The verification lab: catching AI-fabricated authority before it reaches a filing
The situation
Paralegal graduates will be handed AI-drafted research on their first day. The employable skill is not producing the draft — it is catching what is wrong with it before a supervising attorney signs. This workflow makes verification the graded deliverable rather than an afterthought.
Steps
-
Give students a real research question
Use a California-specific procedural question where the answer depends on local rules. General chatbots fail these reliably, which makes the failure legible rather than theoretical.
What you only learn by doing it: Pick a question you already know the answer to. You need to be able to see what the tool got wrong faster than the students do.
-
Generate the research memo
Spellbook Westlaw Advantage with CoCounsel Lexis+ AI with Protégé
Have students produce a memo with full citations. If you can run the same question through both a purpose-built legal tool and a general chatbot, do — the comparison is the lesson.
What you only learn by doing it: Do not tell students in advance that some citations will be fabricated. The discovery is the point.
-
Pull every authority by hand
CEB AccessLaw Clio Duo and vLex Vincent AI
Every case, every statute, every pin cite. Students document for each one: does it exist, does it say what the memo claims, is it still good law. This is the deliverable.
What you only learn by doing it: Require a table with a row per citation and a column for 'verified how.' Students who skip a row always skip the one that was fabricated.
-
Write the reconciliation
Students write a short memo to the supervising attorney: what the AI produced, what verification found, what they changed and why. This is the actual work product a firm wants.
What you only learn by doing it: Grade this, not the research memo. The research memo is the raw material; the reconciliation is the professional skill.
-
Read the sanctions record
Close with the New Mexico case. In September 2026 the state Supreme Court held attorney Stephen Aarons in contempt over an appellate brief in a murder appeal that cited four witnesses who do not exist and attributed false testimony to three real people. $5,000 sanction, barred pending discipline, briefs struck, the defendant's appeal reassigned. He said he expected a bulletproof summary.
Most sanctions cases involve fabricated citations. This one fabricated facts in the record of a criminal appeal, and a client bore the consequences.
What you only learn by doing it: Point out that the public database now holds over 2,000 such cases and roughly 1,175 involve self-represented litigants, not attorneys. The lesson is not that lawyers are careless — it is that unverified AI output is a systemic failure mode.
Where this breaks down
Be precise with students about the numbers. Stanford's RegLab study measured Lexis+ AI hallucinating on about 17% of queries and Westlaw's AI-Assisted Research on roughly 33% — and these are the purpose-built, retrieval-grounded, paid legal tools marketed as hallucination-free. A general chatbot was worse, around 43%. There is no correlation between how confident the output sounds and whether it is correct.
The access constraint is real and worth stating to students: most legal AI academic programs are limited to ABA-accredited law schools. Spellbook explicitly includes paralegal students; CEB AccessLaw is free to paralegal programs at participating California institutions. Do not promise students Westlaw or Lexis AI access without written confirmation.
Finally, the professional-responsibility frame. Using these tools does not authorize a paralegal to exercise independent legal judgment, and the supervising attorney remains accountable. California's 2026 State Bar guidance is blunt: no paper filed in court should contain a citation the responsible attorney has not personally read.