AI drafts fast and invents confidently. The skill is knowing which is which.
After this, you'll be able to sort legal tasks into where AI genuinely helps and where its raw output is risky, and you'll run every answer through one human verification step before it leaves your desk.
The idea
If you suspect AI is overhyped and cannot handle real legal work, start here: it already turns a forty-page contract into one clear page in about a minute, and it drafts a first-pass memo while you are still finding the file. That is not a demo trick. It is boring, billable work it takes off your desk today, which is the honest reason to keep reading.
Here is the other half, the part the headlines got right: the same tool will invent a court case that never existed and hand it to you with total confidence. Both are true at once, and the whole skill, the one this track keeps coming back to, is telling them apart before anything goes out the door.
Start with what it does well. AI is a fast first-pass drafter and summarizer. It can turn a blank page into a rough memo, condense a long contract into plain English, or take a stack of documents and pull out the parts you asked about, in minutes instead of hours.
Those are first-draft tasks, where you already know enough to judge the result. That is the safe zone, because you are the expert checking the work, not a beginner trusting it.
The dangerous zone is anything you cannot personally check. Here is the failure that has ended up in court filings: AI invents legal authority. It produces a hallucination (a confident, made-up answer that looks real) in the shape of a real case, with a plausible name, a plausible court, and a plausible quote, none of which exist.
It does this because a model (the AI system behind tools like ChatGPT or Claude, also called an LLM, a large language model) predicts text that fits the pattern of real law, not text it has verified. A fake citation matches the form of a real one perfectly, which is exactly why you cannot catch it by reading.
This is not theoretical. Lawyers have faced sanctions (formal punishment from a court, such as fines or a public reprimand) for filing briefs that cited cases the AI made up. In the widely reported Mata v. Avianca matter (2023, in the Southern District of New York), attorneys submitted fabricated cases that opposing counsel and the judge could not find, because the cases did not exist. Filing fabricated authority is how routine AI use becomes malpractice (a professional failure serious enough to expose you to liability).
Here is the before and after: in the risky version, you ask AI for cases supporting your position, paste its answer into a brief, and file it. In the safe version, AI writes the draft, then you open each cited case yourself, confirm it exists and says what the AI claimed, and only then does a qualified human sign off and file.
That safe version is not extra steps bolted on. It is a short, repeatable pipeline you run on every AI answer before it leaves your desk.
The pattern that makes AI safe is the same every time. AI drafts, a qualified human verifies and signs off, always.
Get that habit down here, because the next lesson adds the rule that protects your client before you even hit send: what privilege is, and the single careless paste that can quietly waive it.
| Safe zone (you can check it) | Danger zone (you cannot check it) | |
|---|---|---|
| Best for | First drafts, plain-English summaries, first-pass review | Inventing cases, citations, holdings, or exact clause wording |
| Why | You are the expert judging a rough draft | The AI predicts plausible text, so fabrications look real |
| How to catch errors | Read it the way you would review a junior's draft | You cannot, by reading; you must open the primary source |
| Before it leaves your desk | A qualified human verifies and signs off | A qualified human verifies and signs off |
The rule is the same in both columns. AI drafts, a qualified human verifies and signs off, always.
Try it (16 min)
Watch out for
Paste this into Claude
Pick a small, low-stakes legal task you already understand well, so you can judge the result yourself. Good examples: "summarize this publicly available statute in plain English," or "draft a short, generic client-update email about a filing deadline." Do NOT use anything client-confidential for this first run. Setting up a safe tool and protecting client data is the very next two lessons; for now you stay on public, non-client text on purpose, so you can focus on the one habit this lesson teaches, verifying the output. Paste this into your AI tool (ChatGPT, Claude, or your firm's approved tool): "You are helping a lawyer with a first draft only. I will verify everything before it is used. Task: [paste your task, e.g. 'Summarize the attached public statute in plain English for a non-lawyer client.'] Rules: 1. If you state any rule, deadline, case, or citation, list each one separately at the end under a heading 'CLAIMS TO VERIFY'. 2. For anything you are not certain about, say so plainly instead of guessing. 3. Do not invent case names, citation numbers, or quotes." When the answer comes back, take the 'CLAIMS TO VERIFY' list and check each item against a real source (the actual statute, the court's website, or an official database). Mark each one confirmed or wrong.
What good looks like
When this breaks
You can now
Run a real (non-confidential) legal task through AI, get a draft plus a list of every factual claim, and verify each claim against a primary source before using any of it. If you find yourself trusting a citation because it looks right rather than because you opened it, that is the one thing that is wrong: a fabricated cite is built to look right, so the only valid check is opening the actual source and reading it yourself.
Key takeaways
AI is a fast first-pass drafter and a confident fabricator at the same time. It helps most on tasks you can personally check, and it is dangerous on anything you cannot. The fix never changes: AI drafts, a qualified human verifies and signs off. Next, the rule that protects your client before you even hit send: privilege, and the one paste that waives it.