Short answer: Lawyers get sanctioned for AI because the tool made something up and nobody checked. Six lines added to your prompts fix most of it: make the AI verify its own work, give it permission to say “I don’t know,” make it ask questions before it starts, ban it from inventing facts, demand a citation for every claim, and get a second opinion from a different model. Copy the exact prompt lines below.
Written by Ed Elmahdy, founder of Rocket Launch Media. Last updated September 28, 2026.
Why this matters for an immigration firm specifically
Immigration work is heavy on filings, deadlines, and country-specific facts that change. That makes ChatGPT and Claude useful, and it also makes them dangerous. A model will happily produce a BIA decision that never existed, cite a USCIS policy memo with the wrong date, or summarize a client’s I-130 timeline with a step that is not in the file. It sounds confident every time.
Courts have stopped treating this as a novelty. The AI Hallucination Cases database maintained by Damien Charlotin now tracks more than 2,000 court decisions worldwide where fabricated citations or AI-generated errors showed up in filings, and the list grows weekly. Sanctions have ranged from fines to referrals to the bar. Every one of those lawyers thought the output looked fine.
The six rules
1. Make it verify its own work
Models are better at catching errors on a second pass than they are at avoiding them on the first. Tell the tool to review what it just wrote, and when it is pulling from documents you uploaded, make it point at the exact page or paragraph it used. Then go check that page yourself. If it cannot show you where a fact came from, treat the fact as unverified.
2. Give it permission to say “I don’t know”
A lot of hallucinations happen because the model does not have the answer but has been trained to be helpful, so it produces one anyway. Telling it up front that “I don’t know” is an acceptable answer removes the pressure. You would rather hear “I couldn’t find that” than get a confident paragraph about a regulation that does not exist.
3. Make it ask questions before it starts
When a prompt is unclear, the model assumes. Assumptions are where the output starts drifting toward things you never asked for, and in bad cases, toward things that are not true. One line fixes this: before it does anything, it asks you what it is unsure about. You answer, and then it works from your answers instead of its guesses.
4. Ban it from inventing facts
This is different from rule 2. Rule 2 waits for the model to admit uncertainty. Rule 4 is a standing order you give at the start: if a case, statute, memo, or date does not exist in the sources you provided, it is not allowed to create one. Put this in your custom instructions or project settings so it applies to every conversation, not just the ones where you remember to type it.
5. Demand a citation for every claim
A fake case reads exactly like a real one. The party names sound right, the reporter cite is formatted correctly, the holding is what you were hoping for. The only defense is a rule with no exceptions: no citation, no use. And the citation has to be something you can click or pull up and read yourself. A cite you cannot open is the same as no cite. This is where most of the sanctions in that database came from, and it is the easiest one to enforce.
6. Get a second opinion from a different model
Different models are trained differently and make different mistakes. If ChatGPT and Claude give you the same answer to the same question, that is a decent signal. If they disagree, that is your cue to stop and dig before you trust either. The manual version is copying the question from one window into another. It takes thirty seconds and it has caught errors for us more than once.
| Rule | What goes wrong without it | Where to put it |
|---|---|---|
| Verify its own work | Errors survive because nobody asked for a second pass | End of any research or drafting prompt |
| Permission to say “I don’t know” | Model guesses instead of admitting a gap | Custom instructions / project settings |
| Ask questions first | Model assumes and drifts off task | Top of any multi-step task |
| Ban invented facts | Fabricated cases, memos, and dates | Custom instructions / project settings |
| Citation for every claim | Fake authority makes it into a filing | Any legal research prompt, no exceptions |
| Second opinion | One model’s blind spot goes unchallenged | Before anything gets filed or sent to a client |
How we use this at Rocket Launch Media
We build marketing systems for immigration law firms, and AI is in most of what we produce: ad copy, intake scripts, follow-up sequences, blog content. Rules 3, 4, and 5 sit in the custom instructions of every project we run for a client. Rule 6 runs on anything that names a law, a form, or a deadline before it goes live on a client’s site. The rules were written for lawyers, and they apply just as well to the marketing side, where a made-up statistic on a landing page is its own kind of liability.
Get the free AI guide for immigration law firms
Frequently asked questions
Can immigration lawyers use ChatGPT for legal research?
Yes, with the rules above in place and with every citation checked by a human before it is relied on. General-purpose models are not connected to a legal research database, so they can produce cases that do not exist. Treat them as a drafting and summarizing tool, and verify anything that would go in a filing.
Is Claude safer than ChatGPT for lawyers?
Both hallucinate. Claude tends to be more willing to say it does not know, which helps, but neither is safe without the rules above. The safest setup is using both and comparing answers on anything that matters.
Where do I put these rules so I do not have to retype them?
In ChatGPT, use Custom Instructions or a Project’s instructions. In Claude, use a Project’s custom instructions or a Style. Rules 2, 4, and 5 belong there permanently. Rules 1, 3, and 6 are easier to add per task.
What happens if a fake citation gets into a filing?
It depends on the court, but the recorded outcomes include monetary sanctions, orders to notify the client and the judge whose name was attached to the fake case, continuing education requirements, and bar referrals. The AI Hallucination Cases database linked below tracks the individual decisions.
Do I need a paid legal AI tool to be safe?
No. Paid tools that pull from real case databases reduce the risk, but they do not remove it, and the six rules cost nothing. Start with the rules. Add a paid tool if your research volume justifies it.
Sources: AI Hallucination Cases database, maintained by Damien Charlotin · The six-rule framework is adapted from a September 2026 video by Andrew Stickel (SMB Team); the prompt wording and immigration examples are ours.
Related reading: AI guide for immigration law firms · Law firm client intake · Law firm marketing automation · AI receptionist vs. answering service
