A few years back, as the AI boom was just beginning to rumble, then-interim and now University of Miami School of Law Dean Patricia Sanchez Abril proposed the creation of a University lab to prepare and empower the next generation of lawyers for the seas of change to come and those already reshaping the legal terrain.
A search for a law and technology fellow led to Or Cohen-Sasson, whose professional background as a geneticist by then had transitioned to computer science and law. Cohen-Sasson was living in Tokyo, Japan, at the time, completing postdoctoral research on the intersection of law and how intellectual property regimes may impact AI development—some of the seminal academic research in that arena.
"It seemed a wonderful opportunity to lead and pioneer in the AI and law domain because back then there was only one technology and law lab in the country," said Cohen-Sasson, today a lecturer and director of the Miami Law and AI Lab (MiLA). "We decided that we could be the first, not only in terms of research, but also in preparing the next generation of lawyers. It was clear to me, too, that we needed to act on the educational level as well."
The lab, officially established in July 2024, has grown to include approximately 40 students, about half from the law school along with others from computer science, engineering and the business school. Students work in small teams of diverse skill sets, each pursuing a research or development project.
Students can receive academic credit for their participation, and some have been hired as research assistants and developers.
Cohen-Sasson classified the lab's endeavors into three main "buttons": research, development and education. Research is both academic and geared to formulate policy.
"The questions we're interested in are mostly in two directions: How AI could or already is actually impacting legal frameworks and the way we think about concepts, and the other is the way that the law should or could regulate the way we build AI so that future AI systems are not detached from the social principles we want to keep in society," Cohen-Sasson explained.
A paper recently presented at the International Conference on Machine Learning, one of the top three conferences worldwide on AI, focused on the issue of access to justice and AI.
Cohen-Sasson highlighted that with the advent of AI, an increasing number of AI-assisted claims are being filed with the courts. MiLA Lab researchers analyzed a data set of 3 million plus cases for the period 2008 to 2025 and documented a 50 percent increase in these pro se litigant filings since 2023, along with a 100 percent hike for the last quarter of 2025.
"One of the promises for AI in legal scholarship was that once AI becomes publicly available, people without previous access to the courts because of lack of know-how or financial constraints now would have tools to help access different forms and courts, and to hopefully accomplish the remedy they're looking for," he noted.
Yet while there has been a surge of pro se litigant filings, the percentage of claims being rejected by the courts has spiked.
Previously a lawyer would act as a preliminary judge filtering claims for validity and substance before they were filed, Cohen-Sasson pointed out.
"Now that this first filter has dissolved, the entity that needs to do the filtering is the court itself, so it has many more cases—and a higher proportion are weaker—so you actually get more dismissed," he explained.
The phenomenon of "sycophancy" is at play here, he noted. AI acts as a "yes man"—it aligns with your argument and reinforces your approach, opinion and bias.
"This is one example of how AI impacts the law. The lawyer [as filter] used to be objective but now AI actually strengthens your belief, making you overconfident. You feel encouraged to go and file what may be a very weak complaint," said Cohen-Sasson.
Another MiLA Lab research initiative has focused on "computer threshold." This regulation monitors the computation power that AI models consume—and serves as a proxy to how dangerous a model could be, Cohen-Sasson explained.
"One of the main problems in the AI domain is alignment. If the model is misaligned with the values of our society, it could go against society, starting with small problems such as not carrying out the task you asked for and potentially creating far more dangerous scenarios," said Cohen-Sasson, adding that "the smarter the model, the harder it is to make it compliant with our instructions."
The lab has conducted research on the current regulation, which aims at only large models such as Claude Fable 5, Anthropic's flagship autonomous AI model, using floating-point operations, or FLOPs, a computational measurement tool.
The assumption is that the more FLOPs, the riskier the model, Cohen-Sasson noted.
"But this regulation can easily be bypassed. Small models working separately operate well below the FLOP threshold, but when we let them work together, they can outperform the larger model—the one being regulated—so you can create a small army of AI agents that are more dangerous than the larger monitored model," he said. "Basically, we show that the compute threshold regime is not successful and not even relevant in the age of agentic AI."
In addition to this novel research, the lab is developing AI tools for the legal domain—scholarship, education and legal practice.
One of the main concerns in the legal domain is confidentiality. The lab's Attorney's Encryption Guard for Information Security (AEGIS) tool serves as a filter or layer between the AI model and the user, encrypting and redacting all the personal information in your prompt.
"Our AEGIS model only sends the encrypted version to Claude or ChatGPT, then whatever model you're using sends back a response. Everything AEGIS does stays on your computer so that your confidentiality and privacy are protected," he noted.
Moot Court XR is another tool the lab has developed. Moot courts competitions are major events nationally and worldwide, and the School of Law has historically sported a strong moot court team—currently #1 in Florida and #3 nationally.
Through collaboration with the University's Virtual Experiences Simulation Lab (VESL), students preparing for moot court competitions use immersive virtual reality to practice .
"We can predesign the sessions for whatever theme they're working on, then based on info the models will be able to ask relevant questions," Cohen-Sasson explained. "We can provide a full 20-minute intensive training for less than $1. You get feedback and a score, and you can do it as many times as you want."
In the educational realm, the lab has developed an automated tool that has a patent pending. Lawyers presenting any scholarly paper must follow a rigorous specific standard for citations: the Blue Book. Students or professionals can spend hours translating their citations into the Blue Book format. The lab is building a software to automate this laborious process.
Cohen-Sasson shares the concerns that many have that the increasing use of AI may ultimately sap or debilitate other critical skillsets.
"One of the main problems with AI, not necessarily for law students but for the younger generation that are in school, is that it's so easy to go to AI first without thinking," he said. "You skip the very important step of creating your own opinion or strategy, and so gradually you may lose this capacity to think about the task for yourself.
"What I'm trying to do with my students is to encourage them to use AI—though not in a detrimental way, in a way that will de-skill. Rather to use it as a collaborator or an assistant as opposed to something that should dictate how they carry out a task," Cohen-Sasson said.
MiLA lab is targeting events for non-student audiences such as judges and clerks as well as collaborations with other University departments such as computer science; institutes such as UHealth's Bioethics and Health Policy, researching the effects of AI on health law; and The Launch Pad, the University's entrepreneurial hub.
