See, e.g., In re Insulin Pricing Litig., 2025 WL 1112837, at *2 (D.N.J. Apr. 11, 2025) (“[C]ourts have mandated some level of transparency and validation of TAR methodologies in consideration of ‘the complexities of TAR’ and Rule 26(g)’s obligations for a producing party to undertake a reasonable inquiry in discovery.”); Progressive Cas. Ins. Co. v. Delaney, 2014 WL 3563467, at *10 (D. Nev. July 18, 2014) (“The cases which have approved technology assisted review of ESI have required an unprecedented degree of transparency and cooperation among counsel in the review and production of ESI responsive to discovery requests.”); Berger v. Graf Acquisition, LLC, 2024 WL 4541011, at *4 (Del. Ch. Oct. 21, 2024) (allowing use of TAR “so long as [the party is] transparent with the plaintiff about their computer-assisted review process”); Moore v. Publicis Groupe, 287 F.R.D. 182, 192 (S.D.N.Y. 2012) (“[T]ransparency allows the opposing counsel (and the Court) to be more comfortable with computer-assisted review, reducing fears about the so-called ‘black box’ of the technology. This Court highly recommends that counsel in future cases be willing to at least discuss, if not agree to, such transparency in the computer-assisted review process.” (footnote omitted)), adopted by 2012 WL 1446534 (S.D.N.Y. Apr. 26, 2012); In re Diisocyanates Antitrust Litig., 2021 WL 4295729, at * 7 (W.D. Pa. Aug. 23, 2021) (“Transparency transcends cooperation. It does not mean merely that parties must discuss issues concerning the discovery of ESI; it requires that they disclose information sufficient to make those discussions, as well as any court review, meaningful.”), adopted by 2021 WL 4295719 (W.D. Pa. Sept. 21, 2021). 

Transparency, at a minimum, includes disclosing the use of AI to the opposing party. See, e.g., In re Valsartan, Losartan, & Irbesartan Prods. Liability Litig., 337 F.R.D. 610, 618 (D.N.J. 2020) (faulting party for failing to disclose “that it might use TAR . . . at th[e] time it was objectively reasonable and foreseeable”). It has also extended to disclosing details about the software and validation methodologies used. See, e.g., Kaye v. N.Y.C. Health & Hospitals Corp., 2020 WL 283702, at *2 (S.D.N.Y. Jan. 21, 2020) (disclosing “detailed information regarding the collection criteria they used, the name of their continuous active learning (‘CAL’) software, their CAL review workflow, and how they intend to validate the review results” is “sufficient . . . to make the production transparent”); Livingston v. City of Chicago, 2020 WL 5253848, at *3 (N.D. Ill. Sept. 3, 2020) (“The City has disclosed the TAR software— Relativity’s AL—it intends to use and how it intends to validate the review results, which in this case is sufficient information to make the production transparent.”). 

However, some courts have stopped short of requiring disclosure of the seed sets used to train TAR software and of requiring the opposing party’s involvement in the review and validation processes. See, e.g., In re Biomet M2a Magnum Hip Implant Prods. Liability Litig., 2013 WL 6405156, at *2 (N.D. Ind. Aug. 21, 2013) (“I won’t order Biomet to reveal which of the documents it has disclosed were used in the seed set, but I urge Biomet to re-think its refusal.”); Livingston, 2020 WL 5253848, at *3 (“Plaintiffs’ insistence that the City must collaborate with them to establish a review protocol and validation process has no foothold in the federal rules governing discovery.”). 

Given the newness of generative AI as a tool for defensive document discovery, there is less clarity on the level of disclosure required. But the body of case law that has developed regarding the use of TAR suggests that a similar level of transparency will likely be required, including transparency with respect to the use of generative AI itself, the specific software used, and validation processes and metrics. There is an open question on the disclosure of the prompts used to instruct generative AI tools—are these akin to search terms, which are routinely shared and negotiated by parties? Or are these akin to the seed set of documents used to train TAR, which are sometimes voluntarily shared between parties, but which some courts have refused to order be disclosed? Or are these more akin to document review protocols, which are widely regarded as work product and rarely shared? Courts will likely be forced to grapple with these issues as the use of generative AI becomes more prevalent and disputes inevitably arise. 

Professional Responsibilities When Using Generative AI 

The use of generative AI in discovery also triggers new applications of existing professional responsibilities. The American Bar Association addressed this in detail in Formal Opinion 512, “Generative Artificial Intelligence Tools,” on July 29, 2024 (“ABA Formal Opinion 512”). Among the primary professional responsibility rules implicated when generative AI is used in defensive document discovery are the following:  

Competency. Rule 1.1 of the Model Rules of Professional Conduct requires a lawyer to “provide competent representation to a client.” This obligation extends to understanding how to use AI tools and the accompanying risks. As this applies to generative AI, ABA Formal Opinion 512 describes this as a duty to “acquire a reasonable understanding of the benefits and risks of the [generative AI] tools that they employ in their practices” or else a duty to “draw on the expertise of others who can provide guidance about the relevant [generative AI] tool’s capabilities and limitations.” This also requires “an appropriate degree of independent verification or review” before relying on a generative AI tool’s output. ABA Formal Opinion 512.  

Confidentiality. Rule 1.6 of the Model Rules of Professional Conduct mandates the confidentiality of client information. Using paid versions of software and ensuring they have appropriate privacy protections in place to avoid using prompts or uploaded documents to train the model, thereby protecting confidentiality, privilege, and work product, is critical.  

Supervision. Rules 5.1 and 5.3 of the Model Rules of Professional Conduct dictate obligations for lawyers who manage or supervise other lawyers and non-lawyers. When generative AI is being used, these obligations extend to the supervision of the use of those AI tools. “Managerial lawyers must establish clear policies regarding the law firm’s permissible use of [generative AI], and supervisory lawyers must make reasonable efforts to ensure that the firm’s lawyers and nonlawyers comply with their professional obligations when using [generative AI] tools. Supervisory obligations also include ensuring that subordinate lawyers and nonlawyers are trained, including in the ethical and practical use of the [generative AI] tools relevant to their work as well as on risks associated with relevant [generative AI] use.” ABA Formal Opinion 512 (footnotes omitted). ABA Formal Opinion 512 also notes the importance of ensuring that generative AI providers and tools are appropriately vetted for both competency and their ability to protect confidential information.  

Client Communications. Rule 1.4 of the Model Rules of Professional Conduct mandates that a lawyer “reasonably consult with the client about the means by which the client’s objectives are to be accomplished.” ABA Formal Opinion 512 advises that “lawyers should consider whether the specific circumstances warrant client consultation about the use of a [generative AI] tool” and that, even if not required, explaining the use of a generative AI tool “may serve the interest of effective client communication.” When the use of generative AI is being evaluated or used, we coordinate closely with our clients in assessing the benefits, risks, costs, level of disclosure, and expected results.  

As of early 2026, a significant and growing number of states have also issued some form of AI guidance for practitioners, creating a patchwork of jurisdiction-specific obligations that counsel must navigate—particularly in multi-jurisdictional litigation where multiple states’ ethical rules may apply. In the next section, we provide a set of best practices for complying with these rules if generative AI is utilized. 

Best Practices 

Best practices when it comes to using generative AI to facilitate defensive document discovery are likely to evolve as the technology develops and becomes more widespread, and as disputes arise, prompting court decisions that will provide further guidance. Even in these early stages, however, there are a number of best practices to ensure a party using generative AI for defensive document discovery is well-situated for success in using the software and obtaining the desired results, for compliance with professional obligations, and for compliance with any potential disclosure requirements that may arise during the life of a case. 

Involvement and oversight by a senior attorney with subject matter expertise. Appoint a senior attorney with a thorough knowledge of the case to oversee the entire review process. Specifically, this attorney should own the prompt development process—drafting initial prompts, analyzing the AI’s narrative reasoning on sample sets, revising prompts based on error patterns, and signing off on final validation metrics before the prompts are applied at scale. The attorney who should be put in this role is akin to the attorney typically entrusted with developing a document review protocol and overseeing human reviewers in a manual human review of documents. 

Awareness of limitations of any software used. Promptly after deciding to utilize generative AI for defensive document review, identify any documents the software may not be able to process (e.g., image files, Excel files, audio files, or documents with a large amount of data) and any other limitations of the software being used (e.g., limitations in reviewing documents in multiple languages). If there are limitations to the software being utilized, develop a separate workflow and strategy for the review of documents that are incapable of being read by the generative AI tools, or are otherwise ill-suited to the generative AI workflow.  

Well-documented processes and procedures. Ensure that the processes and procedures used in employing generative AI are well-documented. Given the uncertain disclosure requirements when it comes to utilizing generative AI in defensive document discovery, and considering the focus on transparency when it comes to TAR, this is crucial to ensuring compliance with any possible disclosure requirements and for dealing with possible challenges to those processes by an opposing party or tribunal. In our experience, this is best accomplished by drafting a memorandum that memorializes, at a minimum, (a) the specific tool used (including the specific version and all settings applied), (b) the complete prompt history showing each iteration along with the final prompts applied across the review universe, (c) the validation methodology and results (including precision, recall, and F1 scores, where applicable) along with any sampling or quality control steps taken; (d) the level of human oversight and the attorneys providing that human oversight; and (e) any documents or file types excluded from the AI workflow with the basis for exclusion and the alternate workflow utilized for reviewing those documents. 

Early discussions and disclosures with opposing counsel. Disclose any intention to use generative AI for defensive document discovery to opposing counsel early in the life of a case. Although the disclosure requirements remain uncertain, early disclosure is the most prudent path forward. In our experience, these disclosures naturally arise during Rule 26(f) conferences (and state equivalents) or otherwise in the context of negotiating an ESI (electronically-stored information) protocol. Coming to an agreement between the parties on the use of generative AI, the level of disclosure required, and the validation metrics to be utilized, among other things, can help ward off discovery disputes, which can be costly, timeconsuming, and lead to surprises and unanticipated results.  

* * * 

 When used responsibly and with appropriate oversight, generative AI has the ability to substantially compress the time required to complete defensive document discovery. But realizing those benefits requires in-house legal departments to engage proactively—understanding the tools their outside counsel are deploying, asking the right questions about validation and documentation, and ensuring that ESI protocol negotiations address AI use early in the life of a case before disputes arise. Firms and in-house legal departments that invest in understanding and deploying these tools today will be best positioned to manage the cost, speed, and quality demands of complex discovery for years to come.