Build an authority table first, from opinions you have read, and give the assistant only those authorities when it drafts from your outline. Then cite-check every authority against the primary source before filing: the decision exists, the quotation is exact, the pin cite is right, the proposition holds, and the decision is still good law. Nothing goes in unverified.
Fabricated citations are the failure courts punish. A public database of court decisions involving AI-hallucinated content, last updated on 2 November 2026, lists 2,262 cases, 897 of them involving lawyers. This routine makes invention structurally hard: the assistant sees only verified authorities, and a person checks every cite against the source.
Keep the drafter and the verifier apart
Three roles run through every filing. The drafter is the lawyer working with the assistant. The verifier is a different person, an associate or paralegal where the firm has one, who checks each authority and record cite against the source. The signer is the lawyer whose name is on the brief and who answers to the court. In a one-lawyer firm the same person may fill two roles, but not in the same sitting: verify the next morning, with the authority table open and the draft closed.
The split exists because an assistant cannot check its own work, and neither can a person who watched it being written. Both are primed to see what they expect. It also changes the expectation about time. Verification usually takes longer than drafting, and the worked example below shows by how much. A link inside a legal research platform is useful, but a click-through to a real decision is the first check, not the last.
Step 1: an outline that carries the argument
The assistant writes what the outline tells it to write, so the outline needs to be tight. Use full-sentence point headings that state the proposition, and under each one list the rule, the authority numbers from the table, and the record cites that support the facts. Add the standard of review and the relief sought at the top. Here is the shape for an illustrative opposition to a motion to dismiss, in a supply dispute between a wholesale distributor and its supplier:
- Standard. The complaint need only allege facts that make the claim plausible. Authorities A1 and A2.
- Point I. The supplier's delivery logs show 41 late shipments in six months, which plausibly alleges breach. Authorities A3 and A4. Record R12 to R19.
- Point II. The notice clause was satisfied by the emails of 3 March and 9 April. Authority A5. Record R21 and R22.
- Point III. The damages claim does not depend on a consequential-loss theory. Statute S1. Authority A6.
- Relief. Deny the motion; in the alternative, allow leave to amend.
Notice what the outline leaves out: every authority is a number, never a name from memory. If a point has no number beside it, that is the gap to fill by research, not by asking the assistant what the law is.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowStep 2: the authority table is the control point
One row per authority, built by a person who has opened the opinion. Seven columns do the work: a number, the full citation, the court's level (binding or persuasive), the proposition it supports in one sentence of your own, the exact quote copied from the opinion's text rather than retyped, the pin cite, and a verified column holding initials and a date. Only rows with initials go to the assistant. Where the rows come from matters less than how they are filled: a research platform, a free public case-law database or the court's own site all work, so long as the quote is a paste and the pin cite was read off the page where the passage sits.
| No. | Authority | Level | Proposition | Quote and pin | Verified |
|---|---|---|---|---|---|
| A1 | [Appellate decision on the pleading standard, full citation] | Binding | A complaint must plead facts, not labels | "[sentence pasted from the opinion]", p. 12 | Initials, 4 Nov |
| A2 | [Same court, later decision applying A1 to a contract claim] | Binding | Repeated late delivery can state a breach claim | "[sentence pasted from the opinion]", p. 8 | Initials, 4 Nov |
| A3 | [Decision from another jurisdiction on notice clauses] | Persuasive | Email notice can satisfy a written-notice clause where no method is prescribed | "[sentence pasted from the opinion]", p. 5 | Pending: not sent to the assistant |
The brackets stand in for the real citation and quotation, which you copy from the opinion. An invented case name in a sample would defeat the point of the table. The "Pending" row shows the rule in action: A3 is on the list because it looks useful, but until a person has read it, the assistant cannot see it, and so cannot cite it.
Step 3: drafting from the table, not from memory
Draft one section at a time, about 600 words each, with the outline, the authority table and the record list in the prompt. A whole brief in one run invites the assistant to improvise connective tissue, and improvisation is where invented authority enters.
You are drafting one section of a brief from the outline, authority table and record list below.
Rules:
1. Cite only authorities in the table, by table number in square brackets, for example [A3]. Never cite anything else. Never invent a case, statute, rule or quotation.
2. Quote only text from the "quote" column, word for word, inside quotation marks. Otherwise paraphrase, and paraphrase only what the "proposition" column says the authority holds.
3. Every statement of fact needs a record cite from the record list, for example [R12]. If a fact has no record cite, write [NEEDS RECORD CITE].
4. If a point has no supporting authority in the table, write [NEEDS AUTHORITY] and move on. Do not fill the gap from memory.
5. Do not call an authority controlling, binding or on all fours unless the table says so.
6. Do not use "id." or "supra". Use table numbers only.
7. About 600 words, formal register, point heading first.
Outline for this section: [paste]
Authority table (verified rows only): [paste]
Record list: [paste]A typical first output looks like the paragraph below (illustrative). Read it with the table open, because it contains four defects a quick skim would miss:
A complaint need not plead evidence; it must allege facts that make the claim plausible. [A1] The court explained that "labels and conclusions" will not do, id., but that detailed factual allegations are unnecessary. [A2] Here, the supplier's delivery logs record 41 late shipments in six months. [R12] Courts routinely deny dismissal in comparable supply disputes. [NEEDS AUTHORITY]
- The quotation "labels and conclusions" is not in the quote column for A1. The assistant supplied it from training memory. Search the opinion for the string: if it is there, add it to the table; if not, delete the quotation marks.
- "id." slipped in despite rule 6. Once a sentence moves, an id. chain points at the wrong authority, so short forms are added at the very end, after the structure is final.
- A2 now carries a proposition the table never gave it (detailed allegations are unnecessary). Cut it back to the wording in the proposition column, or research whether A2 really says so and update the table.
- "Courts routinely deny dismissal" is the assistant doing what rule 4 asked, which is flagging a gap. The drafter either finds a verified authority and adds row A7, or deletes the sentence.
After the structure is stable, replace each bracket token with the full citation from the table in one find-and-replace pass, and only for rows whose verified column is filled. That replacement is the last mechanical step before the cite-check proper.
Step 4: the cite-check, nine checks per authority
The verifier works through the table and the draft side by side, one authority at a time, with the opinion itself open. A citation passes only when every check that applies to it passes.
| Check | What the verifier does | The failure to expect |
|---|---|---|
| 1. Exists | Finds the decision by name and citation on a primary source: the court's site or a database you trust | A fabricated case, or a real name attached to the wrong court |
| 2. Citation form | Matches volume, reporter, first page, court and year to the decision | A plausible volume number with the wrong year |
| 3. Quotation | Searches the opinion's text for the quoted string; checks ellipses, brackets and emphasis | A paraphrase presented as a quote, or headnote text quoted as the court's words |
| 4. Pin cite | Confirms the page or paragraph where the passage actually sits | Off by one page, or a star page confused with a reporter page |
| 5. Proposition | Reads the surrounding paragraphs: does the court hold this, merely discuss it, or reject it? | Dicta cited as a holding, a dissent cited as the majority, a losing argument cited as the court's view |
| 6. Later history | Checks treatment in a citator: reversed, overruled, superseded or distinguished | A decision reversed in part, still cited for the reversed part |
| 7. Level and place | Confirms binding or persuasive status for this court | Persuasive authority described as controlling |
| 8. Statutes and rules | Reads the current text and effective date from the official source | A superseded section number, or a quotation taken from a summary |
| 9. Record cites | Checks each fact against the record page cited | A fact that is not in the record, or a page index off by one |
Free tools can do the first filter. CourtListener's citation lookup service checks citations against its database (more than 18 million citations when its documentation was checked on 4 November 2026) and describes itself as a guardrail against hallucinated citations. It can parse every citation in a block of text, up to 250 citations a request and 60 valid citations a minute. A result of 200 means the citation is valid and found, 404 means it is validly formed but not in the database, 400 means the reporter is not recognised, and 300 means more than one match. It does not look up statutes, journal articles, "id." or "supra".
Read the limits carefully. A 200 says only that the citation exists; it says nothing about the quote, the pin cite or the proposition, which are checks 3 to 5. A 404 is a reason to investigate, not proof of fabrication, because every database has gaps; search the court's site and a second database before deciding. The tutorials on stopping AI inventing case law and on a citation-checking workflow for research cover the research side of the same problem in more depth.
Record the result for each authority in a log the signer can read in a minute. A filled-in extract looks like this (illustrative):
| No. | Result of the nine checks | Action |
|---|---|---|
| A1 | Exists, form, quote at p. 12, proposition and later history all pass | None |
| A2 | Exists and form pass; the quoted sentence differs from the opinion by two words | Replace with the exact sentence |
| A5 | 404 on lookup; not on the court's site; not in a second database | Delete, record in the file, find replacement authority |
A 20-page opposition with 31 authorities, from outline to filing
Return to the opposition outlined earlier, with illustrative figures throughout. The finished brief runs 20 pages and cites 31 legal authorities: 22 decisions, 6 statutes or rules and 3 secondary sources, plus 28 record cites in the facts. The authority table held 27 verified rows when drafting began. The assistant's five section drafts produced 6 [NEEDS AUTHORITY] markers and 2 [NEEDS RECORD CITE] markers; four of the authority gaps were closed by researching and verifying four more authorities (rows 28 to 31), two sentences were cut, and the record gaps were closed from the record.
| Stage | Who | Time | Result |
|---|---|---|---|
| Outline | Drafter | 1.5 hours | Five points, authorities by number |
| Authority table | Drafter | 5.0 hours | 27 verified rows |
| Assistant drafts five sections | Drafter | 1.0 hour | 6 authority gaps, 2 record gaps flagged |
| Research to close the gaps | Drafter | 1.0 hour | 4 new verified rows; 2 sentences cut |
| Editing for argument and length | Drafter and partner | 4.0 hours | 20 pages |
| Cite-check of all 31 authorities and 28 record cites | Verifier | 4.5 hours | 7 authorities and 3 record cites corrected |
| Final read, AI-use log, any certificate | Signer | 1.0 hour | Filed |
| Total | 18.0 hours |
The cite-check found problems in 7 of the 31 authorities. Two quotations were paraphrases sitting inside quotation marks. One pin cite was a page out. One proposition was overstated: the court had discussed the point without holding it. One decision had been reversed in part, and the cited passage was in the reversed part. One statute section had been renumbered, and one secondary source carried the wrong edition year. Three of the 28 record cites pointed to the wrong page of the record.
No authority was fabricated, because the assistant never saw a name it had not been given. The seven failures are the ordinary kind that human drafting produces too, and the cite-check exists to catch both sorts. Look at the time column as well: the assistant's drafting took 1.0 hour and the verification 4.5, more than four times as long. A firm that budgets for the first and not the second is the firm that files the bad citation. For facts taken from transcripts, record cites need page and line, and the page-line test for deposition summaries shows how far to trust an assistant's references there.
Court AI rules: disclosure, certificates, records and bans
Rules differ by court and often by judge, and judges keep adding and amending them. Commentary on current orders sorts them into four kinds, and a single court may combine several:
- Disclosure. The filing states that generative AI was used, sometimes under a required heading and sometimes identifying which parts.
- Certification. The signer certifies that a person independently reviewed and verified the AI-assisted portions. One order allows AI-drafted language only if it has been checked for accuracy, using print reporters or traditional legal databases, by a person. Some certificates also ask the signer to confirm that no confidential or proprietary information was disclosed to an unauthorised party through the tool.
- Use and records. The filer identifies the tools used and keeps records of prompts and of the text the tool produced, ready to show the court if asked.
- Prohibition. Some judges bar AI in preparing filings, with narrow exceptions such as AI built into traditional research products. Read the exact wording, because the exceptions decide what is allowed.
Consequences named in these orders include striking the filing, economic sanctions, contempt and dismissal of the lawsuit. At least one order treats hallucinated content as attaching to the signature on the filing, so the signer cannot point to the drafter or the tool.
The routine for any filing is five questions, asked at the start of the matter and again before filing, because orders get amended: Which court rules and which judge's orders apply? Does any rule restrict AI for drafting? Does any require disclosure or a certificate, and who signs it? Does the certificate cover confidentiality? What records must we keep? Then keep an AI-use log in the file:
AI-USE LOG Matter [CODE] Filing [title]
Tool and plan: [for example ChatGPT Business] Data setting: no training by default
Used for: [outline | first draft of sections 2 and 3 | edit for length]
Not used for: [research | citations | record facts]
Verifier: [name] Date: [ ]
Method: every citation checked against the opinion text; quotations searched in the opinion; record cites checked against the record
Court rule checked: [rule or order number] Date checked: [ ]If a certificate is required, use the court's own form. To show the substance only, the wording often looks like this:
I certify that any text in this filing drafted with the assistance of generative AI has been independently reviewed by a lawyer of this firm, and that every citation and quotation has been checked against the primary source.
Mistakes that survive a careful read
Most of the damaging errors are not invented cases. They are real authorities that are cited wrongly, and a lookup tool passes every one of them. Learn their shapes:
- A real case with an invented quotation. The name and citation check out, so the lookup passes. Only a search of the opinion text for the quoted string catches it.
- A real case with the opposite holding. The assistant summarises the losing party's argument as if the court had accepted it. Check 5 exists for this.
- A dissent or concurrence quoted as the majority. The words are real, but they are not the court's. The verifier reads the page header where the passage sits.
- Headnote or syllabus text quoted as the court's words. A publisher's summary is not authority, and it often reads convincingly.
- A string cite padded with plausible extras. Ask an assistant for "more support" and it will oblige. The table rule, which allows only verified rows, is the defence.
- A renumbered rule. The rule exists but under another number, or in a version no longer in force. Read the current text from the official source.
- A record fact that has been smoothed. The record says the supplier's logs show 41 late shipments; the draft says the supplier admitted them. Check each fact against the page cited.
- An authority that changed after the research. A decision can be reversed between drafting and filing, so run check 6 again on the day of filing, not the day of research.
Where to draft, and what to keep out of the prompt
General assistants and legal research platforms can both draft from a table you supply, and both need the cite-check. Among general assistants, ChatGPT Business and Claude Team each cost $20 a seat a month billed annually ($25 monthly), with a two-seat minimum and no training on your content by default. Research platforms with built-in AI add linked citations and their own databases, which helps with the research stage; the tutorial on CoCounsel and Lexis+ with Protégé compares two of them for a small firm.
Ethics guidance on generative AI (Formal Opinion 512, July 2024) adds two points that apply either way. Informed client consent is required before client information goes into a self-learning tool, and general boilerplate in an engagement letter does not count. And output needs an appropriate degree of independent verification, scaled to the tool and the task, because the tool cannot replace the lawyer's judgement. In practice, draft with the argument and the authorities, not the client's confidences. Leave out sealed material, settlement authority, the names of minors and anything else the task does not need, and keep the firm's written policy on AI use close to hand; the AI use policy clauses tutorial is a starting point for drafting one.
Pilot the routine on one filing with no urgent deadline. Run it alongside your normal process, log how long the verification takes, and note what the verifier changed. That record is the best argument for giving cite-checking its own line in the budget, and it is the evidence you will want if a court ever asks how the firm uses AI.
Questions litigators ask about AI drafting and cite-checking
Is a citation-checking tool enough on its own?
No. A lookup tool can confirm that a citation points to a real decision. It cannot tell you whether the quotation is exact, the page is right, the proposition holds or the decision is still good law. Use it as the first filter, then check each authority against the opinion itself before anything is filed.
Do I have to disclose that I used AI in a filing?
It depends on the court and often on the individual judge. Some require disclosure, some a signed certificate that a person verified the content, and a few restrict AI use altogether. Check local rules and the judge's standing orders at the start of the matter and again before filing, and keep a record of how AI was used.
Can AI write the statement of facts?
Only from record material you supply, and every sentence needs a record cite that someone has checked. Assistants smooth over gaps, so they tend to add a plausible detail the record does not contain. Rearranging facts you have already cited is a safer use than asking for a narrative from scratch.
What if a reviewer finds a bad citation after filing?
Treat it as urgent. Tell the supervising lawyer at once, check whether other citations were verified the same way, and consider correcting the filing promptly. Duties of candour to the court vary, so involve your firm's ethics partner rather than deciding alone.
Which assistant should we draft in?
Pick one that keeps your content out of training by default, such as ChatGPT Business or Claude Team, or a legal research platform your firm already licenses. The routine does not depend on the brand, because verification happens against primary sources and not against the assistant.
Further reads
- Spotting AI-Generated Evidence: A Checklist for Litigators — Check exhibits as carefully as citations before they go in.
- Best AI Chatbot for Lawyers: General Models vs CoCounsel — General assistants against legal research platforms for drafting.
- Claude or ChatGPT for a Law Firm: Which One to Pay For — Which general assistant to pay for in a small firm.
- ChatGPT Business for Law Firms: Seats, Privacy Settings and Cost — Seats, privacy settings and cost on ChatGPT Business.
- Claude v ChatGPT for Lawyers: A Same-Task Test on Your Own Documents — Run the same task through two assistants on your own documents.
- Paralegals and AI: How to Divide the Work in Your Firm — Who drafts, who verifies and who signs in a small team.
- Best AI for Solo Lawyers: A Stack That Fits One Person — Assistant, practice-management AI and contract tool for one lawyer: three priced stacks from $120 to $275 a month, and what to leave out.
- Best Legal AI for Criminal Law: A Small Defence Firm's Shortlist — Which AI to use for bodycam transcripts, research, drafting and intake in a small defence practice, with list prices and a vendor data questionnaire.
- AI Discovery Review for Defence Firms: Index, Search and Cross-Check — Index discovery, map each statement to the disputed issues, build a contradiction table with page cites, then verify every cite by hand before filing.
- AI Legal Chronologies for Family Cases: Events Linked to Evidence — One dated table for a family case: extract events from messages, statements and orders, tie each row to a source and page, and keep a gaps list.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: a public database of court decisions involving AI-hallucinated content (updated 2 November 2026), the CourtListener citation-lookup documentation, published commentary on court standing orders on generative AI (April 2026) and a law-firm summary of a confidentiality certificate, Formal Opinion 512 on generative AI tools (July 2024), and the ChatGPT Business and Claude Team pricing pages. Checked 4 November 2026.