The Bayan Ambassador Program
Where Egyptian law meets AI.
[date] · [author] · 5 min read
When Harvey launched its ambassador program in law schools, it asked students to be their classmates' go to resource on AI: running training sessions, organising events and meeting faculty, for pay and about 32 hours a term. We took the shape of the idea: a paid, flexible program with a new cohort each term and no prior experience required. We left out the part we do not need. A Bayan ambassador does not teach anyone or promote anything. Ambassadors work with us on the same problem we are building Bayan to solve.
The problem
A general model writes about Egyptian law fluently and gets it wrong confidently. It learned from text in which Egyptian law is rare, and it predicts the most likely word, not the correct one, so it fills what it does not know with what it saw often in other systems. Only someone who knows the law can see the error: an invented article looks real to someone who has not read the Civil Code, and a dropped condition goes unnoticed by anyone who does not know it was there. That is why Arabic legal AI needs lawyers as much as it needs engineers, and engineers who understand why «الفسخ» is not «الانفساخ».
What you will understand
Participants leave the program understanding three things together, because a single mistake in a model's answer often involves all three.
- The model: how it writes one word at a time, why it sounds sure even when it is wrong, where it is weak on Egyptian law, and what can be measured in its answer.
- The law: the text in force on the date of the facts, the right article with its conditions and exceptions, and the ruling with its full citation.
- The language: the barriers of Arabic legal text that mislead a machine: near synonyms with different rules, such as rescission and dissolution; the many spellings that grow from one root; a question in Egyptian Arabic answered from sources in formal Arabic; and translation between two legal systems.
One example holds all three. A general model answers: "Yes, the buyer may rescind the contract at once under Article 147, and dissolution occurs by operation of law." Confidence with no source is a mistake of the model. Article 147 instead of 157, and no formal notice, is a mistake of the law. "Dissolution" instead of "rescission" is a mistake of the language. Someone who sees all three at once is who we are looking for.
What ambassadors work on
- Writing Bayan Bench questions from real issues, with model answers and their sources.
- Grading model answers, and checking every citation against the text it claims.
- Logging and sorting where models fail in Arabic legal language: an invented article, a repealed text, a term in the wrong place, a dropped condition.
- Building the Arabic and English glossary, with the difference between the two systems and the article that governs each term.
- Tracking amendments, and knowing which text was in force on each date.
- Writing articles for the Bayan blog under their own names, reviewed by the team.
Two tracks
The legal track is for law students in their final years and recent graduates [Sherif confirms the years]. The engineering track is for engineering and computer science students, working on retrieval, evaluation and Arabic text processing. The two tracks work together: the lawyer writes the question and judges the answer, and the engineer builds what measures it and improves it.
What ambassadors get
- A monthly stipend [Sherif confirms the amount].
- Full access to Bayan for the length of the program.
- Their name on the work that is published: in Bayan Bench and on articles.
- Direct work with Bayan's legal and engineering team.
- A certificate and a reference at the end, and priority for roles at Bayan.
The details
[Sherif confirms before publication.] One term, renewable. About [number] hours a week, scheduled around your studies. Remote, with a meeting in Cairo [every two weeks]. [Number] ambassadors a term. You do not need to have used Bayan before: curiosity, care for accuracy, and the readiness to say "I don't know" when you don't are enough.
How to apply
- Apply on the program page and answer one question: an Egyptian legal question a general AI model got wrong, and why it got it wrong.
- A short task taken from the program's own work.
- A short interview with the team.
- Start with the cohort at the beginning of term.
We read applications in the order they arrive, so apply early.
Why a general model gets this wrong
A model cannot correct itself on what it does not know. Many teams use another model to grade answers instead of people, because it is faster and cheaper. But the grading model learned from the same kind of text, so it misses the same errors: it cannot tell an article is invented if it has never seen the real one. It also has documented biases: researchers publishing at NeurIPS in 2023 found that models judging other models' answers tend to prefer the longer answer, and the one shown first. In other words, they reward fluency again. A measure that rewards accuracy needs someone who knows the text, and that is an ambassador's work.
In Bayan
The program brings what used to be separate internship and ambassador programs into one. What ambassadors build goes into Bayan itself: the Bench questions, the glossary, and the error log the engineers use to know where to start. Apply at bayanlaw.ai/ambassadors, or write to ambassadors@bayanlaw.ai.
Sources. Harvey, "Law School Ambassador Program" · Zheng et al., "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena", NeurIPS 2023.