Thinking like a lawyer in the age of AI
Five common assumptions about how lawyers are formed, and what the reality is.
[date] · [author] · 8 min read
Ibn Khaldun wrote in the Muqaddimah that a skill, a malaka, is a settled quality that comes only from doing a thing again and again until its form takes root. Legal judgment is that kind of skill. Nobody learned to read a text the way a lawyer reads it from one lecture. They learned it from the hundreds of times they read an article and read it again, searched for a ruling and failed and then found it, and drafted a statement of claim that came back from a senior lawyer with red ink in every margin. So what is left of that skill when a machine does the repetition for us?
Harvey put this question to the dean of Texas Law, the director of AI at BU Law, a professor of legal practice at Georgetown and a partner at a US law firm. The conversation produced five common assumptions that do not hold as stated. We look at them here from Egypt: from law faculty lecture halls that seat thousands, and from the years of training in law offices where a lawyer learned the job by doing it.
Assumption 1: AI is making law students worse thinkers
Reality: the picture is more complicated.
Critical thinking is hard to measure in the first place, and faculties rely on imperfect proxies for it. An exam that rewards recalling the text measures memory before understanding, with or without AI. A tool can even expose gaps in understanding instead of creating them: a student who asks it to argue against their position sees what they missed, and where their conclusion rests on something they never proved. The real risk has a name, cognitive surrender: handing the tool the part of the work the student should have thought through, until they read conclusions instead of reaching them. The worry is as old as writing itself; Plato feared that writing would weaken memory. The answer then was not to stop writing, but to learn to write and remember at once.
Assumption 2: AI is automating junior lawyers out of a job
Reality: the routine work that used to teach the job is shifting, so the training has to be redesigned.
An Egyptian lawyer learned the profession in the training years: searching the Court of Cassation's Technical Office collections page by page, drafting a statement of claim and getting it back corrected, attending hearings and following the court list, and hearing a senior lawyer explain why one defence went first and another last. Much of that slow work was the teaching, even if nobody called it that. When a tool does the first search and the first draft in minutes, the junior lawyer does not disappear, but the path they learned on does, unless another one is designed on purpose. Every firm should ask: what must a trainee still do by hand, even when the tool is faster?
Assumption 3: prompting is just a technical skill
Reality: a strong prompt applies legal judgment.
Type "what is the rule on penalty clauses?" and you get a general answer that fits every contract and none in particular. Type "An apartment sale contract governed by Egyptian law; the buyer is two instalments late, and the contract sets a monthly penalty of twice the instalment. Which articles govern the penalty? What must each side prove? Then argue against your own conclusion" and you get an analysis. The difference is not technique. It is knowing what matters in law: the decisive facts, the governing law, who carries the burden of proof, and the counter argument. A strong prompt tests the argument, surfaces its weaknesses, asks for uncertainty to be stated plainly, and leaves the final decision to the lawyer.
Assumption 4: schools must choose between AI literacy and traditional legal reasoning
Reality: both can be built at once, if they are built deliberately.
The two goals are not in a zero sum relationship. One exercise teachers who have tried this suggest: the student analyses a problem first without any tool, then returns to it with the tool, then writes down what changed in their analysis and why. Alongside the work with the tool, protected time for independent thinking: an oral discussion of a ruling, a clause drafted by hand, an article read until the student can explain it without looking.
Assumption 5: the goal is to make legal work faster
Reality: the value is in making it better.
Speed alone is an incomplete target. AI helps most when it lets a lawyer test more angles, find the weak points, and reach a more complete analysis. The minutes a tool saves belong to checking and thinking, not to one more file done in a hurry.
Why a general model gets this wrong
A general model invites surrender by the way it is built. It predicts the most likely next word and is then tuned toward the answers people prefer, so its text comes out smooth and confident every time, right or wrong. And people read smoothness as truth: in a well known study published by Reber and Schwarz in 1999, participants judged the same statement more likely to be true when it was easier to read. So one answer can carry fluency that reassures the reader and an error that only someone who knows the article will see. It gets harder because models tend to agree with whoever is asking, as Anthropic's researchers found in 2024: ask "isn't this clause void?" and the model leans toward yes. A 2025 study by Michael Gerlich found that heavier reliance on AI tools went with lower critical thinking scores, through what he calls cognitive offloading. That is a correlation and does not prove cause on its own, but it is reason enough for care. So reading the answer is not enough. Ask for the counter argument, go back to the text of the article yourself, and phrase the question neutrally so it does not suggest its own answer.
In Bayan
Bayan is built against surrender, not for it. Every answer comes with the article as published and the ruling with its full citation, so checking is one click, not a new search. When the sources are not enough, Bayan says so instead of filling the gap. And in the Bayan Ambassador Program, law students work on the same question from the other side: measuring where models get Egyptian law wrong, and why.
Sources. Harvey, "Thinking Like a Lawyer in the Age of AI", 25 September 2026 · Reber and Schwarz, "Effects of Perceptual Fluency on Judgments of Truth", Consciousness and Cognition, 1999 · Sharma et al., "Towards Understanding Sycophancy in Language Models", Anthropic, ICLR 2024 · Gerlich, "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking", Societies, 2025.