Practice Notes: Artificial Intelligence as a Judge Advocate’s Force Multiplier

(Source: Chief DIgital & Artificial Intelligence Office, U.S. Department of War)

Practice Notes

Artificial Intelligence as a Judge Advocate’s Force Multiplier

---

By Captain Bridget R. Reineking

The commanders whom Army lawyers advise are increasingly making decisions at machine speed, in domains shaped by adversaries who themselves use artificial intelligence (AI). The judge advocates (JAs) who cannot keep pace with that tempo do not remain neutral observers of the technology; they become a brake on operations. This is the practical meaning of the Secretary of War’s January 2026 directive that the Army become an “AI-first” warfighting force, emphasizing speed, experimentation, and operator-level adoption.1 The mandate does not frame AI as a future aspiration or a niche technical capability.2 It casts AI as a present operational reality—one that demands bold, informed, and creative adaptation.3

This institutional posture matters for Army lawyers. JAs do not operate outside the tempo of modern warfare; they advise within it. As operational timelines compress and commanders confront increasingly complex environments and directives to employ disruptive technologies, their counselors must keep pace. AI—and large language models (LLMs) in particular—offer a means of increasing legal speed, consistency, and analytical depth.4 With training and willingness (and a certain degree of self-study), members of the Corps can and will develop the competence to employ AI deliberately and responsibly. In this environment, abstention is not neutrality. It is a decision to fall behind.

The institutional architecture for that engagement is already in place. The Department of War (DoW) has approved AI tools for use in Government environments,5 and the Judge Advocate General’s (JAG) Corps Interim Artificial Intelligence Governance Policy declares that the Corps “will exploit AI responsibly,” encouraging Army JAs to engage with authorized LLMs as a matter of professional competence.6 This policy direction reflects a clear strategic vision: properly deployed, LLMs are force multipliers.7 For supervisory attorneys in particular, authorized AI systems offer a means to scale legal judgment across dispersed formations without diluting individual responsibility or accountability. The task ahead is not to automate legal judgment but to augment it, freeing JAs to focus cognitive resources on the analysis, counsel, and advocacy that remain irreducibly human.8 But policy permission and a force-multiplier framing are not, by themselves, competence. Using these tools well—and avoiding failures that have already embarrassed lawyers in civilian practice9—starts with understanding what they are.

Master Your Tool: LLMs and How They Function

Like any professional instrument, LLMs reward users who understand both their capabilities and their limitations. An LLM is a type of “foundation model”—an AI model trained on vast amounts of data that can be adapted to various tasks and applications.10 Foundation models encompass multimodal systems like Google’s Gemini (which processes text, images, and audio),11 specialized generative models such as OpenAI’s DALL-E (which focuses on text-to-image generation),12 in addition to LLMs. An LLM can “recognize, summarize, translate, predict, and generate text and other forms of content based on knowledge gained from massive datasets.”13 These models can be general-purpose or fine-tuned for specialized tasks, including in domains like law and medicine.14 

The JAG Corps Interim AI Governance Policy is not an invitation to experiment. It is a professional imperative. LLMs offer JAs a means to multiply effectiveness without compromising the judgment, ethics, and accountability that separate trusted counselors from legal processors—but tools require mastery.

Many readers familiar with ChatGPT may not understand the underlying foundation model (an LLM created by OpenAI) or its information processing mechanisms. Unlike search engines, traditional LLMs do not retrieve specific information from existing databases.15 Instead, these algorithms are “trained” to identify linguistic patterns by processing billions of pages of text, generating responses based on statistical probabilities rather than direct retrieval.16 After an LLM is trained on a sufficiently large and diverse body of works—hundreds of billions to trillions of tokens17 derived from pages of news reports, books, essays, stories, and similar documents—the model has encountered a given word so often, across so many contexts, that it can predict statistically likely word sequences in response to prompts.18 Many (but not all) modern LLMs now also incorporate retrieval-augmented generation (RAG), which lets the model pull accurate, up-to-date information from curated sources when generating a response, improving factual accuracy while preserving fluent text generation.19

Two characteristics of this architecture matter for attorneys. First, LLMs do not “know” facts in the human sense.20 An LLM does not reason deductively, form intent, or understand truth.21 Instead, it generates language that resembles expert discourse because it has been trained on large volumes of expert discourse. When asked to operate beyond reliable context, it may confidently invent citations or misstate facts—the phenomenon referred to as “hallucination,” which is a predictable consequence of probabilistic text generation rather than a flaw unique to AI.22 Second, an LLM’s output is bounded in large degree by what was in its training data and shaped by the quality of a prompt.23 The structure, specificity, and framing of a prompt determine the usefulness of the output.24 A model that has seen many appellate opinions will draft persuasively in that genre; a model that has seen relatively little of, say, the Defense Federal Acquisition Regulation Supplement, will draft less reliably in that domain.

For members of the Army JAG Corps, that second point has particular force. With few exceptions, as of this writing, JAs are not currently positioned to train LLMs or other AI models for daily operations. But JAs may leverage authorized models available through the Department of War’s GenAI.mil platform (GenAI) and the Army Enterprise LLM Workspace, which is powered in part by Ask Sage.25 The practical levers available to all JAs, therefore, are (i) prompts themselves and (ii) supporting material uploaded with a given prompt. Both function as a form of in-context conditioning: they tell the model not only what to do but also, by supplying authoritative context, narrow the statistical space from which it draws.26 This is why an uploaded regulation or policy memo can dramatically improve output quality over an unsupported question.

It is worth pausing on a contrary view. Certain industry commentary suggests that prompt design matters less than it once did, because newer models infer user intent more reliably.27 That observation describes the experience of users running general-purpose queries on the most advanced commercial models. JAs operate on a constrained set of authorized platforms and cannot fine-tune the models they use. In that environment, careful prompting becomes more important.

Critically, these limitations do not displace the supervisory role of lawyers; they reinforce it. JAs are already trained to verify authorities, assess relevance, and exercise professional judgment under conditions of uncertainty, all while supervising subordinate work product.28 Properly used, LLMs accelerate those functions by reducing the administrative and cognitive burden associated with information synthesis, drafting, and organization. The question, then, is not whether these systems can replace legal judgment—they cannot—but how they can be employed tactically across the demands of military legal practice.

Five Tactical Applications Across JA Billets

For JAs, that tactical employment occurs in the daily friction of practice, as time constraints, operational tempo, and dispersed formations shape legal support. The following applications illustrate how LLMs can be employed right now across common JA roles to amplify speed and analytical efficiency in a manner that is consistent with application guidance and limitations on currently available platforms.29

Continuity Books Between Transitioning JAs

One of the most persistent challenges in military legal practice is continuity. Frequent personnel rotations disrupt institutional memory, particularly in complex billets where local practice, commander preferences, and recurring legal issues matter. Traditional continuity books—often static binders or shared drives—rarely capture nuance or remain current.

LLMs offer a means to transform continuity from a static artifact into a living knowledge base. Outgoing JAs can use approved AI tools to create clear, searchable products with substantially less effort than building traditional continuity binders. Incoming JAs can then query this material to understand not only what tasks exist, but how and why they are performed. The work lies in selecting the most helpful and timely authorities, and redacting, as necessary, precedent legal reviews or legal products. Below are examples of prompts a JA can input into GenAI.

Create with Prompt on GenAI: Review the attached continuity materials and redacted legal products. Generate a structured continuity guide for the incoming administrative law attorney organized into the following sections: (1) Recurring Legal Issues (with frequency and resolution patterns); (2) Commander Decision-Making Preferences (including risk tolerance and preferred briefing formats); (3) Common Pitfalls and Lessons Learned; (4) Key Points of Contact and Their Roles; and (5) Seasonal or Cyclical Legal Requirements. Use only information explicitly contained in the provided materials—do not infer, extrapolate, or add context. Flag any areas where the provided materials are incomplete or unclear.

Augment with Workbook on Ask Sage: Use the same prompt (above) and then save the output as a memo titled “Core Continuity Guide.” Create the Workbook on Ask Sage, classify it as CUI, and add the continuity memo and all the reference materials (redacted as necessary). An incoming attorney can open the Workbook and can leverage it to ask continuity questions (“how does this command typically handle first-time DUI offenses?”) or (“what deadlines should I expect this quarter?”).

Used this way, LLMs preserve institutional knowledge and allow continuity to scale beyond individual personalities, reducing friction during transitions and improving legal support without substituting for human judgment.

Quick-Turn Staff Judge Advocate Support

Command-advising attorneys are frequently asked to provide a “quick take” on changes (or proposed changes) to policies or regulations, synthesizing long documents or comparing expansive policy memos under time pressure—sometimes with only minutes before an office call. For instance, a JA may be asked, “How does this new policy change our PT requirements?” or “What are our obligations under this new drone acquisition regulation?” LLMs can assist quickly: upload one or multiple documents and, within seconds, the model will identify changes in authorities, structure them in a readable format, and even frame discussion points.

Rather than drafting final advice, the model functions as an analytical amplifier; it helps the attorney prepare calmly for a meeting in lieu of frenetic speed-reading. Any outputs by the LLM would, naturally, require verification, but being prepared with a 90 percent solution and a solid “I’ll follow up to confirm after verifying a few points more thoroughly” is preferable to arriving empty-handed.

Example Safe Prompt 1: Compare the original policy document (Document X) with the revised policy document (Document Y). Provide: (1) A line-by-line identification of all substantive changes, noting the specific section/paragraph where each appears; (2) Changes categorized by type (new requirements, deleted provisions, modified standards, clarified language); (3) Potential rationales for each major change based solely on the text itself; (4) Questions or ambiguities created by the changes that may require clarification; and (5) Areas where the revised policy creates new decision points for commanders. Present findings in a table format for quick reference. Do not make recommendations or draw legal conclusions.

A Soldier integrates AI into military workflows. (Credit: SFC Shane Smith).

Example Safe Prompt 2: Analyze the following proposed command social media policy against AR 600-20 (Army Command Policy), DoDI 1325.06 (Handling Dissident and Protest Activities), and First Amendment considerations as applied to Service members. The proposed policy includes three provisions: (1) mandatory command approval before posting photos in uniform; (2) prohibition on political commentary while identifying as a Service member; and (3) mandatory disclosure of personal social media accounts to the chain of command.

For each provision, provide: (1) Relevant regulatory citations and legal standards; (2) Potential legal issues or conflicts with existing guidance; (3) Precedent from military appellate courts addressing similar restrictions; (4) A compliance analysis framework organized by: (a) regulatory authority, (b) First Amendment limitations, (c) narrow tailoring requirements, and (d) less-restrictive alternatives. Identify which provisions appear most legally defensible and which present elevated risk. Do not provide a final recommendation, but structure the analysis so I can advise the commander on risk levels and potential modifications.

Note the precision in each prompt. Counter-intuitively, slowing down to craft careful inputs saves time.30 A rushed prompt often generates unhelpful outputs.31 The appropriate mental model is that the attorney is tasking a literal-minded junior attorney who will do exactly what you say and nothing more. Specify what you need, why it matters, and how to structure the output.

Terminology and Issue Spotting

In the cyber domain and elsewhere, our clients are leveraging disruptive technologies and using schemes of maneuver that are jargon-heavy and require advanced degrees to fully comprehend. LLMs allow lawyers to dual-hat as technologists. These models—in the operational space, many of which are classified at the TS level and higher—allow JAs to review concepts of operation (CONOPs) and other proposals carefully. They enable a JA to catch legal triggers early because only in fully understanding an exploit, technical modification, or scheme of maneuver can a JA properly conduct a LOAC review, for example, or advise how some modification to a unit’s drone can be financed.

Example Safe Prompt: Review the following technical description of a cyber capability: [INSERT DESCRIPTION]. Translate this into plain language that explains: (1) What the capability does in functional terms; (2) What systems or networks it affects; (3) What the intended operational effect is; and (4) Any potential unintended or collateral effects mentioned. Use analogies to physical-world actions where helpful (e.g., “this is like cutting a phone line” rather than “this disrupts packet routing”).

Here, the model accelerates awareness without replacing analysis. To verify this output, a JA can have an informed client conversation.

Contract Law: Solicitation Document Review

Installation JAs review numerous contract actions, including construction projects, service contracts, and supply acquisitions. The Federal Acquisition Regulation (FAR) and its supplements create a dense web of requirements that JAs must accurately navigate to manage these actions. Contracting officers can leverage LLMs to identify potential issues before solicitation release.

The models can perform initial compliance checks against standard FAR provisions, though attorneys must verify current regulatory language, assess agency-specific requirements, and redact information as necessary before the LLM’s review.

Sample Prompt: Review the following draft statement of work (SOW) for a base maintenance services contract with these parameters:

• Services: janitorial, grounds maintenance, HVAC repair, pest control

• Estimated value: $2.3M annually
• Contract structure: 5-year base period with two 1-year option years
• Acquisition type: Service contract subject to FAR Part 37

Analyze the SOW and provide:

1. Language Issues:

• Ambiguous or vague performance standards that lack objective measurements

• Conflicting requirements between different SOW sections
• Terms that are undefined or could be interpreted multiple ways
• Missing performance metrics or quality assurance surveillance plan references

2. Required FAR Clauses (cite by number and title):

• Mandatory clauses for service contracts over simplified acquisition threshold

• Service Contract Labor Standards (formerly Service Contract Act) provisions
• Clauses specific to multi-year contracts with options
• Any specialized clauses for HVAC work, pest control, or safety-sensitive services

3. Labor Standards Compliance:

• Wage determination requirements under 41 U.S.C. § 6707

• Compliance with FAR 22.10 (Service Contract Labor Standards)
• Health and safety requirements for service employees on Federal property
• Potential prevailing wage issues

4. Risk Areas for Disputes:

• Scope creep vulnerabilities between the base period and option years

• Insurance and liability provisions (identify what’s missing)
• Performance standards that may be impossible to measure objectively
• Government-furnished property or equipment issues
• Transition-in/transition-out requirements

Organize findings by priority: Critical (must fix before solicitation), Important (should fix to reduce protest risk), and Advisory (best practice improvements). Provide specific SOW section references for each issue identified. Here is the SOW text: [INSERT TEXT, redacting content as necessary].

Once again, the mindful scoping and framing of the prompt will shape the usefulness of the output. Based on the carefully crafted prompt, the LLM can review the SOW and flag vague performance standards, missing safety requirements for HVAC work, the need for Service Contract Labor Standards32 provisions, wage determination requirements, potential conflicts between base year and option year scopes, and insurance requirements, to name a few. The attorney can then verify these observations against current FAR provisions, check DFARS supplements, consult recent GAO protest decisions, and provide the contracting officer with verified recommendations. The LLM handles the initial document review; the attorney applies professional expertise to confirm and refine.

A Soldier integrates AI into military workflows. (Credit: SGT Kayden Bedwell).

“Legalese” Translators for Commanders and Staff

After JAs spend significant time analyzing law, they often need to translate legal concepts into language commanders can use to make decisions. A fifteen-page memo on the First Amendment implications of restricting Soldiers’ social media use lacks usefulness if the brigade commander needs a two-minute brief before a meeting with staff. The translation itself may require time and several iterations to be distilled into actionable guidance. But in a pinch, a JA can use the LLM as a “translation engine.”

Sample Prompt: I’ve completed a legal analysis on whether National Guard forces in Title 32 status can conduct vehicle checkpoints in support of a counter-drug operation. My analysis concludes they can, but only under specific conditions. Below is my legal reasoning [PASTE YOUR ANALYSIS—no names, no operational details, just the legal framework]. Generate three versions of this analysis: (1) Executive Summary (3 bullet points)—Bottom-line answer for the commander who has 60 seconds; (2) Decision Brief Format (1 page, formatted for a PowerPoint slide)—Issue, applicable law, analysis, recommendation, and risk factors; (3) Talking Points (5-7 points)—What I should say during a stand-up brief to explain this to the commander or his staff, anticipating likely questions. Keep all versions at a college reading level. Use active voice. Maintain legal accuracy while improving accessibility.

Bear in mind that a prompt like this can be stored and, with only small changes, reused by copying and pasting into an LLM, with an uploaded legal memo, for quick output. The JA can master oral delivery of the bottom line up front (BLUF) while already prepared to handle the nuanced questions that may follow.

Protocols for Safe and Professional Use

Safe use of LLMs within Government-
approved platforms ultimately rests on four principles: (1) data discipline, (2) task alignment, (3) verification, and (4) transparency. As highlighted in the JAG Corps Interim Artificial Intelligence Governance Policy, JAs should never input classified or privileged information into unauthorized systems.33 Commercial tools like ChatGPT, Claude, or Gemini—while powerful—are inappropriate for Government work absent specific approval. These platforms may retain user inputs for training purposes, potentially exposing sensitive information.34

Even on approved platforms, attorneys must exercise classification and privacy discipline, sanitizing, as appropriate, personally identifiable information or operational details that could compromise security if disclosed.35 When drafting prompts, consider anonymizing facts: replacing real names with generic identifiers, substituting specific locations with broader geographic references, modifing dates and times while preserving relevant temporal relationships, and removing unit designations and operational details.

At all times, prompts should align with tasks the model is suited to perform rather than final judgment.

Output Verification and Professional Responsibility

Every JA remains professionally responsible for all work product, regardless of AI assistance.36 An LLM is analogous to a highly capable but fallible research assistant.37

Ideally, verification steps include: independently verifying all statutory and regulatory citations against current codifications, cross-checking factual assertions against reliable sources, assessing whether legal analysis reflects current law and recent developments, and evaluating whether the output fully addresses the legal question posed.

JAs would be advised to maintain documentation of AI-assisted work. It is best practice to note which tasks involved LLM assistance, preserve the prompts used to generate outputs, record the verification steps performed, and be prepared to explain the AI’s role if questioned. Some jurisdictions require disclosure of AI use in court filings.38 While military practice generally lacks such requirements, transparency is prudent.

Competence and diligence obligations also require staying current with AI developments. As these tools evolve rapidly, JAs should participate in available AI training programs, experiment with approved platforms in low-stakes contexts to build proficiency, monitor emerging guidance from the American Bar Association, state bars, and military authorities, and share best practices with colleagues to elevate collective competence.39

Finally, members of the Corps must recognize what AI cannot do. LLMs lack human judgment. They cannot read confusion on a commander’s face, assess witness credibility in real-time, weigh equities, sense a commander’s risk tolerance, or recognize when pushing a legal argument will cost you a client relationship. AI enhances human capabilities—it does not substitute for professional expertise that separates competent advocates from exceptional ones. The attorney who grasps this distinction harnesses AI’s power while avoiding its most catastrophic failures.

Conclusion

The JAG Corps Interim AI Governance Policy is not an invitation to experiment. It is a professional imperative. LLMs offer JAs a means to multiply effectiveness without compromising the judgment, ethics, and accountability that separate trusted counselors from legal processors—but tools require mastery. A generation from now, military legal historians will scrutinize whether the Corps led the AI transformation or was dragged through it. The answer depends on decisions made today, in every office, by every member of the Corps. TAL

---

CPT Reineking is the Legal Advisor to the Joint Force Headquarters-Cyber (Army) at Fort Gordon, Georgia.

She regularly advises clients on the development and deployment of classified AI tools and provides training to external members of the Joint Force on the use of AI models and the development of the same. She is engaged in USCYBERCOM planning efforts regarding AI integration and is a member of the JAG Corps Information Technology Operational Planning Team (OPT).

---

Notes

1. Memorandum from Sec’y of War to Sr. Pentagon Leadership et al., subject: Artificial Intelligence Strategy for the Department of War 1, 6 (Jan. 9, 2026) [hereinafter SecWar Memo].

2. See id.

3. See id. at 4–5.

4. See generally Thomson Reuters Inst., 2025 Generative AI Professional Services Report (2025) (describing that generative AI accelerates repetitive tasks and allows for deeper research into proprietary, trusted content in the legal profession).

5. SecWar Memo, supra note 1, at 2-3; C. Todd Lopez, Hegseth Introduces Department to New AI Tool, U.S. Dep’t of War (Dec. 9, 2025), https://www.war.gov/News/News-Stories/Article/Article/4355797/hegseth-introduces-department-to-new-ai-tool https://perma.cc/XS2N-U7WW.

6. Pol’y Memorandum from Deputy Judge Advoc. Gen., U.S. Army, to Judge Advoc. Legal Servs. Pers., subject: DJAG Policy Memorandum 26-01 – Interim Artificial Intelligence Governance Policy, para. 1 (Feb. 10, 2026) [hereinafter DJAG Pol’y Memo 26-01].

7. See id.; Colonel Ryan A. Howard, Forging the Bimodal Judge Advocate: Human-Machine Integration and the Future of the JAG Corps, Army Law., no. 4, 2025, at 2, 4.

8. See Howard, supra note 7, at 9 (“AI should be treated as a cognitive teammate, performing tasks it excels at: collecting, analyzing, synthesizing, and drafting with speed and consistency.”).

9. See, e.g., Emily Preston, AI Hallucination Sanctions Grow as Judges Get Punitive, Bloomberg L. (Apr. 7, 2025), https://news.bloomberglaw.com/bloomberg-law-analysis/analysis-ai-hallucination-sanctions-grow-as-judges-get-punitive https://perma.cc/8TJU-DHTF; Mike Scarcella, US Appeals Court Fines Lawyers $30,000 in Latest AI-Related Sanction, Reuters (Mar. 16, 2026), https://www.reuters.com/legal/litigation/us-appeals-court-fines-lawyers-30000-latest-ai-related-sanction-2026-03-16 https://perma.cc/338E-W3RX.

10. See generally Rishi Bommasani et al., On the Opportunities and Risks of Foundation Models, arXiv, Cornell Univ., no. 2108.07258 (2021), https://arxiv.org/abs/2108.07258 (describing creation of foundation models and uses); Jon Schmid et al., RAND Corp., Evaluating Natural Monopoly Conditions in the AI Foundation Model Market (2024) (describing foundation models as a class of AI models “trained on large and diverse datasets and capable of performing many tasks”).

11. Sundar Pichai & Demis Hassabis, Introducing Gemini: Our Largest and Most Capable AI Model, Google (Dec. 6, 2023), https://blog.google/technology/ai/google-gemini-ai https://perma.cc/CZL9-NAPR.

12. See DALL·E: Creating Images from Text, OpenAI (Jan. 5, 2021), https://openai.com/research/dall-e https://perma.cc/M6NX-SHH2 (last visited Apr. 16, 2026).

13. Angie Lee, What Are Large Language Models Used For?, NVIDIA (Jan. 26, 2023), https://blogs.nvidia.com/blog/2023/01/26/what-are-large-language-models-used-for https://perma.cc/B79L-YQJT.

14. See Paul Ohm, Focusing on Fine-Tuning: Understanding the Four Pathways for Shaping Generative AI, 25 Colum. Sci. & Tech. L. Rev. 214, 216 (2024).

15. See Murray Shanahan, Talking About Large Language Models, Commc’ns of the ACM, Feb. 2024, at 68; Matthew Burtell & Helon Toner, The Surprising Power of Next Word Prediction: Large Language Models Explained, Part 1, Ctr. for Sec. & Emer. Tech. (Mar. 8, 2024), https://cset.georgetown.edu/article/the-surprising-power-of-next-word-prediction-large-language-models-explained-part-1 https://perma.cc/H49F-QJWD.

16. See Shanahan, supra note 15.

17. See Tom B. Brown et al., Language Models Are Few-Shot Learners, arXiv, Cornell Univ., no. 2005.14165 (2020), https://arxiv.org/abs/2005.14165 ; Alex Hughes, ChatGPT: Everything You Need to Know About OpenAI’s GPT-4 Tool, BBC: Sci. Focus (Sep. 25, 2023), https://www.sciencefocus.com/future-technology/gpt-3 https://perma.cc/ZL83-PKLT.

18. See Burtell & Toner, supra note 15.

19. See Patrick Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, arXiv, Cornell Univ., no. 2005.11401 (2021), https://arxiv.org/abs/2005.11401 ; Raghav, Beyond ChatGPT: How Retrieval-Augmented Generation (RAG) Is Shaping the Future of AI Applications, Medium (July 29, 2024), https://medium.com/@raghav_47115/beyond-chatgpt-how-retrieval-augmented-generation-rag-is-shaping-the-future-of-ai-applications-f1564a620020 https://perma.cc/QK8J-MMJJ.

20. See Luciano Floridi et al., What Kind of Reasoning (If Any) Is an LLM Actually Doing? On the Stochastic Nature and Abductive Appearance of Large Language Models, arXiv, Cornell Univ., no. 2512.10080 (2025), https://arxiv.org/abs/2512.10080 (“LLMs produce plausible hypotheses, simulate commonsense reasoning, and provide explanatory answers without grounding them directly in truth, semantics, verification, or understanding, and without any abductive reasoning.”).

21. See id.

22. Olga Zern, Hallucination in Large Language Models: A Friendly Insight, Medium (Feb. 1, 2024), https://medium.com/@olga.zem/hallucination-in-large-language-models-a-friendly-insight-49abe22a3d77 https://perma.cc/3887-BKPD.

23. See, e.g., Melanie Sclar, Yejin Choi, Yulia Tsvetkov & Alane Suhr, Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design, arXiv, Cornell Univ., no. 2310.11324 (2024), https://arxiv.org/abs/2310.11324 ; Strategies for Managing Prompt Sensitivity and Model Consistency, PromptHub (Oct. 23, 2025), https://www.prompthub.us/blog/strategies-for-managing-prompt-sensitivity-and-model-consistency- https://perma.cc/9TJ3-XDSV; see Bommasani et al., supra note 10, at 3 (observing that “any flaws in the model are blindly inherited by all adapted models,” and discussing how downstream task performance is shaped by, and limited by, the underlying training data).

24. See generally Federico Errica, Giusseppe Siracusano, Davide Sanvito & Roberto Bifulco, What Did I Do Wrong? Quantifying LLMs’ Sensitivity and Consistency to Prompt Engineering, arXiv, Cornell Univ., no. 2406.12334 (2024), https://arxiv.org/abs/2406.12334 .

25. Army Launches Army Enterprise LLM Workspace, the Revolutionary AI Platform That Wrote This Article, Army.mil (May 15, 2025), https://www.army.mil/article/285537/army_launches_army_enterprise_llm_workspace_the_revolutionary_ai_platform_that_wrote_this_article https://perma.cc/4V33-VZUW; The War Department Unleashes AI on New GenAI.mil Platform, Dep’t of War (Dec. 9, 2025), https://www.war.gov/News/Releases/Release/Article/4354916/the-war-department-unleashes-ai-on-new-genaimil-platform https://perma.cc/RA26-DTN6.

26. See Brown et al., supra note 17 (demonstrating that LLMs can adapt their outputs based on context supplied in the prompt without modification to model weights—the property commonly described as “in-context learning”).

27. See, e.g., Bernard Marr, Why Prompt Engineering Isn’t The Most Valuable AI Skill In 2026, Forbes (Mar. 10, 2026), https://www.forbes.com/sites/bernardmarr/2026/03/10/why-prompt-engineering-isnt-the-most-valuable-ai-skill-in-2026 https://perma.cc/5A5T-XMNT; Prompt Engineering Is Mostly Dead in 2026. Here’s What Replaced It., Dev Comm. (Apr. 18, 2026), https://dev.to/gabrielanhaia/prompt-engineering-is-mostly-dead-in-2026-heres-what-replaced-it-433b https://perma.cc/U397-PUB9.

28. See U.S. Dep’t of Army, Regul. 27-26, Rules of Professional Conduct for Lawyers rr. 5.1, 5.3 (26 Mar. 2025) [hereinafter AR 27-26].

29. JAG Corps personnel must not upload CUI, PII, PHI, or privileged material, client information, or sensitive/classified data into any GenAI tool unless the platform is explicitly authorized for that type of data. As of this writing, and consistent with DJAG Policy Memorandum 26-01, very few JAG Corps personnel have access to models to which they can upload client information in a secure manner or upload classified data; GenAI is authorized for CUI but not yet clearly authorized for the input of privileged information. See DJAG Pol’y Memo 26-01, supra note 6, para. 4.

30. See, e.g., Alara W, How a Good Prompt Could Save Time, Energy, and $$$—The Efficiency Tricks Not Discussed Often, Medium (Jan. 6, 2026), https://pub.towardsai.net/how-a-good-prompt-could-save-time-energy-and-the-efficiency-tricks-not-discussed-often-ef5280fd38e6 https://perma.cc/6EP6-AVAY.

31. See id.

32. See FAR pt. 22.10 (2026).

33. DJAG Pol’y Memo 26-01, supra note 6, para. 4.b.

34. See, e.g., Terms & Policies, OpenAI, https://openai.com/policies https://perma.cc/A2CP-SE5X (last visited May 27, 2026) (prompts are retained to “improve and train” models).

35. See generally U.S. Dep’t of Def., 5200.01, DoD Information Security Program: Overview, Classification, and Declassification, vol. 1 (Jan. 17, 2025) [hereinafter DoDM 5200.01] (classification and handling requirements); 5 U.S.C. § 552a (Privacy Act). In October 2023, the Chief Digital and Artificial Intelligence Officer issued his “Interim Guidance on the Use of Generative Artificial Intelligence.” The memo is marked CUI and not publicly available, but relevant Government personnel should seek it before utilizing commercial AI.

36. See Model Rules of Pro. Conduct r. 1.1 (A.B.A. 2020) (competence); id. r. 5.1 (responsibilities of partners, managers, and supervisory lawyers); id. r. 5.3 (responsibilities regarding nonlawyer assistance); DJAG Pol’y Memo 26-01, supra note 6, para. 4(a).

37. See Howard, supra note 7, at 8 (“[AI] cannot possess a professional ethic; it does not have moral principles or values that guide behavior. The introduction of autonomous [AI] systems will heighten, not reduce, the moral obligations of human lawyers.”).

38. See, e.g., Local Civ. R. N.D. Tex., LR 7.2(f) (requiring disclosure of generative AI use in briefs).

39. See AR 27-26, supra note 28, r. 1.1 cmt. 7 (“To maintain the requisite knowledge and skill, a lawyer should keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology, engage in continuing study and education, and comply with all continuing legal education requirements to which the lawyer is subject.”).