- Courses
- AI & the Law
- Activity 4
Verification as a Craft
Why it Matters
Sanctions for fabricated citations have climbed from $5,000 in 2023 to five-figure fines, fee awards, and bar suspensions, and the standard courts now apply is personal verification of every citation, whoever or whatever generated it. That makes cite-checking the paralegal skill that AI made more valuable, because someone has to do it and it has to be done well. This activity gives you a real brief that got its authors sanctioned and has you find every error three different ways, timing each, so you learn what each method catches, what it misses, and how long it actually takes.
Current Context
An April 8, 2026 EDRM analysis counted at least $145,000 in sanctions for AI-fabricated citations in the first quarter of 2026 alone, including about $109,700 in combined penalties and adverse fees against a single Oregon attorney and $30,000 in punitive fines from the Sixth Circuit, and described a shift from embarrassing anomalies to a measurable enforcement trend. The pattern courts are punishing was laid out in Noland v. Land of the Free, California’s first published opinion on the subject, decided September 12, 2025, where 21 of 23 quotations in the appellant’s briefs were fabricated after counsel ran his work through four different chatbots that amplified rather than caught each other’s inventions; the court sanctioned him $10,000, ordered the opinion served on his client, and referred him to the State Bar. The judges in both jurisdictions used nearly the same words: no filing may contain a citation the responsible attorney has not personally read and verified. The brief you audit in Part 2 is the kind of document that language was written about.
Key Concepts
Existence
The first question for every citation: does this authority exist at all, in this reporter, at this volume and page? Fabricated cases usually have plausible names, real court abbreviations, and correct citation formats, which is why format alone proves nothing.
Holding Accuracy
Whether a real case actually held what the brief says it held. A citation can be genuine and still be wrong because the case decided a different question, went the other way, or was cited for dicta as though it were the rule.
Pinpoint Support
Whether the specific page cited contains language that supports the specific proposition. Verification means reading the cited page, not the syllabus, and confirming that the words there say what the brief claims.
Quotation Fidelity
Whether every quoted phrase appears in the source exactly as quoted. AI tools invent quotations inside real cases as readily as they invent cases, and a false quotation in a real case is harder to catch than a fake case.
Subsequent History
Whether the authority is still good law. A citator such as KeyCite or Shepard’s shows whether a case has been reversed, overruled, or criticized, and a real case that has been overruled is almost as dangerous as a fabricated one.
Verification Cost
The time it takes to check a citation properly, measured in minutes per cite. Knowing the cost is what lets a paralegal tell a supervising attorney, honestly, how long a cite check will take and why it cannot be skipped.
Resources
- Woodward Harbor, L.L.C. v. City of Mandeville, No. 2:23-cv-05824, Doc. 97, Joint Opposition to Second Motion to Dismiss (E.D. La. June 3, 2025) (The brief you audit; filed by licensed attorneys and followed by sanctions.)
- Woodward Harbor, L.L.C. v. City of Mandeville, docket, CourtListener (The docket, where you can find the court’s response to the brief.)
- Westlaw via the Daytona State College Library (Full-text case law and the KeyCite citator for subsequent history.)
- Google Scholar (Free case law search; select Case law and search by party name and citation.)
- Russell v. Mells, No. 2D2024-1560 (Fla. 2d DCA Dec. 10, 2025) (A Florida appellate court addressing fabricated citations; read it for the Florida standard.)
- Noland v. Land of the Free, L.P. (Cal. Ct. App. Sept. 12, 2025) (The published opinion on fabricated quotations that sets out the personal verification duty.)
- California Appellate Court Issues $10K Sanctions in State’s First Published Opinion on AI-Hallucinated Case Citations, McGuireWoods (A short summary of Noland and what it requires.)
- Whiting v. City of Athens, No. 25-5424 (6th Cir. Mar. 13, 2026) (The federal appellate standard and the sanctions that came with it.)
- The AI Sanction Wave: $145K in Q1 Penalties, EDRM (The 2026 tally of sanctions and the cases behind it.)
- AI Hallucination Cases database, Damien Charlotin (Search here for any new decision in your own state.)
- Ethics of Artificial Intelligence for Lawyers, Arkansas Law Notes (2026) (A short scholarly overview of the duties verification serves.)
- GenAI Hallucinations, Thomson Reuters (A vendor’s own account of why legal AI tools hallucinate and how citation checking features work.)
What to Do
In this activity you verify every citation in a real sanctioned brief three ways, time each pass, classify what you found, and build the verification workflow you would use in practice. You post the log, a comparison of the methods, a corrected paragraph, and your workflow.
Part 1: Read the Brief and Set the Baseline
Read the Woodward Harbor brief once without checking anything, and list every citation it contains with the proposition each one supports. Then read Noland’s discussion of what verification requires, so you know the standard you are about to apply. Start a verification log with one row per citation and columns for each method, the result, and the minutes spent.
Part 2: Verify Three Ways
First, verify by hand. For each citation, use Google Scholar or Westlaw to confirm the case exists, read the cited page, confirm the holding and any quotation, and check subsequent history in KeyCite. Record what you found and how long it took. Second, upload the brief to a general chatbot and use the Citation Screener to have it flag problems, then record what it flagged, what it missed, and whether it confidently validated anything you had already proven fake. Third, run the citations through a legal research tool with a citation checking feature, such as the AI tools in Westlaw, and record the same. Keep the timing honest; the point is to learn the real cost of each method.
Part 3: Classify and Rewrite
Use the Error Taxonomist to sort every problem you found into a category, and correct its sorting where your own reading disagrees. Then pick one paragraph of the brief that rests on bad authority and rewrite it with real, verified support found through Scholar or Westlaw, and run your rewrite through the Rewrite Checker with the text of the cases you relied on. Finally, use the Verification Workflow Designer to turn what you learned into a step-by-step process you would actually follow before handing research to a supervising attorney, with the human checkpoints marked.
Part 4: Post Your Deliverables
Create a single new discussion thread containing your verification log, a comparison of about 400 words on what each method caught and missed and how long each took, your rewritten paragraph with its verified citations, and your personal verification workflow of 250 to 400 words. Close with a paragraph on how the comparison changed your understanding of what a human in the loop has to do.
Part 5: Share Chat Link
Include one AI chat link with a 1–2 sentence explanation of what the conversation shows and why you chose to share it.
Suggested AI Prompts
Use these prompts as a starting point, then adjust them to fit your goal. Strong prompting develops through trial, revision, and testing. It’s a foundational skill that grows into more advanced AI work such as context engineering and agent-based workflows.
Citation Screener
Act as a cite-checking assistant. I will upload or paste a legal brief. List every citation to a case, statute, or rule in it, and for each one tell me whether you can confirm it exists, whether the proposition it is cited for is one the authority plausibly supports, and whether any quotation attributed to it raises a concern. Label each citation as confirmed, suspicious, or unable to verify, and explain your reasoning in one sentence. You do not have access to a case law database, so do not claim to have confirmed a case unless you are certain; unable to verify is an acceptable answer and a better one than a guess.
A screen, not a verdict. The prompt tells the model what it cannot do and gives it permission to say so, which is the only way to keep it from validating a fake case with the same confidence it uses for a real one.
Error Taxonomist
Act as a legal research supervisor. I will paste my verification log for a sanctioned brief, listing each citation and what I found when I checked it. Classify every problem as a fabrication (the authority does not exist), a holding misstatement (a real case described wrongly), a pinpoint failure (the cited page does not support the proposition), a false quotation, a subsequent history problem (overruled, reversed, or criticized), or a citation form error. For each, explain in one sentence why it belongs in that category, and flag any entry where my own finding is ambiguous enough that you cannot classify it. Wait for me to paste the log.
Classification on your findings, not the model’s. Because you supply what each check found, the model is organizing evidence rather than generating it, and the categories give you a vocabulary for the comparison in Part 4.
Rewrite Checker
Act as a strict appellate editor. I will paste a rewritten paragraph from a brief along with the full text of each case I cited in it. For every citation, quote the exact language from the pasted opinion that supports the proposition and give the page it appears on, or state that the opinion does not support the proposition. Confirm that every quotation in my paragraph appears verbatim in the pasted text. Use only the opinions I provide; do not add authority or rely on your memory of the law.
Verification confined to supplied text. Pasting the opinions turns the check into a reading task the model does well, and forbidding outside authority prevents it from repairing your paragraph with a citation you have not read.
Verification Workflow Designer
Act as a litigation support manager who trains new paralegals. I have just verified a sanctioned brief three ways and I will describe what each method caught, what it missed, and how long each took. Help me design a step-by-step verification workflow I would follow before handing legal research to a supervising attorney: what gets checked, in what order, with which tool, how long each step should take, and where a human must personally read the source before the work moves on. Make it specific to the tools I actually have, Google Scholar and Westlaw, and challenge any step I propose that skips personal reading of a cited page.
Turning results into a process. The prompt asks the model to build the workflow from your measured experience rather than from generic advice, and the instruction to challenge shortcuts keeps the personal verification standard at the center of it.