AI Learning Hub › Learn AI › Self-Paced Courses › Introduction to AI › Intermediate › Verification and Source Checking

Level 3 of 4 · Intermediate · Lesson 4 of 6

Verification and Source Checking

Checking claims, quotations, citations, legal authorities, calculations, and source traceability in AI output.

About 15 minutes

By the end of this lesson you should be able to

Run a claim-to-source check and record each claim as supported, qualified, contradicted, or unresolved
Distinguish a citation’s format, the source’s existence, its support for a proposition, and its current authority
Explain, at an introductory level, what a citator check adds to reading a legal source, and what it does not replace

Why this matters

Lesson 3 ended with provisional claims and a list of sources. This lesson asks: what must be checked before this output can be relied on? The answer is a small, repeatable routine that works the same way for a summary of fellowship rules, a draft announcement, or a research memo. You are checking whether each claim is supported by a source you can read. You are not trying to work out where the assistant’s wording came from, which the output cannot tell you.

The claim-to-source check

1. Identify the claim

One checkable statement at a time. Split a sentence that makes two claims into two.

2. Locate the source and the passage

Open the actual source, not the assistant’s description of it, and find the passage the claim rests on.

3. Read the context

Read the surrounding paragraph and headings. Conditions, exceptions, and definitions often sit a sentence away.

4. Compare wording and scope

Does the source say this, as strongly, about the same situation? Is a quotation exact?

5. Check the date and version

Is this the version that applies to your question, and is it still current?

6. Record the result

Use one of four labels, with a note.

Supported: The source says this, with the same scope.
Qualified: The source supports it only with a condition, exception, or narrower scope that the claim leaves out.
Contradicted: The source says otherwise.
Unresolved: No source located, the sources you have do not address it, or they conflict in a way you cannot yet settle. Note what further investigation is needed.

Some claims are inferences: a calculation or conclusion drawn from what a source says. Label them as inferences, check each premise against the source, and check the reasoning. Do not record an inference as if the source stated it. And a real source’s statement is evidence, not a guarantee: it can be outdated, mistaken, or not apply to your situation.

Four different questions about a citation

Format: Is it styled correctly, for example in Bluebook form? A perfectly formatted citation can point to nothing.
Existence: Does the source exist, and does the citation lead to it?
Support: Does the cited passage support the proposition it is cited for?
Current authority: For legal sources: is it still good law, and does it bind the court or body you care about?

Passing one check says nothing about the others. If an AI tool fixes a citation’s format, that does not show the citation exists, supports the proposition, or is current authority. Establish each of those separately.

Checking legal sources: a first look

Legal sources need four more questions. This is an introduction; your legal research courses and the library go further.

Jurisdiction. Whose law is it, and from which court? Whether a decision binds a particular court depends on the jurisdiction, the court structure, and the decision’s precedential status; elsewhere it may be persuasive at most.
Procedural context. What was the court deciding: a motion to dismiss, a summary judgment, an appeal after trial? A statement can depend on that posture.
Holding versus other statements. The holding is what the court decided in order to resolve the case. Descriptions of the parties’ arguments, passing remarks, and dissents are not the holding.
Subsequent treatment. Has the decision since been reversed, overruled, questioned, or limited?

That last question is what citators are for. Two citator services are KeyCite, in Westlaw, and Shepard’s, in Lexis. Both require a licensed subscription. The Robert Crown Law Library’s Legal Databases page (opens in a new tab) lists them and says access is limited to the current Stanford Law School community, with individual usernames and passwords; it directs questions to Reference. Thomson Reuters describes KeyCite as a way to “verify whether your case is still good law and to find citing references.” (Thomson Reuters, “Access KeyCite” (opens in a new tab); product help.) LexisNexis explains that Shepard’s signals reflect overall precedential status, are based on editorial treatment “regardless of the point of law or the rendering court,” and that the most negative treatment controls the signal. (LexisNexis Support Center, “Shepard’s Signals and Analysis” (opens in a new tab); product help.)

A signal is a starting point, not a conclusion. Open the citing references and read what they say, and about which point: a negative signal may concern an issue you are not relying on, and the absence of one does not show that the case supports your proposition. Finding an opinion through an ordinary web search or a free case database lets you read it, but it is not the same as a completed citator check.

The library’s session on assessing output in CoCounsel and Protégé applies the same ideas to AI legal research tools, with a three-step habit: verify, evaluate, test.

Worked example: checking an inference

A constructed example. The seminar and syllabus are fictional.

An assistant summarizing a seminar syllabus writes: “The seminar meets 13 times, so students write 13 response papers.” The syllabus says:

The seminar meets for 13 sessions. Students submit a one-page response paper for any 10 sessions of their choice, due by 9:00 a.m. on the day of the session.

Claim 1, “meets 13 times”: supported.
Claim 2, “13 response papers”: an inference, and its premise is wrong. The syllabus ties papers to “any 10 sessions,” not to every session. Contradicted.

The log entry would read: Claim 2 — syllabus, paragraph 1 — “any 10 sessions of their choice” — contradicted (inference from session count; the syllabus sets 10). The correction is 10 papers. The deadline detail is worth noting too, because a summary that dropped it would be incomplete, not wrong.

Try it: verify a summary

A constructed exercise. The fellowship, offices, and documents are fictional. No legal research knowledge is needed.

The assistant’s summary (constructed):

The 2027 Summer Research Fellowship is open to 1L and 2L students. Fellows receive a $6,000 stipend, which works out to $600 a week. Fellows may also hold another paid position during the summer. Applications close February 15, and faculty sponsors must submit recommendations by March 8. The program has funded more than 200 students since it began.

Source A — Fellowship Guidelines 2027 (issued by the Dean of Students office, January 2027)

§ 2. Eligibility. Applicants must be enrolled 1L or 2L students.

§ 3. Stipend. Fellows receive a stipend of $6,000 for ten weeks of full-time work. Fellows who work part-time receive a prorated stipend.

§ 4. Other employment. Fellows may not hold another paid position during the fellowship period.

Source B — Fellowship FAQ (published by the Dean of Students office, updated February 2027)

Q: When are faculty recommendations due? A: Faculty sponsors submit recommendations through the online portal by March 8. This replaces the March 1 date in earlier guidance.

Source C — Announcement email (from the Dean of Students office, January 2027)

Applications for the 2027 Summer Research Fellowship open January 15 and close February 15. Please direct questions to the fellowship coordinator.

Split the summary into separate claims.
Check each claim against the sources and label it supported, qualified, contradicted, or unresolved.
Which claim is an inference, and what does checking it involve?
Write a corrected summary.

Verification log and discussion

“Open to 1L and 2L students.” Source A, § 2: “enrolled 1L or 2L students.” Supported.
“$6,000 stipend.” Source A, § 3: $6,000 is for “ten weeks of full-time work”; part-time fellows receive a prorated stipend. Qualified: the summary drops the full-time condition.
“$600 a week.” An inference, not a statement in any source: $6,000 ÷ 10 weeks = $600. The arithmetic is right and both inputs come from § 3, but it is an equivalent weekly average for full-time fellows only. The guidelines say nothing about how or when the stipend is paid, so “$600 a week” should not suggest weekly payments. Qualified, and it should be presented as a calculation, not as a figure the guidelines state.
“May also hold another paid position.” Source A, § 4: fellows “may not hold another paid position during the fellowship period.” Contradicted. (The summary also says “summer,” which is broader than the fellowship period the guidelines refer to.)
“Applications close February 15.” Source C. Supported.
“Recommendations by March 8.” Source B, which says March 8 replaces an earlier March 1 date. Supported. Source B is the only document in the packet on this point, from the office that runs the program; the earlier guidance it replaces is not in the packet.
“Funded more than 200 students.” None of the sources addresses the program’s history. Unresolved. Remove it, or find a source.

A corrected summary: “The 2027 Summer Research Fellowship is open to enrolled 1L and 2L students. Full-time fellows receive $6,000 for ten weeks of work (an average of $600 per week); part-time fellows receive a prorated stipend. Fellows may not hold another paid position during the fellowship. Applications close February 15, and faculty sponsors submit recommendations through the online portal by March 8.”

Notice that the contradicted claim and the unresolved claim read just as smoothly as the supported ones. Checking the sources is what reveals the difference; the log records what you checked and found.

A verification log you can reuse

Verification log

CLAIM: [one checkable statement]
SOURCE: [document, version or date, page or section]
PASSAGE: [what the source actually says, quoted or closely paraphrased]
RESULT: Supported / Qualified / Contradicted / Unresolved
NOTE: [the missing condition, the correction, whether it is an inference, or the next step]

For a long document, you will not log every sentence. Log every claim that someone will rely on: dates, amounts, requirements, quotations, and anything cited as authority.

Record what you checked. A log shows a reviewer, or a later you, which claims rest on which passages. A claim you could not resolve is removed or clearly flagged, not softened into vaguer wording.

Sources and further reading

Thomson Reuters: Access KeyCite (opens in a new tab) — official product help for Westlaw Precision. Using KeyCite requires a Westlaw subscription.
LexisNexis Support Center: Shepard’s Signals and Analysis (opens in a new tab) — official product help (Document ID HT5208) on what the signals mean and how they are assigned. Using Shepard’s requires a Lexis subscription.
Robert Crown Law Library: Legal Databases (opens in a new tab) — the library’s list of research databases, with access notes for Westlaw and Lexis.
Assessing Output in CoCounsel and Protégé — the library’s self-paced session on evaluating AI legal research answers.