By Mike Holley
First Assistant District Attorney in Montgomery County
Los Angeles County is big. Really big. More people live in Los Angeles County than in 40 of the 50 states. The Los Angeles County Public Defender’s Office (LACPD) is also big, consisting of 1,200 employees, including over 700 attorneys.[1] These 700 attorneys receive reports from 99 law enforcement agencies. These 99 agencies use various formats for reports, including handwritten documents. To digitize those incoming documents and videos, LACPD would need to hire a substantial number of data entry staff.
Instead, LACPD began using an artificial intelligence (AI) system to assist with this process. Under the supervision of Mr. Mohammed Al Rawi, the office’s gifted Chief Information Officer, LACPD developed a system to format the reports in a manner most helpful to the attorneys for bond hearings, docket, trial prep, etc. The AI system quickly identifies information useful to the attorneys by recognizing which arresting agency submitted the report and “understanding” the pattern of each agency. The system extracts relevant information accurately and efficiently, placing it at the fingertips of LACPD attorneys and staff in an “intuitive interface,” significantly enhancing the quality and speed of the important work done there. As of April 2024, the cost to the office for these services, including document recognition, video transcription, and indexing, is approximately $4,000 a month, replacing the costs of about 50 administrative positions.[2]
Friends, AI has arrived in our profession.[3]
AI is a topic that generates both confusion and strong emotions. For about the last 18 months, I have invested time reading about the subject, listening to podcasts, and attending conferences and seminars. I have spoken about AI and prosecution, and I have been utilizing AI tools myself and overseeing the use of AI tools in my office.[4] I have come to believe that the arrival of AI is inevitable, transformative, and deeply concerning. As a supervisor and prosecutor, my hope is to manage these AI tools in a way that enhances our mission without compromising it. This will be a challenge. Part of the challenge is the emotional aspect of this issue. Specifically, I know from many discussions with people that this subject causes people anxiety, frustration, and sometimes outright hostility.
I get it.
Let me say here at the outset that it will be OK. We can start with this article, where I will try to assist you in navigating these new trails. I hope to make this subject as straightforward as possible. The topic is complex, but I will try to use plain language and avoid the overstatements that often accompany discussions of AI. This is a primer, a foundational first step. It’s not the end of the discussion but a start.
General statements about AI
Looking under the hood, you see that AI is just software. That’s it. The software is then fed information—spreadsheets, books, websites, images, speeches, videos, etc.—and put to work in various ways. The software operates as a type of machine learning, and it can do wondrous things. It can also do stupid things. There is an element of randomness to the software, both by design and as a byproduct of how the software works.[5] Exactly how the software does what it does and what data goes into the “machine” is often unknown by the user. For this reason, AI is sometimes considered a “black box” system because we can’t look inside to see exactly how it works.[6] Significantly, the software can convincingly mimic the human mind, but—and this is very important—the software is not a human mind.
There are three main types of machine learning to consider:
1) AI, or “artificial intelligence,” is a system that performs specific tasks that normally require human intelligence, such as playing a game or recognizing human speech. An example of AI is ChatGPT, which is also called a “large language model” (LLM). An LLM specializes in dealing with text. Another example of AI would be DALL-E, which is called a “generative adversarial network” (GAN). A GAN creates realistic synthetic images or audio from textual description. “Deepfakes” would fall in this category.
2) AGI is “artificial general intelligence,” and it is a system much closer to human intelligence. AGI is capable of performing any intellectual task a human can do, and it is not limited to specific tasks. AGI is theoretical at this point and does not currently exist at the time of this article.
3) ASI is “artificial superintelligence.” ASI is a system which would surpass human intelligence in all aspects, including creativity, problem-solving, and emotional understanding. ASI, like AGI, is a theoretical concept and has not yet been achieved.[7]
Many smart, well-informed, educated people worry that AI or, more likely, AGI or ASI, poses an existential threat to humanity. There are different theories as to how this will occur.[8] Other smart, well-informed, educated people are concerned that AI, even as it currently exists, will damage our civilization by, among other things, causing us to lose the ability to think for ourselves or perform other essential tasks (such as reading and writing) or further impair our (already degraded) skills at dealing with one another in person. Others are concerned that AI will create significant unemployment.[9] Sometimes, these people are referred to as AI “safetyists,” “decelerationists,” or “doomers.”[10]
A different group of smart, well-informed, educated people believe AI will lead to a massive, positive transformation of society. They believe AI will lead to innovations in education (what if every child had a world-class tutor who never got tired and never complained?[11]), medicine (what if AI could be a near-perfect doctor on your smartphone or could cure cancer?[12]), and bring improvements to every possible area of life (what if AI could translate every written or spoken language in real time?). They believe AI will allow humans to move from tedious, repetitive, or dangerous work to higher-quality, more fulfilling work. They believe AI will assist us in solving hard problems (sometimes called “wicked problems,”[13] such as homelessness or nuclear proliferation) that we have not previously been able to solve on our own. They believe that we should push AI harder and faster. These people are sometimes referred to as “effective accelerationists” or “e/accs” (pronounced “e-acks”).[14]
Efforts to regulate AI face significant, perhaps insurmountable challenges. For one, AI is proliferating at an astonishing rate, greatly aided by two other powerful technologies: the internet and the smartphone. AI cannot be completely contained because, at its core, it is information, and information is capable of being transmitted quickly, freely, and effortlessly across borders. Moreover, because AI is a powerful technology, governments and individuals aggressively and continuously seek to develop and employ it. Notably, regulating AI in one country or one area will not prevent AI from advancing in another country or area. AI is developing at a breakneck speed, making it as challenging to regulate as holding back a flowing river with your hands.[15] OpenAI, Anthropic, Google, and Meta are the major tech companies developing AI in the United States, but thousands of other businesses and organizations are developing or adopting their products (67,000 companies as of January 2024).[16] In 2022, IBM reported that 35 percent of global companies are using AI, with China having the highest rate of AI adoption at 58 percent.[17] These numbers have doubtlessly increased in the intervening two years.
Criminal law and AI
Our primary interest lies in the intersection of AI and criminal law. Here are some considerations in that regard.
People with criminal intentions are actively and effectively using AI now and will continue to do so. AI “supercharges” known criminal activity, particularly fraud, and creates new types of criminal conduct, some of which our current law does not adequately address.[18] (We know better than anyone that a determined criminal will use powerful technology in any and every way he can.)
AI can create audio or visual reproductions of people, objects, places, or events that are difficult to distinguish from the real ones. This creates two problems. First, we must guard against being deceived by fake or altered videos. These could come from several sources, including complaining witnesses, defendants, and experts. The second, and I cannot stress this point enough, is to find ways to overcome our jurors’ growing doubts about the authentic video and audio we offer into evidence. At some point, jurors may be unwilling to trust anything they see or hear presented in digital form, which is disastrous for our work. We must stay ahead of these challenges in whatever way we can.
An emerging challenge for us as prosecutors is that the evidence we present to juries will be increasingly attacked as to its veracity, with doubt being liberally and creatively sown in the jury box. We must, therefore, develop methods and procedures to repel these attacks and to guard ourselves against deception.
Some AI companies are philosophically averse to using their work to aid law enforcement, which may create challenges for prosecutors seeking applications.[19] There are also operational security issues and privacy concerns with law enforcement’s sensitive information, particularly when we “upload” information into AI systems.
AI brings robust and novel new tools for law enforcement to detect and address crime.[20] Some AI applications—facial recognition, for example—raise questions of fairness, privacy, and racial disparities that must be addressed squarely, skillfully, and honestly.[21]
Practical applications for AI
Given the above, here are some considerations for using AI in your day-to-day work.
AI can digest any information, summarize it, and respond to queries. For example, AI can read a lengthy report and then list witnesses and their significance to the case or create a comprehensive timeline of events. AI can review bank or medical records and then summarize trends or patterns or answer (as if you were speaking to a person) questions about those records. AI could quickly digest jail calls or body cams and answer queries about that information. AI can also translate virtually any language in real time, including victim or witness interviews or defendant statements; transcribe audio and video; and identify nuances of communication that may interest a prosecutor in a particular case.
AI can be positioned to assess probabilities of success for probations and diversions and help us more quickly and accurately assess cases that may need to be directed to specialty courts. AI could also assist with bond recommendations and conditions.
AI can create visual and audio depictions of virtually any event, including crime scene recreations, demonstrative exhibits, etc. A prosecutor using a keyboard and computer can soon accomplish tasks that previously required a film studio.[22]
AI can greatly aid our textual work. Although there have been some widely publicized stumbles involving “hallucinated” cases created for legal briefs, AI has the potential to be an extremely powerful aid to legal research and writing. For example, very soon (perhaps right now?), a prosecutor could receive a defense appeal and then feed the appeal and record of the trial into an AI system. The AI system would then craft a very good response brief in minutes instead of days. The brief would need to be checked and polished, but still, the quality of the work might be on par with (or perhaps even better than) what many prosecutors could do given the other responsibilities we have, and the work would certainly be done much more quickly.[23] AI can instantly draft a bench brief, and soon, AI will be able to suggest (and respond) to objections in a trial in real time.[24] AI can also draft pleadings, warrants, and other legal documents quickly and (increasingly) accurately for our review.
AI is a tremendously powerful idea generator. AI can propose arguments and counter-arguments, suggest voir dire questions, offer trial themes, etc. (In this area, anything a colleague could do for you, AI is likely able to do. Think of AI as an always-on-hand assistant who does not drink the last of the coffee but then refuses to make another pot.)[25] AI can also write anything you are required to write—from an email response to a challenging victim to an opening argument to a speech you give high school students about the dangers of fentanyl.[26] AI can also help train prosecutors and staff in various ways.
Next steps for prosecutors
With the above information in mind, here are possible next steps for you in your office.
For mid-size and large offices, designate an innovation officer and have that person start experimenting with AI. Contract with companies that will help us create self-contained systems that protect the sensitive information in our possession without disclosing it more widely. Determine what AI systems will help us “see that justice is done.” Then, pass this information along to the rest of us.
For smaller offices, start with a large language model such as ChatGPT or Claude. Start with the free version and just experiment with the system as an idea generator. This will be trial and error, but you will see the value with time and a little patience. Treat these LLM’s like you would a 3L intern. Try the Lexis or Westlaw AI features and see if they don’t add value to your work. (There is a real skill in “prompts” that you will need to learn and practice. Be patient with this, and don’t quit too soon.)
We need conferences and serious papers on this subject, and we need to be part of managing AI’s relentless advent to achieve what good can be achieved and mitigate its negative impacts.[27]
We need to work with legislators to protect the people of Texas from criminals who use AI to exploit or endanger them.
We need to stay the course. Whether the e/accs are right, the Doomers are right, or both are right and wrong in one way or another, we need to be clear-headed and committed to our fundamental, most human of missions. We should neither panic nor ignore AI. We should continue to lead and serve.
At the Montgomery County District Attorney’s Office, we have used AI in a number of ways (see this endnote for specifics[28]). We are very interested in continuing this conversation with you, and if you are interested in doing so, please contact us. Whether you reach out to us or not, we wish you every success in this “brave new world” of AI.[29]
Endnotes
[1] Law Offices of Los Angeles County Public Defender. (n.d.). https://pubdef.lacounty.gov/about-us.
[2] Comments made by Mr. Mohammed Al Rawi during the “Prosecution Leaders of Now” Spring 2024 Workshop entitled “AI and Prosecution.” Many thanks to Patrick Robinson (and others) for putting that conference together. You can find several excellent resources about AI and prosecution at www. aiandprosecution.com.
[3] It may be wise in what follows to keep “Amara’s law” in mind, which states that we “tend to overestimate the effect of technology in the short run and underestimate the effect in the long run.” DevIQ. (n.d.). Amara’s Law: A Quick Guide on Technology Predictions. https://deviq. com/laws/amaras-law#what-is-amaras-law.
[4] In our office, I have appointed a “King of AI,” Mr. Chris Seufert, and a “Queen of AI,” Ms. Lianne Baldridge, to assist us in our efforts. Chris leans forward—and is doing a tremendous job at finding ways to use AI, and Lianne is the wise voice of caution. Both are appreciated.
[5] Because AI inherently includes a degree of randomness as a feature rather than a bug, demonstrating the reliability of AI results to court may be challenging. For example, when a judge says, “Show me how you got that result,” the same prompt you used before may result in a slightly different outcome. One Washington state Superior Court judge recently rejected the admission of video exhibits offered by the defense which were “enhanced by artificial intelligence,” partly because the results were not reproducible. Washington v. Puloka (Superior Court of Washington for Kings County, March 29, 2024). This is an unreported decision.
[6] Yasar, K., & Wigmore, I. (2023, March 17). “What is black box AI? Definition from TechTarget.” WhatIs. www.techtarget.com/whatis/definition/black-box-AI. The opposite of a “black box” is a “glass box” which is fully interpretable by people. For more on this subject, see Garrett, Brandon L. and Rudin, Cynthia, The Right to a Glass Box: Rethinking the Use of Artificial Intelligence in Criminal Justice (February 16, 2023). Cornell Law Review, Forthcoming, Duke Law School Public Law & Legal Theory Series No. 2023-03, Available at SSRN: https://ssrn.com/abstract=4275661 or http://dx.doi.org/10.2139/ssrn.4275661
[7] If AGI or ASI comes into existence, I hope it will consider how respectful I’ve been here. I, for one, welcome my new AI overlords.
[8] This group describes the probability that AI will kill everyone either on purpose or on accident as their p(doom). For frame of reference, Elon Musk, who has heavily funded OpenAI, the maker of ChatGPT, reportedly has a p(doom) between 20 or 30 percent. Still, Mr. Musk remains a proponent of AI. Marantz, A. (2024, March 11). Among the A.I. Doomsayers. The New Yorker. www.newyorker.com/magazine/2024/03/18/ among-the-ai-doomsayers.
[9] This is a hotly debated topic with some suggesting AI will create jobs and others claiming AI will destroy jobs. It’s difficult to know who is right. Regardless, consider carefully this idea that often comes up in AI discussions: “AI may not replace the lawyer, but the lawyer who uses AI will replace the lawyer who does not.” Or consider this comment from a recent MIT grad. She obtained degrees and math and computer science but pivoted to a degree in programming that is “closer to the hardware.” She said, “Absolutely it’s about staying ahead of the curve. A good rule of thumb is looking for jobs and skills that require [judgment] or research in some way.” Gen Z is entering the workforce with Generative AI skills. The Washington Post. (n.d.). www.washingtonpost.com/technology/2023/06/28/ai-gen-z-work/. That’s likely good advice.
[10] Marantz, A. (2024, March 11). Among the A.I. Doomsayers. The New Yorker. www.newyorker.com/ magazine/2024/03/18/among-the-ai-doomsayers.
[11] Harvard Business Publishing Education. (n.d.). https://hbsp.harvard.edu/inspiring-minds/ai-as-personal-tutor.
[12] Could AI cure cancer? Worldwide Cancer Research. (n.d.). www.worldwidecancerresearch.org/news-opinion/2024/february/could-ai-cure-cancer.
[13] What’s a Wicked Problem? Stony Brook University. (n.d.). www.stonybrook.edu/commcms/wicked-problem/about/What-is-a-wicked-problem.
[14] Marantz, A. (2024, March 11). Among the A.I. Doomsayers. The New Yorker. www.newyorker.com/ magazine/2024/03/18/among-the-ai-doomsayers.
[15] This is an overstatement. I acknowledge that it is. But not by much.
[16] Cardillo, A. (2024, May 5). How Many Companies Use AI? (New Data). Exploding Topics. https://explodingtopics.com/blog/companies-using-ai.
[17] Id.
[18] UCL (2022, February 2). Policy brief: AI-Enabled Future Crime. Dawes Centre for Future Crime at UCL. www.ucl.ac.uk/future-crime/policy-briefs/policy-brief-ai-enabled-future-crime.
[19] The creators of AI are responsible for the “alignment” of the software. AI alignment is the “process of encoding human values and goals into large language models.” Martineau, K. (2024, March 29). What is AI alignment? IBM Research. https://research.ibm.com/ blog/what-is-alignment-ai. Of course, who chooses these values and who chooses the goals makes all the difference.
[20] Blasius, M. (2024, April 26). How AI is stopping shoplifters, including here in the Valley. ABC15 Arizona in Phoenix (KNXV). https://www.abc15.com/news/local-news/investigations/how-ai-is-stopping-shoplifters-including-here-in-the-valley. A remarkable aspect of AI technology—and if you are reading the endnotes then perhaps you are the type of person to follow up on this—is that AI can be used to detect images from brain waves. Words will not be far behind. You could, for example, take a witness incapable of speech and “see” what he or she is thinking. The applications here are both thrilling and terrifying—but mostly terrifying.
[21] Johnson, A. (2023, October 4). Racism and AI: Here’s how it’s been criticized for amplifying bias. Forbes. https://www.forbes.com/sites/ariannajohnson/2023/05/25/racism-and-ai-heres-how-its-been-criticized-for-amplifying-bias.
[22] I recommend spending some time reviewing OpenAI’s product Sora. You can find it at https://openai.com/index/sora/. Sora creates realistic scenes from text instructions. Sora has its issues (which its creators are working on), but the application reminds me of Arthur C. Clarke’s quote that “any sufficiently advanced technology is indistinguishable from magic.” Some of the Sora images are, frankly, breathtaking, and they have caused a least one major Hollywood power player to put an $800 million studio expansion on hold. Kilkenny, K. (2024, February 23). Tyler Perry puts $800m studio expansion on hold after seeing OpenAI’s Sora: “Jobs are going to be lost.” The Hollywood Reporter. www.hollywoodreporter.com/business/ business-news/tyler-perry-ai-alarm-1235833276.
[23] There is a phrase often used in military circles when speaking about units or weapons: “Quantity has a quality all of its own.” In litigation, the same could be said for speed, where speed has its own quality, particularly in a trial setting.
[24] Perhaps one of you will create a tablet that sits at counsel table and identifies objectionable questions and answers and provides responses to objections using Texas law?
[25] This is actually a decent metaphor for AI. Of course, AI cannot (yet) make coffee, but it doesn’t drink coffee, either. But it can help you write an opening statement, and it never calls in sick.
[26] I did use AI for this article, but only for revisions of particular phrases or sentences. I use AI almost every day, but I still resist using AI for the overall structure for a project like this. That’s because I don’t want to lose what limited abilities I have to write. I would like to be a better writer, and I’m concerned that handing the reins completely over to AI, or even largely over to AI, works against that goal. This is a personal preference, and I could be entirely wrong. You will spot the completely original work of mine easily because it will be ungainly, unwieldy, and ungrammatical. Like that last sentence.
[27] “Listen, and understand! That Terminator is out there! It can’t be bargained with. It can’t be reasoned with. It doesn’t feel pity, or remorse, or fear, and it absolutely will not stop, ever, until you are dead!” Kyle Reese, MGM Home Entertainment. (2001). The Terminator.
[28] We have used AI in the following ways: to draft preliminary pleadings (Notice of Master Case File, Notice of Joinder, Motion to Stack, Notice of Intent to Use Business Records Aff.); to draft charge language in Promotion of Child Pornography Cases; to summarize Arrest Reports for attorney work-product form summaries; to re-write sentences and paragraphs in search warrants to increase clarity and succinctness and improve grammar; to draft language to permit the use of a drone before and during the service of a search warrant; to draft language for search warrants explaining complex topics like NAT’d IP Addresses, HASH collisions, the Onion browser, etc; to list names from reports to be included in discovery responses; to analyze a massive SnapChat spreadsheet (100,000+ lines) to identify useful information and generate word documents with chats between specific users; to analyze sentiment of emails to ensure the tone is what we intend; to create challenge coin artwork; to create PowerPoint graphics to illustrate specific points; to create a graph of the cases filed by each detective in a special division in the last 12 months; to create a transcript of a Kik chat from screenshots; to draft a paragraph for speaker bios to customize it for a particular presentation; to translate audio from Spanish to English and from English to Spanish; to analyze the transcript of an interview and identify inculpatory and exculpatory statements; to summarize jail calls; to create initial drafts for internal reports, position descriptions, etc.; to brainstorm trial strategies; and to age regress images for use in ICAC investigations. Again, special thanks to our King of AI, Chris Seufert, for many of these applications.
[29] “It isn’t only art that is incompatible with happiness, it’s also science. Science is dangerous. We have to keep it most carefully chained and muzzled.” Huxley, A. (2010). Brave New World (11th ed.). Vintage. (Ironically and perhaps fittingly, this is an argument for the advancement of science.)