Legal AI — Module 21.1

How This Lesson Fits the Module & Volume

Healthcare and finance established regulated product patterns. Legal AI is the third: citation-locked research, clause compare, and drafting for licensed counsel—not a chatbot that “is your lawyer.” Research assistants + Vol. 19 hallucination tests are the technical core; invented case citations are infamous failure modes. Vol. 20 copyright, privacy, and transparency constrain corpora, privilege, and client disclosure.

Educational engineering, not legal advice, not the practice of law. Students learn allowlists, matter ACLs, jurisdiction pins, attorney HITL, and audit—then the volume capstone Education AI asks a different integrity question: tutoring vs assessment.

Learning Objectives

By the end of this lesson, students should be able to:

State that this lecture is not legal advice and does not authorize practicing law.
Scope legal AI to counsel-assist tasks: research, compare, draft, summarize—with attorney review.
Require jurisdiction, matter privilege ACLs, and citation support (no dangling reporters).
Reuse Vol. 19 hallucination/attribution tests as release gates.
Apply Vol. 20 privacy, copyright, and transparency to client data and work product.
Know when unauthorized-practice and ethics questions must go to qualified counsel—not the model.

Definition

Legal AI (here) means AI features used by legal professionals or legal ops on matter work product: research memos, contract redlines, discovery triage, and client-facing drafts that a human lawyer still owns. The regulated product pattern is: matter-scoped RAG + citation lock + attorney (or supervised paralegal) HITL + privilege-aware ACLs + audit + no claim that the product is a lawyer or gives legal advice to the public. Privilege / confidentiality is a privacy-and-ethics overlay: retrieval must not leak Matter A into Matter B (Vol. 20 privacy cousin of tenant isolation).

Not Legal Advice

Nothing here is legal advice, an ethics opinion, or authorization to practice law. Court-citation examples are synthetic. Unauthorized-practice rules, privilege, and advertising rules vary by jurisdiction—involve licensed counsel. Do not ship “AI attorney: guaranteed to win” from this lecture.

In-Scope Counsel Assist vs Out-of-Scope Advice Bot

Usually in-scope (attorney HITL)	Usually out-of-scope as the product claim
Case/statute research over a licensed corpus with pinpoint cites	Public chatbot answering “what should I file?” as advice
Clause compare / playbook deviation flags	Binding negotiation conducted solely by the model
First-draft memo or letter for attorney edit	“This email is your legal advice; send it”
Discovery ranking for human reviewers	Unsupervised privilege calls that produce to opposing counsel
Plain-language explainer of attorney-approved text	Invented case law or fake quotations

Unauthorized practice of law is a product and go-to-market issue, not a prompt suffix. If the user is not a licensed professional (or their supervised staff) inside a matter, the safer default is refuse or redirect—same spirit as healthcare’s “not a diagnosis.”

Citation Lock, Privilege ACL, HITL

Citation lock (Vol. 19)

Only cite retrieved, licensed sources
Fail dangling reporter strings
Quote spans must support the claim
Pin jurisdiction + “as of” date

Privilege / matter ACL

Index shards per matter_id
Canary docs across matters (Vol. 20 privacy)
Vendors: no-train, DPA, residency as counsel requires
Work-product labels on logs

Attorney HITL

Edit + sign before client send
Escalate novel issues / conflicts
Disclose AI use when rules/transparency require
Sampled review of “routine” NDAs

Counsel-assist design buys

Speed on rote compare/research
Auditable drafts (governance evidence)
Honest marketing: tool for lawyers, not a lawyer

Advice-oracle costs

Hallucinated cases (famous real-world failures)
Cross-matter privilege leaks
UPL / ethics / malpractice exposure for the firm, not “the model”

Matter-Scoped Draft Gate (Not the Practice of Law)

Illustrative engineering hooks. Not an ethics opinion. Not legal advice. Synthetic citation tokens only.

# Citation-locked research / clause compare — not the practice of law.
# Educational engineering. Not legal advice.

import re

ALLOWED_TASKS = {"research_memo_draft", "clause_compare", "discovery_rank", "plain_language_approved"}
ADVICE_MARKERS = ("you should sue", "you will win", "file this tomorrow", "this is legal advice")

def privilege_acl(user: dict, chunk_meta: dict) -> bool:
    return (
        user.get("matter_id") == chunk_meta.get("matter_id")
        and user.get("role") in {"attorney", "paralegal", "client_authorized"}
    )

def legal_draft_gate(query: dict, retrieved_meta: list, draft: str, reviewer_id=None) -> dict:
    if query.get("task") not in ALLOWED_TASKS:
        return {"ok": False, "reason": "out_of_policy_task"}
    if not query.get("jurisdiction"):
        return {"ok": False, "reason": "missing_jurisdiction"}
    if any(m in draft.lower() for m in ADVICE_MARKERS):
        return {"ok": False, "reason": "sounds_like_advice_to_a_client"}
    allowed_ids = {m["id"] for m in retrieved_meta}
    cite_ids = re.findall(r"cite:([A-Za-z0-9_-]+)", draft)
    dangling = [c for c in cite_ids if c not in allowed_ids]
    if dangling:
        return {"ok": False, "reason": "dangling_citation", "dangling": dangling}
    if not reviewer_id:
        return {"ok": False, "reason": "pending_attorney_review"}
    return {
        "ok": True,
        "disclaimer": (
            "Attorney work-product assist only. Not legal advice to a client. "
            "Licensed counsel must review before use."
        ),
        "privilege_tag": query.get("matter_id"),
    }

# Vol. 19 gate: attribution + contradiction on memo claims vs retrieved opinions.
# Vol. 20: canary chunk from Matter A must never appear in Matter B answers.

Related Lectures

Lecture	Role
Finance AI	Previous regulated sibling
Research assistants	Citation + RAG product core
Document AI	Contract extraction sibling
Hallucination tests	Fake citations fail the build
Copyright / Privacy	Corpus license + privilege/PII
Transparency / Compliance	Disclose AI use; high-level duties
HITL	Attorney sign-off
Education AI	Next: volume capstone

Common Misconception

“If the model cites a case name, it exists.” Attribution tests exist because models invent reporters. Second: a consumer “AI lawyer” is just marketing around the same stack. Third: matter ACLs are optional because the LLM “will only use relevant docs.” Fourth: copyright-clean training data is automatic if you RAG public PDFs. Fifth: this lecture is legal advice or an ethics ruling. Sixth: paralegals can unsupervised-send model output to clients without attorney review (product default: no).

Knowledge Check

Short Answer: Is this lecture legal advice? Answer: No—educational engineering only; not the practice of law.
True/False: Invented case citations should fail the product gate. Answer: True.
Multiple Choice: Privilege ACL means: (a) Matter A chunks stay out of Matter B answers, (b) higher temperature, (c) skip HITL. Answer: (a).
Short Answer: Name two in-scope legal AI tasks under attorney HITL. Answer: Any two of: research memo draft, clause compare, discovery rank, plain-language of approved text.
True/False: A public chatbot claiming to be your lawyer is an in-scope product claim here. Answer: False.
Multiple Choice: Missing jurisdiction on a research query should: (a) block or require pin, (b) invent a circuit, (c) raise BLEU. Answer: (a).
Short Answer: Which Vol. 19 lecture supplies citation/faithfulness tests? Answer: Hallucination tests.
True/False: Cross-matter canary leakage is a privacy/privilege eval failure. Answer: True.
Multiple Choice: Vol. 20 copyright mainly constrains: (a) corpus/output licensing, (b) GPU clocks, (c) Cosine thresholds only. Answer: (a).
Short Answer: Why keep attorney HITL before client send? Answer: The human lawyer owns advice, ethics, and accuracy; the model is assist only.

Key Takeaways

Legal AI here is counsel-assist: citation lock, matter ACLs, attorney HITL—not a lawyer-in-a-box.
Fake citations and privilege leaks are incidents (Vol. 19 + Vol. 20).
Unauthorized-practice and ethics questions escalate to licensed counsel.
This page is not legal advice.
Next: Education AI — Vol. 21 capstone, then Vol. 22 ecosystem.

Trainer’s Guide

Lab: Toy corpus of 8 synthetic “opinions” across two matters. Students implement dangling-cite failure + cross-matter canary. No real client files. No advice to classmates about real legal problems.

Whiteboard: Research assistant \(\to\) legal vertical overlays (jurisdiction, privilege, UPL). Contrast education next: the integrity risk is cheating/assessment, not privilege.

Recap: Legal AI products assist licensed humans with locked citations and matter ACLs—they do not practice law. Continue to the volume capstone Education AI.