Synthetic Legal Data Generation for Saudi Family Law Courts Using Multi-Role Simulation
This paper proposes a multi-role simulation framework using Large Language Models and Human-in-the-Loop validation to generate a high-quality, expert-verified synthetic dataset of Saudi Family Law divorce cases, effectively overcoming the scarcity of real judicial records caused by privacy constraints.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where computers can read, write, and argue like lawyers. This is the exciting, slightly scary frontier of "Legal Artificial Intelligence." But here's the catch: for a computer to learn how to be a lawyer, it needs to read millions of real court cases. In most places, this is easy because court records are public. But in some places, like Saudi Arabia, family court cases are super secret to protect people's privacy. It's like trying to teach a chef to cook a secret family recipe without ever letting them taste the food. Without real data, the computer can't learn the rules, and if it guesses, it might make up laws that don't exist. This paper tackles that exact problem: how do you build a smart legal brain for Saudi family courts when you can't look at the real cases? The answer involves creating a "fake" but perfect world where computers can practice, learn, and get it right without ever breaking anyone's privacy.
The researchers behind this study, Zarah Shibli, Areej Alhothali, and Lulwah Al-Harigy from King Abdulaziz University, decided to build a digital playground to solve this puzzle. They focused on a specific, tricky type of case: when a wife goes to court to prove that her husband actually divorced her, even though he never wrote it down or told her officially. In Saudi law, this is called "Ithbat Talaq" (establishing divorce). It's a legal minefield because the husband might say, "I never said that," or "I was just joking," while the wife says, "He definitely said it!"
To create a training dataset for their AI, the team didn't just ask a computer to "make up a story." Instead, they built a sophisticated role-playing game. They programmed a powerful AI (specifically, a model called Gemini 3.1 Pro) to act out a full courtroom drama. But here's the twist: the AI didn't just play one character. It had to switch hats and play everyone in the room at the same time. It had to be the Judge (who has to be strict and follow the law), the Wife (the plaintiff), the Husband (the defendant), and even the Court Clerk (who writes everything down).
The AI was given a "script" based on real Saudi laws, including the Family Law and the Law of Evidence. It was told to generate 200 different scenarios. In some, the husband admitted he divorced his wife; in others, he denied it. Sometimes the divorce was said out loud, sometimes it was written, and sometimes it was conditional (like, "If you leave the house, you are divorced"). The AI had to figure out the legal outcome for each of these 200 complex situations, strictly following the rules of Saudi law.
But the team knew that a computer playing lawyer might get the facts right but sound like a robot or miss a tiny legal nuance. So, they added a "Human-in-the-Loop" step. This means real human legal experts—six of them, who are specialists in Saudi family law—read every single one of the 200 AI-generated cases. They graded them on a scale of 1 to 5 for things like "Did the judge use the right legal words?" and "Was the reasoning logical?"
The results were impressive. Before the humans stepped in, the AI was already pretty good at the heavy legal lifting, but it sometimes got a bit wordy or used slightly off-key legal phrases. After the human experts tweaked the AI's work, the results were nearly perfect. The experts agreed with each other on the quality of the AI's work 94.6% of the time (when allowing for tiny, one-point differences in scoring). In fact, the AI-generated cases scored a perfect "5" on the experts' rubric in over 70% of the evaluations.
However, the paper is careful not to say this is a magic wand that replaces judges. The AI still struggled a bit with the most complex, edge-case scenarios where the law gets very nuanced, like determining if a divorce was valid when the husband was extremely angry or when the wife was in a specific physical state. The researchers found that the AI is best used as a super-powered assistant that drafts the legal reasoning, which a human expert then reviews and polishes.
In short, this paper shows that by using a multi-role simulation—where an AI acts out a whole courtroom drama—and then having human experts double-check the work, we can create high-quality, privacy-safe legal data. This "synthetic" data is a game-changer because it allows future AI systems to learn how to navigate Saudi family law without ever needing to peek at real, private court files. It's like building a flight simulator for lawyers: the AI can crash and burn in the simulation, learn from its mistakes, and be ready for the real sky, all while keeping everyone's secrets safe.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.