AI for legal basics
Citation and hallucination risks
A citation-shaped string is not authority; verification begins by retrieving the source and ends by checking what it actually supports.
Before you start
Why this matters
You receive a polished memo containing a case name, reporter citation, court, year, quotation, and parenthetical. It looks complete. Without searching yet, list the independent questions required to rely on it. Does the case exist? Is the citation correct? Does the quotation appear in the opinion? Is the quoted speaker the court or a party? Is the holding relevant? Is the authority binding, current, and still valid? Which question does formatting answer? None of them.
1Learn the idea
This lesson is educational and is not legal advice. Citation practices, duties of candor, filing requirements, and source authority vary by forum and jurisdiction. Qualified reviewers must verify work used in real legal matters.
Read
Why legal hallucinations are dangerous
Generative models produce likely language. Legal citations are highly patterned, so a model can produce names and numbers that look authentic even when no source exists. It may also combine attributes from several real decisions, attach a genuine citation to the wrong proposition, invent quotation marks around a paraphrase, or omit language that narrows the rule.
Errors are not limited to cases. A model may invent statutory subsections, regulations, agency guidance, docket events, contractual definitions, quoted evidence, or secondary-source passages. It may confidently claim a source is current without access to a citator or updated database.
Confidence, detail, and repetition do not reduce this risk. Asking the same model “Are you sure?” can generate a stronger explanation of the same mistake. Even retrieval-enabled systems can return the wrong document, use a poor chunk, or synthesize beyond the retrieved text.
Read
Verify every link in the chain
Treat each citation as a chain of claims:
- Existence: the source can be found in an authoritative repository.
- Identity: title, court or issuer, date, citation, and version match.
- Text: the quotation is exact, including omissions and alterations.
- Speaker: the quoted language belongs to the court, majority, regulation, or other claimed authority—not a brief, dissent, or quoted lower source.
- Context: nearby text does not reverse, limit, distinguish, or condition the proposition.
- Proposition: the source supports the exact sentence for which it is cited.
- Authority: jurisdiction, court level, publication status, and source type are appropriate.
- Treatment: later history, amendments, repeal, negative treatment, and effective dates are checked.
- Factual fit: material facts and procedural posture are sufficiently comparable.
- Citation form: the final format meets the relevant style and forum rules.
Complete these checks in trusted primary sources and approved legal research systems. A model-generated validation table is useful only as a checklist; it is not the validation itself.
Read
Use claim-to-source mapping
Break a draft into atomic propositions. One sentence may contain a legal rule, an exception, a factual assertion, and a prediction. Each needs its own support or label.
A claim map can record:
- claim number and exact draft language;
- claim type: fact, law, quotation, interpretation, or prediction;
- cited source and stable identifier;
- exact supporting text and location;
- authority and currency check;
- limitations or contrary material;
- verifier, date, and status.
Mark a claim SUPPORTED, PARTIAL, CONFLICTED, or UNSUPPORTED. Do not let the model convert PARTIAL into a clean rule while rewriting. Remove, qualify, or research unsupported claims before circulation.
Read
Prompt with a closed source set
When using AI for synthesis, supply verified sources and prohibit additions:
Use only the sources provided below. For each proposition, quote the exact
supporting text and identify its source and location. Distinguish holdings,
dicta, party arguments, dissents, facts, and secondary commentary. If support
is missing or partial, mark it UNSUPPORTED or PARTIAL. Do not create, repair,
or complete citations. Do not claim that authority is current or valid.
This reduces, but does not eliminate, error. The reviewer still compares every quotation and proposition with the original. If the model cannot see a full source, disclose that limit.
Read
Check quotations and paraphrases differently
For a quotation, compare character by character, including inserted brackets, ellipses, emphasis, and punctuation. Confirm that omitted material does not change meaning.
For a paraphrase, ask whether the draft preserves the source’s level of certainty, scope, actor, conditions, exceptions, and procedural posture. A model may turn “the court may consider” into “the court must,” or a fact-specific result into a universal rule.
Pinpoint references matter because they let another reviewer reproduce the check. A link to a hundred-page opinion without a page or paragraph is not enough for a consequential proposition.
Read
Design failure to be visible
Require the workflow to stop when a source cannot be retrieved, versions conflict, a quotation is not found, treatment cannot be checked, or the draft relies on an unavailable database. Never replace a missing citation with the model’s best guess.
For court filings, regulatory submissions, formal opinions, and other high-stakes work, use enhanced review: named source verifier, citation report, supervising lawyer approval, and final comparison after formatting. Formatting changes can introduce broken footnotes or mismatched numbering, so the gate belongs after the final document is assembled.