Legal Prompting Fundamentals
Reviewed by Human · Updated October 6, 2026
✗ Wrong way
Write a legal brief arguing that my client's claim isn't time-barred. Include supporting case citations.
The model doesn't retrieve cases from a database — it predicts text that looks like a citation. Ask for citations and it will happily invent case names, reporter numbers, and quotes that read perfectly and do not exist. In 2023 two attorneys were sanctioned after filing a federal brief full of cases a chatbot fabricated (Mata v. Avianca, Inc.).
✓ Better way
Draft the argument structure for a brief arguing my client's claim isn't time-barred. Use ONLY the three cases I paste below. Where a citation is needed but I haven't supplied one, insert [CITATION NEEDED] instead of writing a case name. Do not invent authorities.
You move the model off the one job it can't do (recalling real authorities) and onto the jobs it's good at: structuring an argument and marking exactly where a human must supply verified law.
What is the single most dangerous thing an AI can hand a lawyer?
Can You Use AI for Legal Work?
Yes — for the parts that aren't legal judgment. AI is strong at summarizing, structuring, comparing, and drafting. It is unreliable at recalling authorities and cannot give legal advice or tell you what the law *is* with authority. **Hallucination**: when a model generates confident, fluent output — a case, a statute, a quote — that has no basis in reality. The fix is a habit, not a better model. Every legal prompt should push the model toward analysis you can verify and away from facts it might invent. This builds on role prompting, where you cast the model as a specific kind of reviewer to sharpen its output.
Field note
Early on I asked a model to pull together authority on a fairly narrow indemnification question. The memo it produced was clean, cited four cases, and quoted holdings that fit my argument almost too well. One citation — a state appellate case — I couldn't find on Cornell LII or the court's own site. It didn't exist. Nor did one of the 'quotes.' Since then I run one non-negotiable rule: no citation leaves my desk until I've opened the actual opinion on a primary source and read the language myself. The model can point me at what to check; it never gets to be the source of truth.
Your turn
Rewrite this unsafe prompt so it cannot produce fabricated authorities. Add a caveat instruction and a marker for where a human must insert verified law.
Reflect
Notice how removing the demand for specific citations removed the incentive to fabricate them.
Why is 'list the cases that support my position' a dangerous prompt?
Go deeper
You've finished this module.
Mark it complete to earn your XP and keep your streak alive.
Progress saved locally · Sign up to earn XP