Yuval Avidani
Author
Let's fly high, but with our feet planted deep in legal ground. Picture a work evening: a 40-page contract waiting to be reviewed, a court ruling to summarize, and a demand letter to draft by tomorrow morning. AI, meaning artificial intelligence software that generates text, like ChatGPT or Claude, can cut that evening in half. But that exact same tool can also burn your license if you feed it the wrong thing, or if you trust it where you're not allowed to.
The whole key lies in one distinction: AI is a brilliant research assistant, not a source of authority. What does that mean in practice? A research assistant is a hardworking intern who brings you raw material fast, but you, the lawyer, remain solely responsible for every word that goes out under your name. A source of authority, on the other hand, is something you cite without checking, and that's exactly what AI must never be. This entire guide is really just an expansion of that one sentence.
Why bring a machine into legal work at all
Because the machine doesn't get tired on page 38, doesn't miss a clause because it's 11 PM, and can read 40 pages in seconds and point out patterns. The flip side: it's also "confident" exactly when it's wrong. So the real gain isn't replacing your judgment, it's freeing up your time from repetitive tasks for what a machine can't do: strategy, nuance, client relationships.
The three tasks where AI actually saves you hours
1. Drafting a first version. The hardest part of drafting is usually the blank screen, that moment of staring at an empty page. AI neutralizes exactly that: you give it context, and it hands you back a skeleton to edit. An AI draft is raw material for editing, never a finished product you send as-is. Why? Because the model doesn't know the specific facts of your case, doesn't know what the client actually wants, and sometimes invents phrasing that sounds legal but isn't valid under Israeli law.
2. Summarizing and comparing. Paste in two versions of a contract and ask: "Compare the two versions and flag every material change in liability allocation, liquidated damages, and cancellation rights, in a table." What AI does here is read both texts word by word and map the differences, a tedious task for a person and a fast one for a machine. "Liability allocation" = who bears the damage if something goes wrong. "Liquidated damages" = a sum set in advance in the contract as compensation for breach, without needing to prove actual damage. And still, every difference the machine flags, you verify with your own eyes against the source.
3. Explaining and rephrasing. "Rewrite this clause in plain language for a client who isn't a lawyer" is a great communication tool. But remember, this is a machine's interpretation, and the binding legal interpretation is yours alone.
Now let's see the difference between a weak prompt (= the instruction you write to the model) and a strong one, because this is where most users fall down:
Enemy number one: legal hallucinations
A "hallucination" is the term for when an AI model invents information and presents it with total confidence as if it were fact. Why does this happen? Because the model isn't a legal database that returns citations, it's an engine that predicts the most likely next word based on billions of texts. So when you ask "give me a supporting ruling," the model will generate text that looks exactly like a court decision, complete with party names, case number and year, even if it doesn't exist.
This isn't a theoretical scenario. Lawyers in the US have already been fined and publicly reprimanded after filing court briefs that cited rulings the AI "invented." The damage is double: you lose the case, and your credibility with the judge takes a hit too.
The unbreakable rule: every citation, whether it's a court ruling, a statute, or a regulation, gets verified against the official source before it enters the document. In Israel, that means Nevo, Pador, or the Judicial Authority's website. If you didn't find the ruling in a real database, it doesn't exist, period. The AI gave you a direction to search, not proof.
The red line you don't cross: attorney-client privilege
Attorney-client privilege is your legal duty to keep secret everything the client has told you. This isn't politeness, it's an ethical obligation, and breaching it is a disciplinary offense. And here's the danger: when you type text into a public AI tool, that information leaves your computer and goes to another company's server, and sometimes gets stored or used to train the model.
What does "training the model" mean? A process where the company uses the texts fed into it to improve the next version of the model. In other words, a detail from your client's confidential contract could, in theory, surface in an answer to a different user. Even if the odds are small, for a lawyer the exposure itself is the breach.
What you must never feed into a public tool: real client names, ID numbers, entire confidential documents, active case numbers, or any information whose identification would harm the client.
So what do you do? Anonymize. Anonymization = stripping out every identifying detail before you type. Replace names with "Party A/Party B," delete ID numbers and addresses, and feed in only the abstract legal phrasing. Want more security? Two paths: (a) an organizational arrangement, an Enterprise/Team account with a contractual commitment from the vendor not to train on your data and not to store it; (b) a local model, AI software that runs on the firm's own computer, so no text ever leaves the building. Anonymizing before you type isn't a suggestion, it's the first line of defense for privilege.
A workflow that respects both hats
You're wearing two hats pulling in opposite directions: the efficiency hat (let AI do the work) and the responsibility hat (don't trust it). The following workflow reconciles the two, and every step is a link that protects both your time and your client:
Why this matters, and where to start today
The gap has already opened up. The lawyer who uses AI correctly finishes in one evening what takes a colleague two days, freeing up time for strategy, judgment, and client relationships. But that exact same tool, used carelessly, produces disciplinary and ethical damage that can't be undone. The difference isn't the technology, it's discipline: research assistant, yes; final source of authority, never.
My practical recommendation: don't start with your most sensitive case. Today, take a small, non-sensitive task, say, summarizing a public legal article or drafting the skeleton of a generic clause, build yourself a consistent verification routine, and only then shift up a gear. Let's fly high, without leaking and without inventing.