(Credit: Alexander Limbach-stock.adobe)
Practice Notes
Generative Pre-Trained Transformers and the Department of Defense’s Own Generative Artificial Intelligence Large Language Model
By Lieutenant Colonel Thomas S. Hong
This commentary makes the case that the Department of Defense (DoD) needs to create its own generative artificial intelligence (GAI) large language model (LLM) similar to ChatGPT.1 Before digging in more thoroughly, however, I must start with a full disclosure2 that this commentary was not written with the help of any generative pre-trained transformer (GPT) chatbot or other GAI LLM.3 Given GPT technology’s uncanny ability to create or improve written prose on many subjects, it was certainly tempting to run some ideas through the free version of ChatGPT.4 However, pumping the brakes on the impulse, I decided that I enjoy the mental spectacle of my own thoughts battling it out in my own headspace. Some of those thoughts get kicked with a roundhouse and stuffed in the recycling bin, while others are coddled like rosy-cheeked newborns who make a parent’s heart swell with pride. But to be frank, the temptation to use a GAI LLM tool is strong because time is a precious resource and GPT chatbots like ChatGPT do one thing better than me: they do not procrastinate when prompted to write out a coherent thought. It seems to move out with alacrity and, in seconds, spits out impressive text on general or specialized topics prompted. I will readily admit, it also does so better than most humans when accounting for speed, coherence, and writing mechanics.5
I will also disclose that this introduction survived the mental recycling bin because I thought it could be a discrete way to show that no GAI LLM was used in its drafting due to its first-person narrative.6 I can imagine there will still be retorts of doubt since newer chatbots like GPT-4 are becoming more human-like in style, word choice, creative prose, and can effortlessly provide a first-person commentary with a unique style. However, if some sentences in this commentary appear to be awkwardly constructed, note these were deliberately left that way as indicia of humanness or, at least, the quirkiness of the author.7 Although, it is easy to conceive that newer GPT chatbots may mimic less-than-perfect writing, when prompted, that could be passed off as solely human-created.8
From this initial disclosure and burgeoning debate, this commentary turns to the topic at hand: the future of military law practice in the age of GPT chatbot prevalence. More specifically, the focus here is on whether all attorneys, paralegals, and other legal professionals like judges and law professors should deeply contemplate what GPT chatbots will do to, or for, the legal profession. Within this extended thought process, one of the tantalizing questions to address is, “Does the DoD need its own GPT chatbot?”9 For an emerging group of legal professionals, the answer is a resounding “Affirmative!” However, if you disagree, I look forward to hearing from you after you have finished this commentary and considered the nascent thesis developed here.
Curating a path to an affirmative response for a DoD GPT chatbot can be done with a review of technological breakthroughs in artificial intelligence (AI). One good starting point is to recall when a chess-playing AI named Deep Blue beat a world chess champion in 1997.10 When this made the news, one can imagine legal professionals at the time were only mildly amused as their profession did not depend on how well they did on a chess board.11
Some also probably registered similar reactions when AlphaGo beat a Go grandmaster in 2016 and when DeepMind beat professional gamers in a strategy video game in 2019.12 Again, Go and video game expertise do not impact legal professionals in the workplace. Perhaps forward-thinking legal professionals raised more eyebrows when IBM’s Watson beat brainiac Jeopardy champions in 2011.13 However, it was unlikely that legal professionals would put down the coffee pot with an incredulous look when the news spread.14
In contrast, when GPT chatbots almost passed the bar exam in 2022 and GPT-4 finally did in 2023, we got an eye-opening introduction to the capabilities of GPT-powered chatbots, a technology that looks capable of replacing large numbers of legal professionals.15 After reading numerous articles, conversing with various people, and testing out ChatGPT personally, many were convinced that we, as legal and knowledge professionals, must get smart on this technology and get energized now.16 In fact, Model Rule 1.1 of the American Bar Association’s (ABA’s) Model Rules of Professional Conduct states that “[t]o 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.”17 Given the rapid rollout of various LLMs in the past year, it is easy to say with confidence that GPTs are now considered highly relevant technology for lawyers.
For purposes of this commentary, being “energized” means researching and advocating for the DoD to develop its own GPT for legal professionals and all DoD knowledge professionals.18 As the DoD starts to take ever wider steps to become energized, legal professionals can be thought leaders who come to the table from the users’ perspective with a desire to be valuable members of the development team.19 They can and should be brainstormers, designers, trainers, testers, and overall change agents in the specialized area of GPT chatbots. After all, legal professionals are in the language and knowledge business, and a DoD GPT will raise the language and knowledge capacities of these personnel to a higher degree. How much “higher” will largely depend on the size, quality, and usefulness of the datasets and how powerful the AI engine behind the DoD GPT will be.20
Given our dual training and mastery of both civilian and military law and regulations, coupled with access to vast amounts of data to train a DoD GPT, including controlled unclassified information (CUI), DoD legal professionals can become skilled users of a DoD GPT.21 Once they become highly proficient users, legal professionals can also assist other, non-legal DoD knowledge professionals in the use of a DoD GPT in their respective fields. As a professor quipped during a legal research lecture at the U.S. Army’s Judge Advocate General’s Legal Center and School (TJAGLCS), legal professionals are the undisputed experts at searching through other staff sections’ regulations and telling them the answer when asked a non-legal question couched as a legal one.22 With a DoD GPT that has the capabilities of similar GPTs in existence or in development, the turnaround time from “legal” may be drastically reduced.23 This would be just one of the many beneficial use cases for a DoD GPT. The next consideration is how we can bring this capability to the DoD now.
For legal professionals to take on this unofficial yet important goal to join or even lead the effort to bring about a DoD GPT, teamwork and collaboration will be essential. To further explore and develop this multifaceted and multilayered idea, TJAGLCS formed the TJAGLCS AI Study Group (TAISG).24 Volunteer faculty, center staff, and resident students—both long- and short-term students at TJAGLCS—formed this informal and open study group. The purpose of the group is to provide a digital and in-person venue for legal professionals at TJAGLCS to discuss all topics related to AI and to collaborate on projects that may benefit the Judge Advocate General’s (JAG) Corps specifically and the DoD generally. The study group members have shared articles and resources to stay informed on the latest developments in AI and how they may apply to the study and practice of military law. Moving forward, this group has discussed the desire for further action. Although readers may see a need for additional lines of effort, the initial seven areas discussed by TAISG members are outlined below.
1. Self-Directed Non-Technical Education in AI and Its Myriad Subdisciplines
Fully aware that many legal professionals in the U.S. Army JAG Corps, other Services’ JAG Corps, and the DoD have studied AI technologies before the advent of ChatGPT, the immediate goal should be to drastically enlarge the number of legal professionals who will self-study AI and its application to DoD legal practice. Although AI for use in the military can take many forms and applications (such as machine learning (ML) and machine vision for military intelligence or targeting purposes), the ideal area for growth would be in the study of technology behind what makes LLMs and GPT-powered chatbots capable of passing the bar exam and possibly becoming an AI legal expert. This desire for more students of AI technology from the legal profession is to build up a large group of informed users who, eventually, can act as expert users and advisors to AI scientists, engineers, and technicians.
(Credit: Iconimage-stock.adobe)
For those seeking external motivation to start a program of self-study, keep in mind that your efforts are in line with the 2018 DoD’s AI Strategy, which states that we “will harness the potential of AI to transform all functions of the Department positively, thereby supporting and protecting U.S. [Service members], safeguarding U.S. citizens, defending allies and partners, and improving the affordability, effectiveness, and speed of our operations.”25 Further plans for educating select members of the DoD workforce are included in the 2020 DoD AI Education Strategy.26 Even though there are no current plans to resource formal AI education for legal professionals, those interested can start learning through self-study. Although not the focus of this commentary, there are many readily accessible means to acquire knowledge in AI technologies for the non-technical user. For example, Army legal professionals can take advantage of free audio and e-books, videos, and courses from several educational resources from the DoD.27 Interested legal professionals can reach out to the author for a growing list of educational and training resources.28
2. Conduct Simple Experiments with Commercially Available GPT Chatbots
Knowing that many legal professionals have already experimented with GPT chatbots, the effort advocated here is to be deliberate and coordinated in capturing and sharing individually learned insights on how to apply GPT chatbots in the practice of military law. Combined with this is the need to share the lessons within the DoD to exponentially increase the collective insights into the capabilities and limitations of using existing GPT chatbots to conduct the DoD’s legal support missions.
After the DoD establishes official policies on employee use of commercial GPT chatbots, conducting experiments could mean using the identical inputs (prompts) over time to see if outputs (answers) change or using identical inputs between GPT chatbots to gauge the accuracy or “expertise” levels of different GPT systems.29 These inputs and outputs would be saved and used to test a DoD GPT chatbot if, or when, one is developed.30 This is just one example of how legal professionals can design experiments to test and train future GPT chatbots. Sharing these lessons learned with other legal professionals will not only help many to understand the capabilities but also help them become connoisseurs of GPT chatbots and know how to rate them appropriately.31
One of the key factors to successfully developing a DoD GPT chatbot may be how well legal professionals, as collective users, plug into the ML operations (MLOps) methodologies that the AI community is developing.32 One possible way for legal professionals to plug into the process would be to become active and engaged team members of the integrated product teams (IPTs) to help develop and shape the DoD GPT chatbot.33 They would also act as liaisons to the legal community, not just passive legal reviewers after the product is ready to be rolled out. The input and feedback from legal professionals in the IPTs can be invaluable to developers, especially if there is a robust discussion about GAI LLMs and GPTs among the larger legal community. After integrating legal representatives into the MLOps process, the next line of effort moves into researching technology companies and their GPT chatbot products.
Before moving on, the point made above is re-emphasized here because of its importance: individuals must wait for the official DoD policies and guidance on how DoD personnel can use commercial chatbots to experiment with Government information, even if it may seem sterilized. The Navy’s Chief Information Officer (CIO) issued a memorandum in September 2023 to “offer interim guardrail guidance when considering the use of Generative [AI] or [LLMs].”34 In the guidance, the CIO states that “[t]he responsibility and accountability for user-induced vulnerabilities, violations and unintended consequences incurred by the use and adoption of LLMs ultimately resides with each individual organization’s respective leadership.”35 However, the guidance warns that that “aggregation of individual generative AI prompts and outputs could lead to an inadvertent release of sensitive or classified information.”36
In October 2023, the DoD’s Chief Digital and Artificial Intelligence Officer (CDAO) issued his “Interim Guidance on Use of Generative Artificial Intelligence” to the senior Pentagon leadership, combatant command commanders, and defense agency and DoD field activity directors.37 The guidance is marked controlled unclassified information, so the author can only share it with other Federal Government personnel. Needless to say, it is the responsibility of individual Government personnel to read this guidance and receive approval from the appropriate leader before proceeding with using commercial GPTs for Government work.
A useful mental framework to have before engaging a commercial GPT chatbot is to think that all your inputs and outputs will be publicly available. In other words, what you think are private interactions with the chatbot may not be. Just imagine that every keystroke inputted may one day be broadcasted on the jumbotron on One Times Square.38 Even without reading the DoD CDAO’s interim policy memorandum, reviewing the Navy’s warning above should make Government personnel pause before deciding to ask a commercial GPT chatbot a question or give it a task that might be incorporated in their Government work. As more Government personnel become tempted to use commercial GPT chatbots to do their work, it is imperative that leadership in every organization set up its own policies to protect Government information from inadvertently weakening operational security.
3. Market Research on GPT Chatbots and Requirements Development
Using the methods of market research prescribed by the Federal Acquisition Regulations (FAR) and other techniques used in Federal procurement (such as Broad Agency Announcements and Commercial Solutions Opening), legal professionals can assist by conducting informal market research on GPTs and in requirements development. Legal professionals would bring their trained minds and critical thinking skills to the effort. A discrete way to assist would occur at the requirements development phase when legal professionals can help draft key performance parameters (KPPs). One of the KPPs could be that the DoD GPT chatbot must be able to achieve certain tasks, such as pass the Multistate Bar Examination (MBE) and Law School Admission Test (LSAT) by a certain percentage or draft a legal review on certain topics, as a baseline capability.39
Although market research and requirements development on a DoD GPT chatbot may seem daunting, the DoD’s CDAO is providing resources to DoD organizations such as Tradewinds, a “framework for sourcing, funding, and developing solutions to challenges in the [AI]/[ML], digital, and data analytics space.”40 According to the website, Tradewinds “launched the Tradewinds Solutions Marketplace, a ground-breaking digital repository of post-competition, readily-awardable pitch videos, which address the Government’s greatest challenges in the AI/ML, digital, and data analytics space.”41 This online resource was built so DoD organizations can find companies and contract vehicles focused on AI solutions.42 These resources and tools will allow legal professionals with little experience in market research to act as the eyes, ears, and brains in the search for a DoD GPT solution.
According to news reports, the DoD has already started developing a ChatGPT-like prototype chatbot called Acqbot.43 It is self-described as “[p]owering the evolution of Government contracting” and “mak[ing] it easy for contracting officers to manage the contract lifecycle.”44 Little is known about how the DoD is developing this capability, but there may be developmental synergies between Acqbot and a general-use DoD GPT. Moreover, the issues of datasets and security requirements, discussed in the next line of effort, will be relevant to both Acqbot and a DoD GPT.
4. Identification of Datasets and Security Requirements for a DoD GPT Chatbot
Initial research on what makes GPT chatbots so impressive points to the LLM and the huge datasets behind the “leap ahead” level of performance of GPT-4 and its competitors. However, surveying the level of transparency on what constituted the datasets, it is unclear exactly what sources the GPT chatbots used to produce the substantive and fact-based outputs.45 Some GPT chatbots may be able to provide sources they used in the output but basic questions remain as to how ChatGPT and other LLM products can fully and accurately cite all sources used in the output.46 For example, it would be impressive to see what sources a GPT chatbot used to “know” the law and legal procedures on, say, qui tam lawsuits. Was it “fed” volumes of cases, treatises, law review articles, legal blogs, Wikipedia entries, and so on? If so, a full list should be readily available for the users to review. Here, the legal professional can fact-check and judge the adequacy of the sources as well as look for updates or changes to the law.
In addition to checking references, legal professionals can also be value-added in evaluating the level of reasoning, complexity, and integration of various subjects in the outputs. It may be debatable whether a GPT chatbot can actually “reason” and “integrate” multiple ideas like a human, but the examples that some GPT developers have provided seem to closely mimic human thought, such as understanding why a meme is funny.47
Beyond the issues of checking references and assessing a GPT’s skills in reasoning, many questions related to GPT databases and security are left to explore. The questions span the gamut, including these that come to mind: (1) How large should the raw data and training data be to achieve the best result? (2) What training models work best with the datasets used? (3) What is to be made of open source databases and nonpublic, secured data? (4) What can a programmer do about older datasets that may be useful on one level but not as useful in another? (5) What about updates and new data created from the outputs? These questions, issues, and more will need to be discussed with AI data experts who can shed light on these matters. After all, we may be asking the wrong questions due to a lack of technical and scientific knowledge.
One major benefit of developing a DoD GPT would be the DoD’s retention of control over selecting the datasets used to train the GPT. This is called “federated learning,” and the DoD could choose the universe of data with more deliberation after consulting with AI scientists regarding the optimal size of the data or number of datasets.48 In other words, if the dataset universe is too small, it could affect the quality, accuracy, completeness, and, ultimately, the utility of a DoD GPT. Too large and other factors like time and cost could increase and swing the cost/benefit analysis in the opposite direction.
A significant concern that requires research from a multitude of angles is cybersecurity and the protection of data in the development and deployment of a DoD GPT. The required security at the different stages of development could drive the focus, from algorithms to computation to the storage of data, including inputs and outputs once the GPT reaches full operational capability. The security protections needed to prevent any compromise of a DoD GPT must be front and center.49
LTC Hong speaks at the Contract Attorney’s Course at The Judge Advocate General’s Legal Center and School in Charlottesville, VA. (Credit: Billie Suttles, TJAGLCS)
Any deliberation on the possible use cases for a DoD GPT includes considerations of when there should be hard limitations or restrictions on the use of a DoD GPT. This research area could dovetail with the existing research and legal analysis on the ethical use of AI in national security and military operations.50 One concern that comes to mind is the possibility of unauthorized users gaining access to input data, even within the DoD. For example, if both Government prosecutors and defense counsel are using a DoD GPT and they input data into the system (such as parts of a case file for analysis) in preparation for trial, is it possible that one side can use the DoD GPT to glean information on what the adversarial side is asking the GPT and what outputs were given?51 This assumes that input data fed into a DoD GPT is looped back into the dataset for use in computing answers to future prompts.52 Again, these questions and issues only skim the tip of the iceberg, but if the best and brightest AI and data experts are partnered with the DoD, the hope is that a strong security solution will be implemented and a DoD GPT chatbot can be deployed.
5. Application and Use Cases for a DoD GPT Chatbot
In this commentary, the foundational questions of whether a DoD GPT chatbot can and should be used in military legal practice have been presumed. However, the hope is that a robust discussion can occur within the legal community as many legal professionals take up interest in researching and writing about how a DoD GPT chatbot can affect their area of practice. One way to frame the inquiry into whether a DoD GPT can make an impact is to divide up the query by the main practice areas found in the U.S. Army JAG Corps, such as legal assistance, administrative law, ethics, military justice, contract and fiscal law, national security, labor and employment, environmental, and other highly specialized areas such as intellectual property law.53
Some have already considered possible use cases for a GPT by asking ChatGPT how it can be used in the legal profession.54 The typical use case list would contain common legal work such as: summarizing investigative files; drafting findings and recommendations; developing direct- and cross-examinations based on the investigative files and sworn statements; and routine ethics opinions like gifts and post-Government employment letters.55 The answers provided by ChatGPT only scratch the surface. If more minds were put to this task, the probability of greater insights and recommendations would be netted and shared, and this would lead to further insights and recommendations; thus, the start of a legal-focused AI spring or boom coming out of the DoD.56
The issues with the proper application and limits on use cases of a DoD GPT would need to be examined and expanded upon once possible use cases are set for discussion and debate. It is easy to envision that the debate on the acceptable and unacceptable uses of a DoD GPT can also inform the application of the DoD’s ethical principles for AI on other systems beyond LLMs (such as AI and computer vision).57 In any case, the DoD will likely need to partner with private technology companies leading the field of AI to explore the possibilities.
6. Assist in the Development of Acquisition Strategies
Congress has provided the DoD with new procurement authorities to accomplish its mission to modernize for the future fight.58 Since developing a DoD GPT will most likely involve the private sector, many legal professionals, as both future users and experts in contract and fiscal law, can assist the procuring arm of the DoD in developing an acquisition strategy.59 Although acquisition strategies are only required for major system acquisition programs (FAR-based contracts) that rise to a certain total cost level, a strategy can be developed for any procurement to promote the goals of effective, economical, and timely decisions.60 Even though Other Transactions (OTs) may be the preferred method for partnering with industry on a DoD GPT, and thus not requiring an acquisition strategy, developing an acquisition strategy nonetheless will help shape how the DoD can negotiate an OT agreement.61
(Credit: emojoez-stock.adobe)
Informed by personal study, market research, and individual and collective experimentations, legal professionals immersed in the GPT arena can help draft the statement of need, product or service descriptions, logistic and security considerations, and many other areas addressed in successful acquisition plans. Of special importance for GPT technology procurements will be how to negotiate the data rights and intellectual property (IP) elements. This emerging area intersects Federal acquisitions law with IP, licensing arrangements, and, perhaps, inventions and patent law. The DoD will need the experts at the DoD IP Cadre to weigh in on how best to negotiate all relevant aspects of a DoD GPT that fall into their areas of expertise.62
7. Collaboration within the DoD
As mentioned above, teaming with many DoD individuals, groups, and offices will be key in this effort to bring about a DoD GPT. As with all successful initiatives, a good starting point would be the creation of a central digital gathering point within the DoD for legal professionals to share and collaborate on projects with like-minded people. This starting point could be a webpage on the DoD’s Milsuite.mil or a Government Microsoft Teams channel that is open to all DoD legal professionals.63 There are many different ways to create a central digital gathering point, but the main point is that the time to create a network of legal professionals is now.
(Credit: emojoez-stock.adobe)
Creating something informal yet effective could be done with little cost and effort. The pandemic years have taught us that computer-based communication tools are available so people can share their thoughts within a large group with a few mouse clicks, a microphone, and a video camera. If the central hub is set and opened, optimism exists in many corners of the DoD that it will be available for all DoD personnel to join and participate. Although these ideas are quite simple, thrilling questions remain as to who will create this digital media space, when, where, what it will consist of, and how they will create it. After these questions are answered, still more questions are surely to follow related to resourcing and organizational sponsorships.
Conclusion
To those who agree with the thesis of this commentary—that the DoD should create its own GPT—know there are strong desires in certain legal circles to join in that effort. For now, these people are informally engaging with one another in the lines of effort outlined above. However, it is reasonable to think that if anything of consequence is to see the light of day, it will only be because of tremendous individual initiative, teamwork, and collaboration. If enough legal professionals in the DoD see the urgency to act on these (and other future lines of efforts) and are open to working with others, there may be positive short-term gains achieved without any significant investment of capital or loss of time.
Another idea to foster is to take the above lines of effort and turn them into a conference agenda. If a DoD GPT symposium can attract hundreds of legal professionals to gather to discuss their thoughts, research, and progress throughout the DoD, a groundswell of action can develop. The hope is, even if the DoD is coming into the GPT capability development late in the game, once the collective minds of legal professionals are united on bringing a DoD GPT into the world, the American people will welcome the news. They will welcome it because it will be known that the incredible power of GPT technology is being studied, developed with the assistance of, and utilized by the trusted legal professionals of the DoD; legal professionals who, rather than operating with a profit motive, stand for principled counsel, mastery of the law, stewardship, and servant leadership.64 TAL
LTC Hong is the Chair of the Contract and Fiscal Law Department at The Judge Advocate General’s Legal Center and School in Charlottesville, Virginia.
Notes
1. This commentary was originally drafted in April 2023 before GPT-4 Turbo was rolled out to the public, but I will keep the original ChatGPT reference here as it is a generic label for an AI capability. Moreover, the proposal is for the Government to create, train, and maintain a GAI rather than just acquire a custom DoD application programming interface (API) that connects to a commercial product.
2. We are now living in the era where disclosure of this nature may become standard practice. It was reported that over 200 e-books listed ChatGPT as an author or co-author in Amazon’s Kindle store; however, the question remains—how many written works since ChatGPT’s public debut have incorporated GPT outputs and not properly credited it? See Greg Bensinger, ChatGPT Launches Boom in AI-Written E-Books on Amazon, Reuters, (Feb. 21, 2023, 3:34 PM), https://www.reuters.com/technology/chatgpt-launches-boom-ai-written-e-books-amazon-2023-02-21.
3. The term “GPT” is being used generically in this commentary. The term “generative pre-training” was introduced in a paper written by four OpenAI personnel in June 2018. See Improving Language Understanding by Generative Pre-Training, OpenAI.com (June 11, 2018), https://openai.com/research/language-unsupervised. The paper can be downloaded at the link above. The term “generative pre-trained transformer” was also coined by OpenAI and is the combination of two ideas: unsupervised pre-training and transformers. Id. A “chatbot” is defined as “a bot . . . that is designed to converse with human beings” by the Merriam-Webster online dictionary. Chatbot, Merriam-Webster, https://www.merriam-webster.com/dictionary/chatbot (last visited Jan. 2, 2024).
4. ChatGPT is a product created by OpenAI, which describes the product as a model that “interacts in a conversational way.” See Introducing ChatGPT, OpenAI.com, https://openai.com/blog/chatgpt (last visited Jan. 2, 2024).
5. The next generation of ChatGPT is GPT-4 (which requires the pay-to-use ChatGPT Plus) and is another OpenAI product that it promotes as being able to “solve difficult problems with greater accuracy, thanks to its broader general knowledge and problem-solving abilities.” GPT-4 Is OpenAI’s Most Advanced System, Producing Safer and More Useful Responses, OpenAI, https://openai.com/gpt-4 (last visited Jan. 2, 2024).
6. See Jack Dunhill, GPT-4 Hires and Manipulates Human into Passing CAPTCHA Test, IFLScience (Mar. 16, 2023), https://www.iflscience.com/gpt-4-hires-and-manipulates-human-into-passing-captcha-test-68016 (describing GPT-4 using very human-like hacking plans and execution, such as hiring a human to solve a Completely Automated Public Turing Test to Tell Computers and Humans Apart (CAPTCHA) for it).
7. Some companies are claiming that its product can detect whether a written work is generated by AI, but the author’s personal testing on free sites showed that the detection technology is not consistent and, thus, not useful.
8. Imperfect writing may be a strategy to pass written work product as human-produced, similar to the story of when a computer simulated speaking like a thirteen-year-old boy to pass the Turing test. See Computer Simulating 13-Year-Old Boy Becomes First to Pass Turing Test, The Guardian (June 9, 2014), https://theguardian.com/technology/2014/jun/08/super-computer-simulates-13-year-old-boy-passes-turing-test.
9. In this question, reference to the DoD would include all military and Civilian legal personnel from all the Service components and civilian agencies under the DoD umbrella.
10. Deep Blue Defeats Garry Kasparov in Chess Match, History, https://www.history.com/this-day-in-history/deep-blue-defeats-garry-kasparov-in-chess-match (last visited Jan. 2, 2024).
11. See Deep Blue, IBM, https://www.ibm.com/history/deep-blue (last visited Jan. 2, 2024).
12. Steven Borowiec, AlphaGo Seals 4-1 Victory Over Go Grandmaster Lee Sedol, The Guardian (Mar. 15, 2016), https://www.theguardian.com/technology/2016/mar/15/googles-alphago-seals-4-1-victory-over-grandmaster-lee-sedol; James Vincent, DeepMind’s AI Agents Conquer Human Pros at StarCraft II, The Verge (Jan. 24, 2019), https://www.theverge.com/2019/1/24/18196135/google-deepmind-ai-starcraft-2-victory.
13. IBM’s Watson Supercomputer Crowned Jeopardy King, BBC (Feb. 10, 2011), https://www.bbc.com/news/technology-12491688.
14. “Putting down the coffee pot” is a reference to a movie scene in Glengarry Glen Ross, a 1992 American movie adapted by David Mamet of his play by the same name. See MovieClips, Put That Coffee Down! – Glengarry Glen Ross (1/10) Movie CLIP (1992) HD, YouTube (May 23, 2012), https://www.youtube.com/watch?v=r6Lf8GtMe4M.
15. Michael James Bommarito & Daniel Martin Katz, GPT Takes the Bar Exam (2022), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4314839; GPT-4, OpenAI, https://openai.com/research/gpt-4 (last visited Jan. 2, 2024). As for passing the bar exam, see OpenAI’s technical report that shows that GPT-4 passed the Uniform Bar Exam (UBE) with a score of 298/400. GPT-4, supra.
16. See, e.g., Andrew Perlman, The Implications of ChatGPT for Legal Services and Society, The Practice, Mar./Apr. 2023, https://clp.law.harvard.edu/knowledge-hub/magazine/issues/generative-ai-in-the-legal-profession/the-implications-of-chatgpt-for-legal-services-and-society.
17. Model Rules of Pro. Conduct r. 1.1 (Am. Bar Ass’n 2020).
18. The term “knowledge professional” can essentially cover all DoD personnel because without general and specialized knowledge, DoD personnel would not be capable of conducting the department’s day-to-day business and accomplishing its missions. Even if the term is limited to language-based knowledge, it would essentially cover all DoD personnel.
19. In August 2023, the DoD “announced the establishment of a generative [AI] task force, an initiative that reflects the DoD’s commitment to harnessing the power of artificial intelligence in a responsible and strategic manner. Deputy Secretary of Defense Dr. Kathleen Hicks directed the organization of Task Force Lima; it will play a pivotal role in analyzing and integrating generative AI tools, such as . . . LLMs, across the DoD.” Press Release, U.S. Dep’t of Def., DOD Announces Establishment of Generative AI Task Force (Aug. 10, 2023), https://www.defense.gov/News/Releases/Release/Article/3489803/dod-announces-establishment-of-generative-ai-task-force.
20. There are many articles on how much data is needed to create an effective GPT chatbot. See, e.g., Ameya Paleja, Alpaca AI: Stanford Researchers Clone ChatGPT AI for just $600, Interesting Eng’g (Mar. 21, 2023), https://interestingengineering.com/innovation/stanford-researchers-clone-chatgpt-ai.
21. The availability of CUI information to interface with a DoD GPT can potentially make the DoD GPT the most valuable GAI LLM in existence; thus, care must be taken from the ground up to protect and secure a DoD GPT.
22. The professor was Lieutenant Colonel Craig Scrogham, Vice-Chair of the Contract and Fiscal Law Department at The Judge Advocate General’s Legal Center and School, and he consented to this attribution. I believe the quote is: “We are the best Control-F’ers out there!”
23. The task of finding sources of information to answer non-legal questions may become easier with the development of DoD’s GAMECHANGER. GAMECHANGER is a natural language processing website that allows DoD personnel to “[s]earch over 50,000 policy documents.” GAMECHANGER, https://gamechanger.advana.data.mil/#/gamechanger (last visited Feb. 12, 2024) (requiring a Common Access Card). Based on its tagline, it would make Control-F’ers’ lives easier. However, its utility may be hampered by lackluster results; for example, the author searched the database with the terms “generative ai,” “generative artificial intelligence,” and “large language model,” and netted no DoD policy documents—only executive orders and congressional hearing transcripts and similar documents came up. If you have a Common Access Card, you can try GAMECHANGER at the URL supra.
24. The Judge Advocate General’s Legal Center and School Artificial Intelligence Study Group (TAISG)’s first meeting was held on 13 February 2023.
25. U.S. Dep’t of Def., Summary of the 2018 Department of Defense Artificial Intelligence Strategy: Harnessing AI to Advance Our Security and Prosperity 4 (2019) [hereinafter 2018 DoD AI Strategy Summary], https://media.defense.gov/2019/Feb/12/2002088963/-1/-1/1/SUMMARY-OF-DOD-AI-STRATEGY.PDF.
26. U.S. Dep’t of Def., 2020 Department of Defense Artificial Intelligence Education Strategy (2020) [hereinafter 2020 DoD AI Education Strategy], https://www.ai.mil/docs/2020_DoD_AI_Training_and_Education_Strategy_and_Infographic_10_27_20.pdf.
27. Resources like Percipio (usarmy.percipio.com) and DoD MWR Libraries (www.dodmwrlibraries.org) are available for free to Army personnel. Another valuable resource is from the Air Force and its Digital University (https://digitalu.af.mil/), which provides access to MIT Horizon, “an enterprise-level content library to cover the latest emerging technology topics.” MIT Horizon, MIT Open Learning, https://openlearning.mit.edu/courses-programs/mit-horizon (last visited Feb. 8, 2024). Lawyers should find MIT Horizon’s purpose statement enticing: “Convenient micro-assets are designed to help technical and non-technical learners stay informed about the latest technologies to drive impact and innovation.” Id.
28. The study group has posted resources in its Microsoft Teams Channel.
29. It may be the case that outputs will never be identical even with the same input used each time because the GPT chatbots go through a massive number of calculations for each input and the calculation pathway may differ. See Jaron Lanier, There Is No A.I., New Yorker (Apr. 20, 2023), https://newyorker.com/science/annals-of-artificial-intelligence/there-is-no-ai (“There is only a giant ocean of jello—a vast mathematical mixing.”).
30. Security concerns on using a commercial GPT like ChatGPT by Federal employees need to be addressed because even commercial companies are starting to ban employee use of ChatGPT. See Mark Gurman, Samsung Bans Staff’s AI Use After Spotting ChatGPT Data Leak, Bloomberg (May 2, 2023, 1:54 PM), https://www.bloomberg.com/news/articles/2023-05-02/samsung-bans-chatgpt-and-other-generative-ai-use-by-staff-after-leak?leadSource=uverify%20wall.
31. Another way for DoD attorneys to get involved in testing and evaluating chatbots is to participate in testing and voting activities like LMSYS Chatbot Arena Leaderboard. Chatbot Arena: Benchmarking LLMs in the Wild, https://chat.lmsys.org (last visited Feb. 12, 2024). The activity created by LMSYS is to compare the outputs of two chatbots and vote for the one that is better. The technology behind this website seems to be an API connected to all the mainstream chatbots and the user interface is designed to make it convenient to compare many chatbots. The purpose of the website is stated as: “Our mission is to build an open crowdsourced platform to collect human feedback and evaluate LLMs under real-world scenarios.” Id. at “About Us.”
32. See Harshil Patel, MLOps vs. DevOps vs. ModelOps, Censius.ai, https://censius.ai/blogs/mlops-vs-devops-vs-modelops (last visited Jan. 2, 2024) (“[Machine learning operations (MLOps)] is a way for data scientists and operations experts to collaborate and communicate in order to manage the production ML lifecycle.”).
33. The 2020 DoD AI Training and Education Strategy describes IPTs as “multidisciplinary groups of [p]roduct [m]anagers and AI developers whose roles are central to delivering AI capabilities.” 2020 DoD AI Education Strategy, supra note 26, at intro. The strategy calls for the creation of a cadre of IPTs who are “composed of product managers, data scientists, AI/ML engineers, IT technicians, and UI/UX designers from the ‘Create AI’, ‘Drive AI’, and ‘Embed AI’ archetypes.” Id. at 5 & n.14.
34. See Memorandum from U.S. Dep’t of Navy, Chief Info. Officer, subject: Department of the Navy Guidance on the Use of Generative Artificial Intelligence and Large Language Models para. 1 (6 Sept. 2023), https://www.doncio.navy.mil/ContentView.aspx?id=16442.
35. Id. para. 3(e).
36. Id. para. 3(c).
37. The memorandum was sent to the author and is marked CUI, thus, it is not publicly available. Government personnel must get a copy of this memorandum and read it before utilizing commercial GPT chatbots for any task, even if one thinks the contemplated prompt or input does not implicate Government information or work. The difficulty will be knowing when an input would cross into the area of Government information or work (vice personal matters), thus, triggering required reviews and approvals. These difficulties would be eliminated for the most part if a DoD GPT chatbot existed. Even with a DoD GPT chatbot, however, questions would arise as to how certain inputs can be compartmentalized and access limited by or to others, such as when electronic walls must be erected between the prosecuting offices and trial defense attorneys.
38. More research is needed, but the retrieval of individual user’s inputs and sessions may be possible even if the chatbot states that it does not retain previous interactions for the sake of a user’s privacy. For all we know, the data is still on the company servers and the response that it does not retain previous interactions could be just another guardrail placed by the company. Like hacking into a traditional network, people have tried, and will continue to try, to hack GPT chatbots to get behind the guardrails.
39. Other standardized legal tests could be more military law-centric, such as the exams given to graduate students at TJAGLCS. These exams could include multiple-choice, short answer, or essay questions.
40. About Tradewinds, Tradewinds, https://www.tradewindai.com/about (last visited Jan. 2, 2024).
41. Id.
42. See id.
43. Jon Harper, AI Bot Developed to Help Defense Department Write Contracts Faster, DefenseScoop (Feb. 8, 2023), https://defensescoop.com/2023/02/08/ai-bot-developed-to-help-defense-department-write-contracts-faster.
44. AcqBot, https://acqbot.com (last visited Feb. 9, 2024).
45. See Kyle Barr, GPT-4 Is a Giant Black Box and Its Training Data Remains a Mystery, Gizmodo (Mar. 16, 2023), https://gizmodo.com/chatbot-gpt4-open-ai-ai-bing-microsoft-1850229989.
46. According to a company called Scale AI, its LLM for U.S. National Security can provide “citations” or sources used in the output. See Donovan: AI Digital Staff Officer for National Security, Scale, https://scale.com/federal-llm (last visited Jan. 2, 2024). OpenAI is also working toward the same goal as seen in their developer forum, which talks about “ChatGPT-4 with Citations / Sources.” See ChatGPT-4 with Citations / Sources, OpenAI, https://community.openai.com/t/chatgpt-4-with-citations-sources/164323 (last visited Feb. 9, 2024).
47. See Stephen Johnson, GPT-4 is Surprisingly Good at Explaining Jokes, Freethink (Mar. 18, 2023), https://www.freethink.com/robots-ai/gpt-4-jokes.
48. See What Is Federated Learning?, IBM, https://research.ibm.com/blog/what-is-federated-learning (last visited Feb. 9, 2024).
49. See supra note 30.
50. See Matthew Ivey, The Ethical Midfield in Artificial Intelligence: Practical Reflections for National Security Lawyers, 33 Geo. J. Legal Ethics 109 (2020).
51. If this was possible, ABA’s Rule 1.6 (Confidentiality of Information) could be implicated for the attorney that used the GPT chatbot and inputted certain details that could be considered a disclosure of client information. See Model Rules of Pro. Conduct r. 1.6 (Am. Bar Ass’n 2020).
52. This hypothetical is only focused on teasing out the technical capabilities (or vulnerabilities) of a GPT and legal professionals’ actions in gray areas. It would be easy to see that attempting to secretly discover an adversarial party’s legal research, even if it focused on search terms directed to a legal data base, would be unethical.
53. For more details on these practice areas, see U.S. Dep’t of Army, Field Manual 3-84, Legal Support to Operations (1 Sept. 2023).
54. See, e.g., Pearlman, supra note 16.
55. Lieutenant Colonel Sean B. Zehtab contributed this list.
56. The reference to spring or boom is to contrast with AI winters. See Valeriia Kuka, The Story of AI Winters and What It Teaches Us Today (History of LLMs. Bonus), Turing Post (June 30, 2023), https://www.turingpost.com/p/aiwinters.
57. See 2018 DoD AI Strategy Summary, supra note 25. The example of computer vision needs to be further developed, but there may be corollaries between how LLMs can “hallucinate” and give a wrong answer with how AI-powered computer vision classifies images and makes errors in the identification of objects or people. See Brian Hayes, Computer Vision and Computer Hallucinations, Am. Scientist, Nov./Dec. 2015, at 380. The question will be what percentage of error will be acceptable or unacceptable for deployment.
58. 10 U.S.C. §§ 4021, 4022; see also Douglas Steinberg, Leveraging the Department of Defense’s Other Transaction Authority to Foster a Twenty-First Century Acquisition Ecosystem, 49 Pub. Cont. L.J. 537 (2020).
59. An “acquisition strategy” is defined in the Federal Acquisition Regulation as “the program manager’s overall plan for satisfying the mission need in the most effective, economical, and timely manner.” FAR 34.004 (2023).
60. The Defense Federal Acquisition Regulation Supplement requires a written acquisition plan for any developmental acquisitions with a total cost of $10 million or more or when deemed “appropriate by the department or agency.” DFARS 207.103(d)(i)(A), (C) (Dec. 2023).
61. The DoD Other Transactions Guide contains useful information on negotiating Other Transactions. See Off. of Under Sec’y of Def. for Acquisition & Sustainment, U.S. Dep’t of Def., Other Transactions Guide (2023), https://www.acq.osd.mil/asda/dpc/cp/policy/docs/guidebook/TAB%20A1%20-%20DoD%20OT%20Guide%20JUL%202023_final.pdf.
62. See Intellectual Property Cadre, Off. of Assistant Sec’y of Def., Acquisition, https://www.acq.osd.mil/asda/ae/ada/ip-cadre.html (last visited Jan. 2, 2024).
63. Searching Milsuite.mil for groups focused on GPT technology netted zero results. There are many communities under the search term “artificial intelligence” but those groups are based on organizational affiliations or not targeted to legal professionals.
64. The U.S. Army JAG Corps has created these “Four Constants” but these standards are readily applicable throughout the DoD legal profession. For an explanation of the Four Constants, see The Judge Advoc. Gen.’s Corps, U.S. Army, Four Constants (n.d.), https://www.jagcnet.army.mil/Sites/JAGC.nsf/46DCA0CA1EE75266852586C5004A681F/$File/US%20Army%20JAG%20Corps%20Four%20Constants%20Smart%20Card.pdf.