Please read pages 707-71-, 718-27, and 728-31.

Please answer any five of the following 17 questions and submit your answers on Brightspace:

1. What are the benefits and costs of prosecutorial discretion?

2. The first comment to Rule 3.8 asserts that “A prosecutor has the responsibility of a minister of justice and not simply that of an advocate.” What part of the text of Rule 3.8 do you think is most inconsistent with that idea and why?

3. The first comment to Rule 3.8 asserts that “A prosecutor has the responsibility of a minister of justice and not simply that of an advocate.” What from today’s readings (or listenings/watchings) is most inconsistent with that idea and why?

4. Describe the standard for initiating criminal charges and explain why you think it gives prosecutors either an appropriate amount of discretion or too much discretion.

5. Describe the standard for disclosing evidence to the defense and explain why you think it gives prosecutors either an appropriate amount of discretion or too much discretion.

6. Describe the standard for revisiting a wrongful criminal conviction and explain why you think it gives prosecutors either an appropriate amount of discretion or too much discretion.

7. Describe the standard that applies to potentially unreliable evidence and explain why you think it gives prosecutors either an appropriate amount of discretion or too much discretion.

8. Both the textbook and the Krasner interview highlight the role of the media in influencing the prosecutor’s decisions. What was meaningful to you about that point and why?

9. The textbook reading and the John Oliver segment include a few different statistics about prosecutors – which of those struck you the most and why?

10. Larry Kranser identifies alternative metrics for measuring a prosecutor’s success – to replace a reliance on conviction rates. Why do you think these alternative metrics are or are not useful? Are there any other metrics you would suggest?

11. In the Krasner interview, the interviewer describes a “tagline” of the progressive prosecutor movement is that “the job of the prosecutor is not just to secure convictions, it’s to seek justice.” Describe and explain your reactions to that tagline.

12. If we had to pick one area of prosecutorial power to reform first, which of the following would you pick and why: (a) charging decisions; (b) disclosure obligations; (c) use of unreliable evidence; (d) plea bargaining power; or (e) jury selection?

13. The John Oliver segment highlights the relationship between prosecutorial power and jury selection. What was meaningful to you about that point and why?

14. You may have fellow CUNY-Law colleagues who are planning to become prosecutors. Regardless of your own career goals, try to imagine why they would be drawn to that work and describe their possible motivations.

15. The John Oliver segment emphasizes the importance of running progressive candidates for prosecutor positions. Explain why that approach does or does not feel meaningful to you.

16. Please describe any specific ideas you have for reforming the role of the prosecutor.

17. Please describe any other reactions you had to the reading that was not captured by any of the above questions.

(a) The prosecutor is an administrator of justice, a zealous advocate, and an officer of the court. The prosecutor’s office should exercise sound discretion and independent judgment in the performance of the prosecution function.

(b) The primary duty of the prosecutor is to seek justice within the bounds of the law, not merely to convict. The prosecutor serves the public interest and should act with integrity and balanced judgment to increase public safety both by pursuing appropriate criminal charges of appropriate severity, and by exercising discretion to not pursue criminal charges in appropriate circumstances. The prosecutor should seek to protect the innocent and convict the guilty, consider the interests of victims and witnesses, and respect the constitutional and legal rights of all persons, including suspects and defendants.

(c) The prosecutor should know and abide by the standards of professional conduct as expressed in applicable law and ethical codes and opinions in the applicable jurisdiction. The prosecutor should avoid an appearance of impropriety in performing the prosecution function. A prosecutor should seek out, and the prosecutor’s office should provide, supervisory advice and ethical guidance when the proper course of prosecutorial conduct seems unclear. A prosecutor who disagrees with a governing ethical rule should seek its change if appropriate, and directly challenge it if necessary, but should comply with it unless relieved by court order.

(d) The prosecutor should make use of ethical guidance offered by existing organizations, and should seek to establish and make use of an ethics advisory group akin to that described in Defense Function Standard 4-1.11.

(e) The prosecutor should be knowledgeable about, consider, and where appropriate develop or assist in developing alternatives to prosecution or conviction that may be applicable in individual cases or classes of cases. The prosecutor’s office should be available to assist community efforts addressing problems that lead to, or result from, criminal activity or perceived flaws in the criminal justice system.

(f) The prosecutor is not merely a case-processor but also a problem-solver responsible for considering broad goals of the criminal justice system. The prosecutor should seek to reform and improve the administration of criminal justice, and when inadequacies or injustices in the substantive or procedural law come to the prosecutor's attention, the prosecutor should stimulate and support efforts for remedial action. The prosecutor should provide service to the community, including involvement in public service and Bar activities, public education, community service activities, and Bar leadership positions. A prosecutorial office should support such activities, and the office’s budget should include funding and paid release time for such activities.

(a) A lawyer should provide competent representation to a client. Competent representation requires the legal knowledge, skill, thoroughness and preparation reasonably necessary for the representation.

(b) A lawyer shall not handle a legal matter that the lawyer knows or should know that the lawyer is not competent to handle, without associating with a lawyer who is competent to handle it.

(c) A lawyer shall not intentionally:
(1) fail to seek the objectives of the client through reasonably available means permitted by law and these Rules; or
(2) prejudice or damage the client during the course of the representation except as permitted or required by these Rules.

Comment 8: To maintain the requisite knowledge and skill, a lawyer should (i) keep abreast of changes in substantive and procedural law relevant to the lawyer’s practice, (ii) keep abreast of the benefits and risks associated with technology the lawyer uses to provide services to clients or to store or transmit confidential information, and (iii) engage in continuing study and education and comply with all applicable continuing legal education requirements under 22 N.Y.C.R.R. Part 1500

This appeal is, in most respects, unremarkable. Plaintiff filed a complaint alleging a variety of employmentrelated claims, and the trial court granted defendants' motion for summary judgment, finding no triable issues as to any of those claims. Plaintiff challenges the grant of summary judgment on several grounds, none of which raises any novel questions of law or requires us to apply settled law in a unique factual context. In short, this is in most respects a straightforward appeal that,
under normal circumstances, would not warrant publication.

What sets this appeal apart-and the reason we have elected to publish this opinion-is that nearly all of the legal quotations in plaintiff's opening brief, and many of the quotations in plaintiff's reply brief, are fabricated. That is, the quotes plaintiff attributes to published cases do not appear in those cases or anywhere else. Further, many of the cases plaintiff [*2] cites do not discuss the topics for which they are cited, and a few of the cases do not exist at all. These fabricated legal authorities were created by generative artificial intelligence (AI) tools that plaintiff's counsel used to draft his appellate
briefs. The AI tools created fake legal authoritysometimes referred to as AI "hallucinations"-that were undetected by plaintiff's counsel because he did not read the cases the AI tools cited.

Although the generation of fake legal authority by AI sources has been widely commented on by federal and out-of-state courts and reported by many media sources, no California court has addressed this issue. We therefore publish this opinion as a warning. Simply stated, no brief, pleading, motion, or any other paper filed in any court should contain any citations- whether provided by generative AI or any other source-that the attorney responsible for submitting the pleading has not
personally read and verified. Because plaintiff's counsel's conduct in this case violated a basic duty counsel owed to his client and the court, we impose a monetary sanction on counsel, direct him to serve a copy of this opinion on his client, and direct the clerk of the [*3] court to serve a copy of this opinion on the State Bar. 

. . . 

DISCUSSION

I. Plaintiff's counsel's reliance on fabricated legal authority.

We begin by noting that nearly all of the quotations in plaintiff's opening brief, and many of the quotations in plaintiff's reply brief, have been fabricated. That is, as noted above, although most of the cases to which the quotes are attributed exist, the quotes do not. Further, many of the cases plaintiff cites do not support the propositions for which they are cited or discuss other matters entirely, and a few of the cases do not exist at all. To give just a few examples:

Plaintiff asserts: "In Schimmel v. Levin [*11] (2011) 195 Cal.App.4th 81, the court discussed the legislative purpose behind Section 437c(f)(2), highlighting that it was enacted to prevent abuse of the summary judgment procedure by disallowing multiple motions on the same issues." In fact, Schimmel does not contain a single reference to either summary judgment or section 437c.

Appellant's opening brief also purports to quote Schimmel as follows: "In Schimmel v. Levin (2011) 195 Cal.App.4th 81, 86-87, the court held: 'Section 437c(f)(2) embodies a legislative judgment that a party should not be allowed to bring multiple motions for summary judgment based on the same issues without demonstrating newly discovered facts or circumstances or a change in the law. This policy applies even when the prior motion was denied on procedural grounds.' " The quoted language does not appear in Schimmel-or in any other case of which we are aware.

Plaintiff also asserts: "In Regency Health Services, Inc. v.Superior Court (1998) 64 Cal.App.4th 1496, 1504, the court emphasized: 'A continuance should not be granted when it is sought to facilitate procedural maneuvers rather than to
promote justice.' " Regency does not address the granting of a continuance, and the quoted language does not appear anywhere in the opinion.

Plaintiff further asserts: "The court in Peake v.Underwood, 227 Cal.App.4th 428, 448 (2014), emphasized that filing a second dispositive motion without new facts [*12] or law is frivolous and subject to sanctions." Peake does not address the filing of a second dispositive motion, and the only sanctions at issue in that case were for filing a frivolous pleading. (Id. at pp. 432-450.)

Plaintiff additionally asserts: "As in Goldstine v. LibertyMut. Ins. Co., 2020 WL 6216738 (W.D. Wash. 2020), where sanctions were imposed for similar baseless claims of personal hardship to delay proceedings, Mr. Yadegari's actions warranted sanctions under California Code of Civil Procedure 128.5 for making false statements to obtain an improper
advantage." Goldstine appears to be a fabricated case. And, plaintiff asserts: "The California Court of Appeal in Heckert v. MacDonald, 208 Cal.App.3d 832, 837 (1989), emphasized that sanctions should be imposed where a party uses procedural rules to gain an unfair advantage by engaging in 'frivolous litigation tactics.' " The words "frivolous," "unfair," and "tactics" do not appear in Heckert, which concerns the appellants' claim that the trial court erred by refusing to order their real estate broker to pay their attorney fees as damages.

In total, appellant's opening brief contains 23 case quotations, 21 of which are fabrications. Appellant's reply brief contains many more fabricated quotations. And, both briefs are peppered with inaccurate citations that do not support the propositions for which they are cited.

The extensive [*13] reliance on nonexistent legal authority would justify striking appellant's opening brief or dismissing the appeal. . . . .

III. Sanctions for pursuit of a frivolous appeal.

Prior to oral argument in this case, on our own motion [*23] we issued an order to show cause (OSC) why this court should not sanction plaintiff's counsel, Amir Mostafavi, for filing appellate briefs replete with fabricated quotes and citations. The OSC noted that nearly all of the quotations in appellant's opening brief, as well as many in the reply brief, were fabricated, and it warned that sanctions might include both an award of attorney fees and costs to defendants and an award of sanctions payable to the clerk of this court.

Attorney Mostafavi filed a written response. He acknowledged that he relied on AI "to support citation of legal issues" and that the fabricated quotes were AIgenerated. He urther asserted that he had not been aware that generative AI frequently fabricates or hallucinates legal sources and, thus, he did not "manually verify [the quotations] against more reliable sources." Mostafavi accepted responsibility for the fabrications and said he had since taken measures to educate himself so that he does not repeat such errors in the future. He asserted, however, that "[t]he majority of citations are accurate
and support the propositions that were being advanced"; the appeal is not frivolous; and in spite of fabricated quotations, [*24] the brief "stands on meritorious arguments that are fully supported by the record."

Mostafavi therefore urges that "[s]hould the Court determine that some corrective action is warranted," the appropriate remedy "is correction of the briefs rather than monetary sanctions" because counsel's "citation irregularities," although "regrettable," "do not rise to the level requiring punitive measures given the isolated nature of the problems relative to the briefs' overall representation of the reversible errors made by the trial court based on the cited record and not necessarily in
complete reliance on the cited authorities."

At oral argument, attorney Mostafavi explained that he wrote initial drafts of the briefs, "enhanced" the briefs with ChatGPT, and then ran the "enhanced" briefs through other AI platforms to check for errors. Counsel admitted that he did not read the "enhanced" briefs before he filed them.

For the reasons that follow, we decline to permit the filing of revised briefs and conclude that an award of sanctions against attorney Mostafavi is appropriate.

 . . . 

B. Counsel's reliance on fabricated legal authority renders this appeal frivolous and violative of the California Rules of Court.

Appellant's counsel has acknowledged that his briefs are replete with fabricated legal authority, which he admits resulted from his reliance on generative AI sources such as ChatGPT, Claude, Gemini, and Grok. Counsel says that he was not previously aware of the problem of AI "hallucinations," but he has educated himself about the issue since receiving the OSC.

 . . .

We agree with the cases cited above that relying on fabricated legal authority is sanctionable. As a district judge recently held when presented with nonexistent precedent generated by ChatGPT: "A fake opinion is not 'existing law' and citation to a fake opinion does not provide a non-frivolous ground for extending, modifying, or reversing existing law, or for establishing new law. An attempt to persuade a court or oppose an adversary by relying on fake opinions is an abuse of the adversary
system." (Mata v. Avianca, Inc. (S.D.N.Y. 2023) 678 F.Supp.3d 443, 461, fn. omitted; see also Park v. Kim (2d Cir. 2024) 91 F.4th 610, 615 [quoting Mata].)

To state the obvious, it is a fundamental duty of attorneys to read the legal authorities they cite in appellate briefs or any other court filings to determine that the authorities stand for the propositions [*32] for which they are cited. Plainly, counsel did not read the cases he cited before filing his appellate briefs: Had he read them, he would have discovered, as we did, that the cases did not contain the language he purported to quote, did not support the propositions for which they
were cited, or did not exist. (See Benjamin v. Costco Wholesale Corporation, supra, 779 F.Supp.3d at p. 343 ["an attorney who submits fake cases clearly has not read those nonexistent cases, which is a violation of [the federal equivalent of 128.7]"]; Willis v. U.S. BankNational Association as Trustee, Igloo Series Trust (N.D. Tex., May 15, 2025, No. 3:25-cv-516-BN) 2025 WL 1408897, at *2 [same].)

Counsel thus fundamentally abdicated his responsibility to the court and to his client. (See Kleveland v. Siegel &Wolensky, LLP (2013) 215 Cal.App.4th 534, 559 [" 'It is critical to both the bench and the bar that we be able to rely on the honesty of
counsel. The term "officer of the court," with all the assumptions of honor and integrity that append to it, must not be allowed to lose its significance' "].)

Counsel acknowledges that his reliance on generative AI to prepare appellate briefs was "inexcusable," but he urges that he should not be sanctioned because he was not aware that AI can fabricate legal authority and did not intend to deceive the court. Although we take counsel at his word-and although there is nothing inherently wrong with an attorney  appropriately using AI in a law practice-before [*33] filing any court document, an attorney must "carefully check every case citation, fact, and argument to make sure that they are correct and proper. Attorneys cannot delegate that role to AI,
computers, robots, or any other form of technology. Just as a competent attorney would very carefully check the veracity and accuracy of all case citations in any pleading, motion, response, reply, or other paper prepared by a law clerk, intern, or other attorney before it is filed, the same holds true when attorneys utilize AI or any other form of technology." (See Versant, supra, 2025 WL 1440351, at *4.)

We note, moreover, that the problem of AI hallucinations has been discussed extensively in cases and the popular press for several years. . . . Thus, even a superficial review of the [*34] literature would have alerted counsel to this issue. Further, the State Bar of California released "Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law" nearly two years ago, in November 2023. . . . Additionally, the notes to Rule 1.1 of the California Rules of Professional Conduct expressly provide that "[t]he duties set forth in this rule include the duty to keep abreast of the changes in the law and its practice, including the benefits and risks associated with relevant technology." (See Editors' Note 1, Cal. Rules Prof.
Conduct, foll. rule 1.1.) We therefore do not agree that counsel's failure to educate himself about the limitations the legal tools he relied on makes the imposition of sanctions inappropriate.

C. An award of sanctions is appropriate in this case.

Sanctions may be awarded to the respondent to compensate for the costs of responding to a frivolous appeal, or to the clerk of the court for conduct that unnecessarily burdens the court and the taxpayers. As one court has explained, " 'Respondent[s] . . . 2025 Cal. App. LEXIS 584, *33 are not the only parties damaged when an appellant pursues a frivolous claim. Other appellate parties, many of whom wait years for a resolution of bona fide disputes, are prejudiced by the useless diversion of this court's attention. [Citation.] In the same vein, the appellate system and the taxpayers of this state are damaged by what amounts to a waste of this court's time and resources. [Citations.] Accordingly, an appropriate measure [*37] of sanctions should also compensate the government for its expense in processing, reviewing and deciding a frivolous appeal.' . . . 

Attorney Mostafavi's fabricated citations and erroneous statements of law have required this court to spend excessive time on this otherwise straightforward appeal to attempt to track down fabricated legal authority and then to research the issues presented without plaintiff's assistance. We therefore conclude that an award of sanctions payable to the court is appropriate. . . . In 2013, another appellate court noted that appellate sanctions for frivolous appeals recently had ranged from $6,000 to $12,500, "generally, but not exclusively, based on the estimated cost to the court of processing a frivolous appeal." (Kleveland, supra, 215 Cal.App.4th at p. 560, citing Kim v.Westmoore Partners, Inc. (2011) 201 Cal.App.4th 267, 294.) The costs of processing a frivolous appeal have undoubtedly increased in the intervening 12 years. Nonetheless, because counsel has represented that his conduct was unintentional, and because he has expressed remorse for his actions, we impose a conservative sanction of $10,000. Such sanction shall be payable to the clerk of this court within 30 days of the filing of the remittitur.

Large Language Model Generative AI (GenAI) sometimes produces inaccurate content. When factual accuracy is not essential, GenAI’s capacity to creatively innovate and connect disparate concepts can lead to new approaches and more successful client outcomes. When circumstances require accuracy, the use of GenAI must be contingent on (1) the lawyer having the necessary subject matter expertise and (2) the lawyer committing to independently evaluating the results.

Virtually all law firms are already using some form of Large Language Model Generative AI (GenAI).[1] The percentage of lawyers using GenAI went from less than 20% in 2022 to almost 80% in 2024.[2] The five most used applications include (1) discovery; (2) legal search; (3) document generation; (4) brief and memoranda generation; and (5) prediction of case outcomes.[3]

This explosion has been made possible, in part, due to the plethora of GenAI platforms specifically designed for the legal profession, such as Harvey AI,[4] CoCounsel,[5] Lexis+ AI,[6] and Westlaw Precision.[7] Many major law firms have also created internal versions of these tools, while criminal defense lawyers and smaller firms are also increasingly applying machine-learning techniques to document review and litigation preparation.[8]

When used well, the research across law firms, corporate legal departments, and the legal industry at large consistently finds that GenAI adoption leads to more efficient operations,[9] higher-quality work output, improved client outcomes,[10] and new opportunities for revenue or cost savings.[11] Lawyers across all sectors, from 1,500-lawyer big law to solo criminal defense attorneys are increasingly embracing GenAI, fully utilizing it in their practices and enjoying the immense benefits it provides.

But lawyers are also increasingly finding that the unskilled or improper use of GenAI has a propensity to create immense risk for themselves and their clients. The main culprit is that GenAI often fails to provide accurate output. The common description of this problem is that GenAI models hallucinate. Research shows that these false but plausible outputs generated by AI are not incidental glitches, but an inherent byproduct of how GenAI systems are designed.[12] These systems are built on probabilistic language modeling, where outputs are generated by predicting the most likely next word based on patterns in massive training datasets, not by verifying factual accuracy or understanding meaning.[13]

Because these models operate statistically rather than semantically, they lack any true grasp of context or truth, which results in fluent yet factually incorrect responses.[14] These errors often stem from overgeneralization: GenAI extrapolates patterns learned from data and fills in gaps with information that is statistically likely but not necessarily true.[15] In some cases, as Anthropic has observed, a model like Claude may appear to reason within a conceptual framework shared across languages — suggesting the presence of a kind of universal “language of thought.”[16] But even this apparent sophistication does not confer understanding. The persuasive fluency of AI-generated language makes hallucinations especially dangerous in the legal field, where even minor factual inaccuracies can have significant consequences.[17] Hallucinations are a structural limitation of current GenAI and must be managed through human oversight, rigorous verification, and professional responsibility rather than technological fixes alone.[18]

Therefore, this article proposes a new approach: a modified paradigm for legal practitioners using GenAI — one grounded not in blind trust or total skepticism, but in rethinking what GenAI is and how to best use it. In this paradigm, originally posited by Professor Jules White,[19] GenAI’s capacity to “hallucinate” is reframed as a positive feature of the technology rather than a malicious flaw.[20] Because guessing, sometimes wrongly, is a feature of the architecture, the system is a powerhouse for tasks like brainstorming or summarizing style and tone. In this alternative use paradigm, the emphasis is on navigation over generation, on reasoning over recitation, filtering over abject accuracy and most importantly, on traceability — ensuring that outputs can always be linked back to verifiable sources.[21]

The following sections explore practical applications of this paradigm, beginning with one of the most promising and useful for lawyers who litigate: filtering. The article examines how GenAI can be used not to generate new legal facts or conclusions, but to distill and structure information that the lawyer already has lawful access to. In doing so, it highlights a class of tasks where GenAI is not just helpful, but reliable — provided that the tasks are properly scoped, the prompts carefully designed,[22] and the results always anchored in verifiable input.

What emerges from this exploration is not a rejection of GenAI in legal practice, but a reframing of its role: not as an oracle, but as a sophisticated assistant. One that is powerful, but only when properly directed.

Correctness and the Cost of Error

Before using GenAI, thought should be given to the importance of the output being right. Viewed this way, the use of GenAI becomes a calculation where the correctness and efficacy are inversely proportional. The more important the accuracy, the less efficacious GenAI will be. Eventually one reaches a tipping point where the use of GenAI is untenable for a particular purpose. For lawyers exploring the use of GenAI, then, the threshold question must often be: How much does it matter if the answer is perfectly correct?

This question is essential because in legal work, the cost of being wrong can range from embarrassing to catastrophic. An AI-generated draft that misstates a deadline, misattributes a case citation, or mischaracterizes a statutory requirement is not merely inefficient, it can damage credibility, jeopardize a client’s position, or violate ethical duties.

And this is not merely a philosophical exercise or a simple odds calculation. In practice, much of the legal profession turns on precision. Therefore, where accuracy is paramount, it becomes essential to evaluate not only whether GenAI can produce a seemingly correct answer, but whether it can be easily and reliably checked.

Where Accuracy Is Vital — Seek Verification

Because hallucinations are inevitable, the trustworthiness of GenAI depends on how easily its output can be verified — ideally with minimal time or effort. Where accuracy is critical, GenAI should only be used if its answers can be quickly and reliably confirmed. Its value lies in generating solutions faster than a lawyer could unaided, but only when checking that solution is easier than creating it from scratch.

The reverse is also true. If verifying GenAI’s output requires deep expertise the user does not possess, its use becomes risky. Consider a criminal defense attorney asking GenAI to draft a document involving nuclear regulatory law and Radiation Exposure Standards. The GenAI may produce a fluent and detailed draft, but without domain knowledge, the lawyer has no way to assess whether it is correct. Verifying the content would require consulting a regulatory expert, thereby rendering the GenAI redundant. In such cases, GenAI creates the illusion of insight while offering no real path to truth.

By contrast, many criminal defense tasks permit straightforward validation. If a GenAI tool extracts key facts relevant to a search and seizure issue from a police report, or flags validation errors in chromatographs from an ethanol blood test, a defense attorney with subject matter expertise can readily confirm the accuracy of the result. These tasks are grounded in materials already in the lawyer’s possession or knowledge base, not ones that require external validation.

The key insight is this: A wrong answer is tolerable if it is obvious; it is dangerous when it is hidden. Legal professionals should prioritize use cases where outputs can be checked without relying on intuition, outside experts, or where the factual accuracy of the output is not important, such as in the ideation or brainstorming scenarios discussed below. In that framing, GenAI becomes not a replacement for expertise but a tool that supports it.

Filtering: Surfacing Without Substituting

One of the most immediately useful applications of GenAI in legal practice is the task of filtering. At its core, filtering involves asking the GenAI to sift through a body of information, such as lengthy discovery documents, or a corpus of case law and reduce it to the subset that is most relevant to a particular issue. This is not about the GenAI making autonomous legal decisions or conclusions, but about automating the lawyer’s capacity to review, interpret, and apply information more efficiently.

Risk

If verifying GenAI’s output requires expertise defense counsel does not possess, its use becomes risky.

Importantly, this process presumes that the lawyer already has access to the underlying materials. The GenAI is not being used to determine what information a lawyer should be utilizing, nor is it deciding what material can be disclosed to opposing counsel or to a court. Those questions are fundamentally ethical and procedural, and they must remain firmly in the discretion of the human lawyer.

Imagine, for example, that a lawyer is reviewing 2,000 pages of email communications produced in response to a FOIA directed to a forensic laboratory. The GenAI can be asked to isolate emails containing answers to questions such as, “What did they know and when did they know it?” In addition, counsel can use it to flag every message that contains a particular word or phrase. The GenAI could then be asked to formulate arguments for suppression based on the number of instances that are discovered. Here is a sample prompt to elucidate this use.

Example: Validity of Forensic Techniques — Measurement Uncertainty and Discredited Methods

Context: You are reviewing forensic lab documents for compliance with ISO/IEC 17025, which requires labs to estimate and document measurement uncertainty, including all significant contributors. You suspect the lab may have failed to account for key sources of uncertainty in its analysis.

Sample Prompt Template

Search the uploaded document [e.g., PDF containing lab emails or investigative reports] for instances of [insert specific keywords or phrases, e.g., “operator error,” “instrument drift,” or “chain of custody”]. Return a chart with the following columns: [page number] [matched phrase] [context snippet — 1–2 sentences showing how the phrase is used]. Highlight only relevant or potentially significant occurrences. Ignore irrelevant or duplicate matches unless the context differs meaningfully.

What makes this use of GenAI trustworthy is the ability to trace its output back to the original inputs. If the GenAI searches a set of 200 emails, the chart produced will include a page number allowing the output to be easily confirmed.

This principle of traceability is the bedrock of responsible GenAI filtering. Whether the GenAI is summarizing trial testimony or flagging key clauses in a set of discovery documents, each assertion or excerpt should point back to its source — line numbers, paragraph references, or direct quotations. For example, if the GenAI produces a summary sentence like “The witness admitted to deleting the emails,” it should immediately follow with a reference such as “evidentiary hearing testimony of John Smith, Page 132, Line 14.” This transparency not only enables accuracy but also supports the kind of due diligence that lawyers are ethically and professionally obligated to perform.

A critical distinction exists here between using GenAI to filter and catalog information versus using it to decide something on the lawyer’s behalf. A trustworthy use of GenAI aids human reasoning without supplanting it. The filtering process is particularly well-suited for this kind of support because the output is, by definition, a subset of the input. Nothing new is being created; no legal interpretations are being asserted by the system itself. This containment makes it relatively easy for the human lawyer to verify that the GenAI’s output is complete, relevant, and accurate. The value of the GenAI’s analysis is entirely dependent on the quality, completeness, and accessibility of the source materials the lawyer gives it.

Beyond filtering relevant documents, GenAI can also assist in making sense of the filtered results by identifying patterns, drawing preliminary conclusions, or suggesting how specific records may support a legal argument. Once potentially significant references are identified, lawyers can prompt the GenAI to assess their argumentative value: how a phrase or pattern might support a claim of forensic error, Brady{23}23  Brady v. Maryland, 373 U.S. 83, 83 S. Ct. 1194 (1963). violation, or institutional misconduct. For more nuanced filtering, attorneys can also ask the GenAI to search for the occurrence of a word or phrase (e.g., “operator error”) within a set number of words from another (e.g., “instrument drift”), helping isolate documents where meaningful connections appear. These capabilities turn GenAI from a document scanner into a strategic assistant.

Sample Prompt Template (Drawing Inferences from Filtered Content)

Based on the filtered results from the uploaded document, identify any patterns or themes that could support a defense argument. Consider how the occurrences of [insert keyword or phrase] — especially those appearing within [X] words of [insert related term] — may suggest systemic issues, procedural violations, or credibility problems. Summarize key conclusions and suggest how they might be used in support of motions, cross-examination, or discovery requests.

As legal professionals explore how to integrate GenAI tools into their workflows, filtering presents a practical, low risk starting point. With careful prompt design, such as asking the GenAI to filter, cite, and ground its outputs, lawyers can begin to utilize the scale and speed of GenAI while still exercising full professional judgment over the results. Filtering is not about replacing legal analysis; it is about supplementing it, making legal analysis faster and more focused. This method is especially useful when facing massive volumes of data that would otherwise be overwhelming to examine and prohibitively time consuming to analyze. When grounded in traceability and transparency, filtering becomes a reliable partner in GenAI-assisted legal work.

Ideation — Using GenAI to Spark Thought, Not Final Answers

Ideation is one of the more exceptional ways for lawyers to use GenAI because it allows the benefits of GenAI without directly requiring the output to be exact. In fact, ideation and others like it are among GenAIs most productive legal applications because it arises in precisely the opposite context — when accuracy is irrelevant or, at the very least, of secondary value. In these circumstances, relieved of the obligation to be right, GenAI is instead used as a catalyst for thinking.

Ideation is especially useful because it addresses a common hurdle for many criminal defense lawyers: finding the time to fully explore a range of strategies that might benefit the client. It allows lawyers to think outside the box or consider doing things differently from what worked in the past.

GenAI can assist with brainstorming across many aspects of practice. It can suggest approaches for drafting discovery demands, organizing a trial book, evaluating and preparing motion strategies, or mapping out a litigation timeline. GenAI can offer imaginative strategies for plea negotiations, propose alternative sentencing ideas, or identify creative compromises that may resonate with clients and prosecutors. It can also spotlight subtle ambiguities in a case theory, prompting attorneys to clarify or refine arguments before presenting them in court. Granted, the suggestions are sometimes implausible, but just as often, they are surprisingly useful. In this mode, GenAI’s utility is limited only by the breadth of the user’s imagination.

Even if some options are irrelevant or suboptimal, the lawyer loses nothing. The mere process of reviewing AI-generated suggestions can spark previously unconsidered defenses or inspire useful follow-up. The GenAI tool is not delivering a final plan. It is stimulating the lawyer’s own decision-making process.

Sample Ideation Prompts

Taking the above discussion from the theoretical to the practical, criminal defense lawyers can start with the following prompt patterns tailored to specific legal tasks. These examples invite the GenAI to offer multiple ideas, prioritize them, and explain its reasoning. Thought might be given to how these sample prompts can be modified to best address specific case posture or needs:

1. Plea Negotiation Strategy

Prompt Pattern

List x [number] potential plea negotiation strategies for a client charged with [insert offense] where the prosecution’s case relies heavily on [e.g., forensic evidence or co-defendant testimony]. Rank them in order of likely effectiveness and explain the reasoning behind your ranking.

This prompt encourages a range of ideas while requiring the GenAI to assess and explain the relative strengths of each suggestion, offering a helpful nudge to the lawyer’s strategic thinking.

2. Motion Practice Planning

Prompt Pattern

List x [number] pretrial motions that may be appropriate in a case involving [insert charge], where the key issue is [e.g., an unlawful search or late-disclosed evidence]. Rank them by strategic value and explain why each was placed in that order.

This structure helps attorneys quickly assess tactical options and consider arguments they may not otherwise prioritize.

3. Jury Narrative Strategy

Prompt Pattern

List x [number] ways to frame a defense narrative that emphasizes [insert theme, e.g., mistaken identity, lack of intent]. Rank them by likely persuasive impact on a jury and explain your reasoning.

This prompt helps explore rhetorical approaches and offers a quick read on what might resonate, or fall flat, in front of a jury.

These sample prompts are purposely terse, and lawyers using them should consider that the quality of GenAI output improves significantly when more case-specific information is included. While short prompts are useful for quick ideation, longer prompts that incorporate things like relevant and specific case facts, identified evidentiary issues, or jurisdictional information tend to produce more tailored and useful responses. Also, the use of persona prompting[23] in this context can significantly elevate the output’s value. However, when using open GenAI systems, legal professionals must remain circumspect about excluding sensitive or identifiable case details, always prioritizing client confidentiality and ethical obligations.

Also, when deciding how many suggestions to request (as indicated by the placeholder “x” above), users may find it helpful to constrain the GenAI. For example, asking for five options will provide a more manageable list while leaving the number open might yield broader results that require more sifting.

A significant advantage of employing GenAI in this way is the low-risk, highly productive nature of ideation. Attorneys can rapidly generate numerous ideas, discard those that do not align with their objectives, and focus on the most promising concepts. The speed and abundance of AI-generated suggestions encourage attorneys to venture into unconventional territory, exploring imaginative avenues that conventional brainstorming might never reach.

Crucially, this kind of use does not absolve the lawyer of responsibility, nor should it. Just as one would not rely on GenAI to draft a dispositive motion without review, no one should accept an AI-generated workflow or project plan without independent evaluation. But in this context, that review is easy, fast, and expected. The output is explicitly exploratory and is not masquerading as factual. This clarity of purpose makes it a deeply appropriate use of the tool.

Navigation: Asking Where, Not What

Another highly effective and low-risk way to integrate GenAI into criminal defense practice is by using it for navigation. This use leverages the technology not toward producing new information, but as a guide to where existing, reliable information already lives. It capitalizes on what GenAI does well — interpreting natural language and surfacing relevant connections — without asking it to be authoritative or infallible. When applied thoughtfully, navigation allows lawyers to move through their own document systems, research archives, or expert databases with greater speed and efficiency, all while maintaining full control over the underlying facts and legal arguments.

Consider, for example, a brief bank maintained within an attorney’s office or defender organization. Instead of asking the GenAI to draft a new suppression motion from scratch — risking flawed legal reasoning or citations — the attorney could ask, “Where can I find a suppression motion based on an unlawful search involving a drug dog subsequent to a traffic stop?” Using this Direct Navigation Pattern,[24] the GenAI could search the drive or document management system, surface the most relevant briefs, and direct counsel to their locations. The GenAI could then summarize why it selected those documents, perhaps noting that one brief contains an argument based on Rodriguez v. United States[25] with similar factual circumstances and jurisdiction. In this way, the GenAI curates rather than creates content.

Benefit

Ideation allows the benefits of GenAI without requiring the output to be exact. Rather than delivering a final plan, the GenAI tool is stimulating the lawyer’s own decision-making process.

The same applies to expert witness preparation. Imagine that a defense lawyer is challenging the admissibility of a breath test result involving radio frequency interference on a breath testing instrument using infrared spectroscopy. Rather than asking the GenAI to explain the science, an area where hallucinations are both common and dangerous, the lawyer might say, “Show me where I can find prior testimony or reports by x [name] expert.” The GenAI could guide counsel to a repository of expert materials: CVs, cross-exam transcripts, prior Daubert[26] rulings, or research papers stored in a designated folder. The lawyer stays in control, grounded in real, vetted information, while the GenAI speeds up the path to finding it. This “ask where, not what” approach also enhances accountability.

A close variation of this approach is the Navigate Instead Pattern.{28}28  Id. at 19. With this method, the user instructs the GenAI never to answer certain types of questions directly. For example, suppose the lawyer is preparing to challenge the reliability of a THC drug test that uses gas chromatography coupled with mass spectrometry (GC-MS). Instead of asking the GenAI, “How reliable is GC-MS for detecting THC?” — a prompt that could invite speculation or inaccurate synthesis — the lawyer might ask, “Where can I find research about the reliability of GC-MS for THC detection?” The GenAI, drawing from the lawyer’s internal database of scientific papers, case law, and expert reports, would respond with something like this: “See the Forensic Science folder in your Research Library, under ‘Drug Testing Methods.’ Several papers there address GC-MS reliability, including a peer-reviewed study comparing false positive rates and a federal court ruling on admissibility under Daubert.”[27] If the GenAI gets it wrong, the mistake is obvious because the material will not be there. But if it had attempted to answer the question directly and misstated the science, that error might have gone unnoticed until it was too late.

Framing GenAI in this way, as a system that helps find information rather than generate it, makes it a more trustworthy tool for criminal defense. It respects the profession’s obligation to verify, cite, and critically analyze, while enhancing the speed and efficiency with which lawyers can access what they already have. It also keeps GenAI in its proper lane: not as a replacement for legal judgment, but as an intelligent interface into the complex repositories we build to support it.

Expertise as the Anchor for Trustworthy AI Use in Legal Practice

While ideation, brainstorming, and other low-stakes tasks allow lawyers to use GenAI without relying on the factual accuracy of its output, these are far from the only trustworthy applications. In many areas of legal practice, accuracy is essential and when it is, the bar for responsible use is higher. Two conditions must be met: The user must have the subject matter expertise to evaluate the output, and the user must actively commit to doing so. Without both, GenAI is no more trustworthy than a junior associate left unsupervised. Expertise isn’t just helpful, it is foundational.

Too often, conversations around GenAI focus on what the tools can or cannot do, rather than on what the user is equipped to do with them. When an experienced litigator uses GenAI to generate a draft argument for a suppression motion, the real value is not in whether the GenAI gets everything right. The value is in how quickly it can get the lawyer 50 or 80 percent of the way there while relying on the lawyer’s subject matter expertise to verify, refine, and complete the work.

AI-generated output should be treated like a draft memo from a junior associate. It might be helpful, or even insightful, but only in the hands of someone who knows how to spot flaws, fill in gaps, and elevate it to something court-ready. If a lawyer is reviewing a draft complaint generated by AI and cannot identify a missing element of a cause of action, the lawyer is not in a position to use the tool for that task. But if the lawyer is an experienced attorney who knows exactly what to look for, the GenAI has served its purpose: giving the lawyer a head start.

Bringing this all together, consider another scenario: an advocate is exploring emerging theories of parental or third-party criminal liability in the context of mass shootings, and wants to use the GenAI for ideation. Counsel might prompt the tool to suggest areas worth researching, such as the legal limits of involuntary manslaughter for parental omissions, how courts define foreseeability and causation in third-party acts, the scope of duty based on special relationships, or whether failure to act on warning signs can constitute gross negligence. Here, counsel is not asking GenAI for definitive answers; counsel is using it as a brainstorming partner, a way to generate leads that can be investigated further. Crucially, counsel has the legal training to recognize which leads are worth following, and the practical experience to assess the quality of the sources or ideas it offers. In this sense, GenAI becomes a tool of navigation, not a source of truth, but a guide toward it.

GenAI can also help connect ideas across doctrinal boundaries. Perhaps defense counsel wants to think creatively about how the broader evidentiary rules in civil law might inform new approaches to expanding the more limited tools available in criminal discovery. Manually making those connections could take hours. With GenAI, one can surface analogies in minutes. But again, the user needs the legal acumen to assess whether those analogies hold up. If fluent in both domains, defense counsel will be able to evaluate the reasoning. If not, the lawyer might still use the output to initiate collaboration or consult reliable secondary sources to verify the ideas.

With the right level of subject matter expertise, GenAI can even assist with deeper, more fact-intensive research. At the far end of the spectrum is the use of GenAI’s deep research capabilities to save time and expand the scope of inquiry. For example:

Sample Prompt for Deep Research Regarding the Fermentation Defense

I am preparing a legal argument in support of a neofermentation defense in a DUI case with a reported BAC of .22. The blood sample was left unrefrigerated for 7 days during summer. I need scientific and forensic support for the contention that sodium fluoride in the collection vial would not be sufficient to prevent fermentation prior to testing, potentially leading to a falsely elevated BAC. Please identify and summarize peer-reviewed studies, official forensic standards, lab guidelines, and expert commentary — ranked in order of evidentiary strength — that support this claim. Prioritize materials discussing sodium fluoride’s effectiveness over time, especially in unrefrigerated or warm storage conditions.

Provided the lawyer using this kind of tool has the expertise to properly vet and evaluate the results, this is a perfectly appropriate, and ethical, use of GenAI.

But things get dangerous when lawyers use GenAI outside their wheelhouse and without guardrails. A lawyer asking GenAI to draft a firearms trust when the lawyer has never handled an estate planning matter is no different from the lawyer running someone else’s Python script[28] on his machine without knowing what it does. The lawyer does not know what is missing, what has been misinterpreted, or what unintended consequences may follow. This is where malpractice risk lives, not in the GenAI itself, but in the mismatch between the tool’s output and the user’s expertise.

The ethical use of GenAI in legal practice hinges on honest self-assessment. If an experienced attorney would not feel confident editing a junior associate’s draft in a given area, the experienced attorney should not feel confident relying on GenAI in that area either. But if the veteran attorney would, and if she does so routinely, then the same instincts that serve her well in mentoring younger lawyers can serve her here, too. Evaluate, revise, improve. In domains where a user’s expertise is strong, GenAI becomes not a shortcut to answers, but a fast track to better questions, clearer thinking, and ultimately, stronger work.

Conclusion: A Framework for Trustworthy Use

The limitations of GenAI, particularly its propensity to produce fabricated or inaccurate content, are not defects, but a predictable result of how the technology functions. Properly understood, this limitation becomes the foundation for its most effective use. In contexts where factual accuracy is not essential, GenAI’s immense capacity to creatively innovate and connect disparate concepts can lead to new approaches, novel arguments, and ultimately, more successful client outcomes.

In these settings, the technology becomes a tool for ideation, insight generation, and strategic exploration. Where accuracy is required, its use must be contingent on two conditions: the lawyer must possess the necessary subject matter expertise and must commit to independently evaluating the results. In this respect, trustworthiness is not a property of the GenAI itself, but of the professional judgment applied to its use

© 2025, National Association of Criminal Defense Lawyers. All rights reserved.

About the Author

Patrick T. Barone has cultivated a nationally respected DUI defense practice. His work integrates legal, scientific, and psychological insight, particularly in cases involving forensic evidence. A board-certified psychodramatist (TEP), he is a prolific author and national lecturer on trial advocacy, forensic litigation, and the ethical use of AI in law.

Patrick T. Barone (NACDL Member)

[1] See American Bar Association, When Legal Tech Comes of Age.

[2] What Is AI and How Can Law Firms Use It? Clio.

[3] Melissa Love Koenig, Julie A. Oseid & Amy Vorenberg, OK, Google, Will Artificial Intelligence Replace Human Lawyering? 102 Marq. L. Rev. 1269 (2019). 

[4] See Harvey AI: A Glimpse Into the Future of Legal Technology, LEX247 (Oct. 3, 2023). See also Kate Rattray, Harvey AI: What We Know So Far, Clio Blog  (last visited Nov. 19, 2023).

[5] See CoCounsel: The Legal AI You’ve Been Waiting For, casetext.com (last visited Nov. 19, 2023).

[6] See Lexis+ AI: Transform Your Legal Work, LexisNexis (last visited Nov. 19, 2023).

[7] See Westlaw Precision Now Has Generative AI, Thomson Reuters (last visited Nov. 19, 2023).

[8] See The Times They Are A-Changin’: The Rise of Generative AI in the Legal Profession, Federal Bar Association Blog, May 7, 2024. 

[9] Clio, AI-Powered Legal Practices Surge: Clio’s Latest Legal Trends Report Reveals Major Shift, (last visited Mar. 27, 2025)

[10] Harvard Law School, The Impact of Artificial Intelligence on Law, Law Firms, and Business Models, Harvard Law School Center on the Legal Profession (last visited Mar. 27, 2025).

[11] Thomson Reuters, How AI Is Transforming the Legal Profession, Thomson Reuters Legal (last visited Mar. 27, 2025).

[12] Ziwei Xu, Sanjay Jain & Mohan Kankanhalli, Hallucination Is Inevitable: An Innate Limitation of Large Language Models, arXiv (last visited Mar. 27, 2025).

[13] Victor Habib Lantyer, The Phantom Menace: Generative AI Hallucinations and Their Legal Implications (2025) (unpublished manuscript) (on file with author). 

[14] Id. at 4, 5.

[15] Id.

[16] Jeremy Laird, Anthropic Has Developed an AI ‘Brain Scanner’ to Understand How LLMs Work and It Turns Out the Reason Why Chatbots Are Terrible at Simple Math and Hallucinate Is Weirder Than You Thought, PC Gamer  (last visited Mar. 28, 2025).

[17] Id. at 6.

[18] Id. at 7, 8. 

[19] Jules White, Trustworthy Generative AI, Coursera (last visited Mar. 27, 2025).

[20] Id.

[21] Id.

[22] See Patrick T. Barone, Mastering Prompt Engineering: Advanced Techniques in AI-Powered Criminal Defense, The Champion, March/April 2025, at 20, and Patrick T. Barone, AI-Powered Advocacy: Transforming Criminal Defense Through Prompt Engineering, The Champion, January/February 2025, at 12.

[23] Id., supra note 22.

[24] Id., supra note 18.

[25] Rodriguez v. United States, 575 U.S. 348, 135 S. Ct. 1609 (2015).

[26] Daubert v. Merrell Dow Pharms., Inc., 509 U.S. 579, 113 S. Ct. 2786 (1993).

[27] Id., supra note 27.

[28] Wikipedia contributors, Python (programming language), Wikipedia: The Free Encyclopedia. (last visited Apr. 2, 2025).

First, please answer the following two questions:

1. In your opinion, what are the most important risks or problems with using AI in your future legal work?
2. In your opinion, what are the most important benefits or value of using AI in your future legal work?

Second, please answer any three of the following ten questions:

1. Ivy Grey recommends “the sandwich approach” for responsible use of AI: what do you think are the strengths or weaknesses of that approach?
2. How would you define (in your own words – not a quote from any of the readings) what it means to be “competent” with AI?
3. Explain how the duty of confidentiality could potentially be breached by the use of AI?
4. What does a lawyer have to tell their client about the lawyer’s use of AI?
5. In your opinion, why were the sanctions in Nolan v. Land of the Free appropriate or not appropriate?
6. Using Claude or ChatGPT or any other AI tool, ask the AI tool to answer one of the following questions and (a) paste in AI’s response and (b) explain why you think the answer is or is not accurate:  Are bitemarks a reliable form of evidence? Is it ethical for a lawyer to strike gay people from the jury in New York? Is it ethical for a lawyer to advise a client not to take a breathalyzer test? [or Any other legal question that you already know the answer to]
7. Take one paragraph from your research memo (or any of your Writing Reflections for this class) and put it into Claude or ChatGPT or any other AI tool and ask the AI tool to revise it. Please (a) paste in the original and revised paragraphs and (b) explain why you think the revision is or is not an improvement.
8. What steps does a lawyer need to take to verify AI-generated work? (This question was generated by AI.)
9. What obligations do attorneys have when submitting AI-assisted work to courts? (This question was generated by AI.)
10. Explain how you feel about the idea of your professors using AI to generate test questions or lessons plans for their courses?