You ask a model for three cases supporting a proposition. It gives you three, with court names, years, reporter citations and a tidy sentence on each holding. They read exactly like real authority. One of them doesn't exist, one exists but says something different, and one is fine. Without opening the sources you cannot tell which is which, and the polish of the output gives you no clue.
That is the whole problem with AI in legal research, and it's why this post is organised around one rule: drafting and verification are separate jobs, and nothing leaves the draft stage until the verification log is complete. The prompts below help you build a memo faster. The workflow stops the fabricated parts from reaching a court or a client.
This is general information, not legal advice. Rules on AI use, disclosure and confidentiality differ by jurisdiction and court; check yours. For broader prompts (contracts, client emails, deposition outlines), see the earlier prompt engineering for lawyers post. This one is deliberately narrower: research memos and their checks.
What actually happened when citations were fabricated?
The best-documented example is a federal case in New York. In Mata v. Avianca, Inc. (S.D.N.Y., No. 22-cv-1461), Judge P. Kevin Castel issued an Opinion and Order on Sanctions on 22 June 2023. I checked the quotations below against the text of the order (Document 54, 43 pages, a RECAP/CourtListener copy of the court filing) on 2026-10-09. Its opening says there is "nothing inherently improper about using a reliable artificial intelligence tool for assistance," but that "existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings." It found the respondents "abandoned their responsibilities when they submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT, then continued to stand by the fake opinions after judicial orders called their existence into question." The court imposed a $5,000 penalty, jointly and severally, on the lawyers and their firm.
Two lessons carry over to any jurisdiction. First, the court did not treat AI use as the offence; it stressed the lawyers' duty to check what they file. Second, the sanction rested on a finding of bad faith by the individual lawyers, based on what the order calls "conscious avoidance and false and misleading statements to the Court" after the citations were questioned (the order says they "doubled down"). If a citation is challenged, stop and verify before replying. This is a US federal ruling under Rule 11; it is not a statement of what any other court would do.
I am only relying on this one order. There have been other reported incidents, and courts and bar bodies have issued guidance since, but I haven't read those primary sources and won't describe them. Search your own court's current practice directions; they change.
The workflow: draft, extract, verify, log
Four stages. The model helps with the first two. You do the third and fourth.
- Draft the memo from authorities you supply, or as a scaffold with every authority marked unverified.
- Extract every authority and every quotation into a list.
- Verify each item against a primary or trusted source.
- Log the result and only then promote the memo to "reviewed".
Prompt 1: a memo scaffold that refuses to invent authority
The safest research prompt supplies the authorities. You pull the cases from your database, paste the relevant passages, and have the model organise them.
You are helping me draft an internal legal research memo.
QUESTION PRESENTED: [QUESTION]
JURISDICTION AND COURT LEVEL: [JURISDICTION]
FACTS (anonymised, no client names): [FACTS]
AUTHORITIES I HAVE PROVIDED (the ONLY authorities you may rely on):
<authorities>
[PASTE EACH CASE OR STATUTE EXTRACT WITH ITS CITATION AND PINPOINT]
</authorities>
Write: Question Presented, Brief Answer, Facts, Analysis, Conclusion.
Rules:
- Cite only the authorities in the block above, using the citations exactly
as I gave them.
- If the analysis needs authority I haven't provided, write
[AUTHORITY NEEDED: describe the proposition]. Do not name a case.
- Quote only words that appear in my pasted extracts. Mark each quote with
the pinpoint I gave.
- Where my authorities cut against the conclusion, say so.
- Flag any place where you are inferring rather than reading.
The [AUTHORITY NEEDED] marker is the useful part. It turns a model's urge to fill the gap into a research to-do list.
Prompt 2: issue spotting without citations
When you're starting cold, ask for issues and search terms, not cases.
Facts (anonymised): [FACTS]
Jurisdiction: [JURISDICTION]. Area of law: [AREA]
List the legal issues a court could treat as relevant, as questions. For each:
- The elements or test I should look for (describe generally; name no cases)
- Search terms and database filters I could use to find authority
- What facts are missing that would change the analysis
Do NOT cite cases, statutes by section number, or quotations. I will find
authority myself.
This gives you an outline and search strings, which is where a model is strong. It removes the temptation to trust the one output type that is most likely to be fabricated.
Prompt 3: extract every authority into a checklist
Once you have a draft, whether you wrote it, a colleague did or a model did, convert it to a verification list.
Read the memo below. Output a table with one row per authority or quotation:
| # | Citation as written | Proposition it is cited for | Quoted text (if any) | Pinpoint | Where in memo |
Include statutes, rules, regulations and secondary sources, not just cases.
Do not correct, complete or "fix" any citation. Copy exactly what is written.
If a citation is incomplete, write [INCOMPLETE] in the citation cell.
If a quote has no pinpoint, write [NO PINPOINT].
MEMO:
[PASTE MEMO]
"Copy exactly" matters. A model told to tidy citations may silently change a year or reporter, which hides the error you want to catch.
The verification checklist
For every row of that table, do the following in a source you trust: the court's official site, an official reporter, or a commercial service your firm licenses. Don't verify a model's citation by asking the same or another model "is this real?"
| Check | What to do | Fail means |
|---|---|---|
| Existence | Search the exact citation and the party names | Remove it; it may be fabricated |
| Match | Court, year, reporter, volume and page all agree | Fix it, then re-check the rest |
| Text | The quoted words appear in the opinion at the pinpoint | Delete the quote or correct it |
| Holding | Read the passage. Does it support the proposition as written? | Reword the proposition or drop the case |
| Level | Binding, persuasive or irrelevant for your court? | Re-label, or find better authority |
| History | Check later treatment in a citator (appeal, overruled, distinguished) | Reconsider or flag |
| Currency | Statutes and rules: is this the current version? | Update text and date |
Three habits help. Open the opinion and read the paragraph, not just the headnote. Copy quotes from the source, not from the memo. And record the date you checked, because law moves.
Prompt 4: a verification log you fill in, not the model
You can have the model build the log skeleton from Prompt 3's table, but the content in each cell is yours.
Turn this authorities table into a verification log with these columns,
leaving every verification cell blank:
| # | Citation | Exists (Y/N) | Court and year match (Y/N) | Quote found at pinpoint (Y/N/NA) | Supports proposition (Y/Partly/N) | Later history checked (Y/N) | Checked by | Date | Notes |
Do not fill in any cell after "Citation". Do not guess.
The memo's status stays "DRAFT, UNVERIFIED" until every row has a "Checked by" and a date. Put that in the document header in capitals. Habits die; headers don't.
Prompt 5: adversarial review of your own memo
Use a model as a sceptical reader. This catches reasoning gaps, not fake cases.
Act as a sceptical senior associate. Read the memo below and list:
1. Propositions that are stated as settled but seem to depend on a single
authority.
2. Places where the analysis skips an element of the test.
3. Facts the conclusion relies on that the Facts section doesn't state.
4. The strongest counterargument and what authority type (not a named case)
would answer it.
Do not add or suggest citations.
MEMO: [PASTE MEMO]
For the habit of cross-checking anything a model asserts, the general fact-checking checklist and the hallucinations deep dive explain why fluent output and true output are different things.
Confidentiality before you paste anything
Research prompts rarely need client identity. Use hypotheticals or anonymised facts: roles instead of names, rounded figures, no case numbers. Before pasting client material, check your professional duties of confidentiality, any client consent or engagement terms, your firm's policy and the tool's data and retention settings. Our vendor retention table shows how to compare those settings, and what not to paste into ChatGPT covers the habit generally. Whether a given use is permitted is a question for your regulator or ethics counsel, not this page.
Failure modes
| What goes wrong | Why | What to do |
|---|---|---|
| Real-looking case that doesn't exist | The model produces plausible text, not a database lookup | Never accept an authority you haven't opened |
| Real case, wrong holding | Summaries blur facts and rulings | Read the passage, not the summary |
| Quote not in the opinion | Paraphrase presented as quotation | Copy quotes from the source |
| Right case from the wrong jurisdiction | Common-law cases look alike | Check court, level and binding force |
| Overruled authority | The model's knowledge has a cutoff | Citator check with a date |
| Confident answer to a question with no law | The model prefers answering | Use [AUTHORITY NEEDED] markers |
| Verified by asking a model | Circular | Verify only in a source you can open |
What not to automate
- The decision that a citation is good. That check has your name on it.
- Filing or sending anything unreviewed. Courts hold the signer responsible, as the order above shows.
- Case strategy and advice to clients. These need judgement and context no prompt supplies.
- Confidential material in unapproved tools. Policy decides, not convenience.
- Replies to a court or opponent who questions a citation. Verify first, then respond.
If your firm is drafting an internal rule, the AI use policy template is a starting point, and the client disclosure wording post covers what to say to clients. Whether disclosure to courts is required depends on your court; look for current standing orders and practice directions rather than assuming. I did not verify any Indian court or Bar Council position for this post, so if you practise in India, check the current guidance from your High Court, the Supreme Court and the Bar Council directly. More general templates live in the prompt library.
