SkarpChapter 2 of 3
From Plausible Answer to Verified Legal Research
Follow a fictional question through a verification workflow: identify jurisdiction and date, turn claims into research leads, locate primary law, check quotations and citations, and mark uncertainty. See how fluent prediction can produce invented cases or outdated rules.
1. Treat a Fluent Answer as a Research Map
Plausible is not proven
Generative AI predicts likely-looking text. It can generate useful research leads, but it does not establish what the law is.
NIST describes confidently presented false output as confabulation. In law, that may mean an invented case, a misused real case, or an outdated rule.
Treat an AI answer like sticky notes on a map: clues for where to look, not proof until you inspect the underlying authority.
2. Freeze the Question Before Searching
Build a research frame
Before entering search terms, write a one-sentence research frame:
`California; consumer contract; online gym membership; signed April 3, 2026; cancellation rights; question asked September 17, 2026.`
This prevents a common error: finding a correct rule from the wrong place or time. "California law" may include a statute, an agency regulation, a court decision interpreting either one, and sometimes a local ordinance. A rule effective after April 3, 2026 may be informative, but it may not govern an agreement made on that date.
Try it: Circle the fact most likely to change the answer: where the consumer lived, where the business operated, how the agreement was delivered, or the agreement date. The best initial choice is usually the agreement date, because legal versions and effective dates matter. Then flag the others as facts to confirm.
3. Convert Assertions into Checkable Leads
Split the answer
Turn broad conclusions into separate claims: coverage, deadline, and method of notice. Each needs its own evidence.
Ask source-ready questions: Which statute? Which version applied on April 3, 2026? What conditions, exclusions, and notice requirements appear in its text?
Use a claim ledger: `unverified`, `verified`, `contradicted`, or `uncertain`. It makes gaps visible rather than rhetorical.
4. Climb the Authority Ladder
Primary authority
Use summaries to find vocabulary, then move to the law itself: statute, regulation, binding judicial decision, or applicable court rule.
Read the full operative text. Definitions, exceptions, and cross-references often change the apparent rule.
For a cited case, retrieve the decision and read its holding in context. A case name alone is not verification.
5. Verify Time, Version, and Scope
Law has a timeline
Record the enactment date, effective date, and event date. The current text may not be the text that governed a past transaction.
Check whether an amendment is prospective, whether a section was renumbered or repealed, and whether the statute limits its scope.
A quotation can be accurate today yet legally irrelevant to an agreement made earlier.
6. Audit Every Quote and Citation
Three checks
Test every citation for existence, accuracy, and relevance. A real authority may still be quoted wrongly or apply to another issue.
Compare a quotation with the primary source, including subsection, version date, omissions, and surrounding conditions.
NIST recommends verifying AI-generated sources and citations; ABA Formal Opinion 512 warns against uncritical reliance on generated legal content.
7. State What You Know and What Remains Open
Finish with calibrated uncertainty
A verified research note might say: "The cited authority does not yet establish a cancellation right because its covered-agreement definition has not been matched to the contract. The notice-method claim remains uncertain pending review of the operative text and any controlling cases."
That is stronger than a confident guess. Mark missing facts, unresolved cross-references, and sources you could not access. Preserve a research trail: source title, official location, version date, pinpoint section or page, search date, and result.
For lawyers, verification also connects to professional duties. California's practical guidance warns that AI may generate plausible but inaccurate legal citations and emphasizes that the lawyer retains control of strategic decisions and professional judgment. Court-specific procedures may add further obligations: the U.S. Court of International Trade directs attorneys to consider chambers procedures and Rule 11 responsibilities.
Reflection: What changed your understanding: the danger of an invented citation, the importance of a date, or the difference between a lead and proof?
Check Understanding
Choose the best next move when an AI answer gives a confident quotation and statute number.
What should a researcher do first with the quoted statute?
- Open the official or authenticated text, confirm the exact subsection and version in force on the relevant date, then read its scope and exceptions.
- Repeat the same prompt until the AI gives the quotation consistently.
- Use the quotation if it sounds consistent with a search-result summary.
Show Answer
Answer: A) Open the official or authenticated text, confirm the exact subsection and version in force on the relevant date, then read its scope and exceptions.
Verification begins with primary authority and the legally relevant version. Repeated AI output and search snippets may repeat the same error or omit conditions.
Key Terms
- confabulation
- Confidently generated but erroneous or false content; it is also commonly called a hallucination or fabrication.
- effective date
- The date a legal provision becomes operative or legally applicable.
- secondary source
- Material that explains, summarizes, or helps locate law, such as a treatise, article, blog, search result, or AI response.
- pinpoint citation
- A citation to the precise section, subsection, page, or paragraph supporting a claim.
- primary authority
- The actual legal source that establishes or authoritatively interprets law, such as a statute, regulation, binding judicial decision, or court rule.