Citation quality: does the source actually say that

By the end, you can

• Check whether a citation attached to a claim actually supports that specific claim, by reading the source passage itself rather than trusting the citation's presence (AF-5).
• Identify at least one way a citation can look correct and still fail to support its claim — wrong attribution, an altered quote, a quote truncated to broaden its scope, or a source that does not actually address the claim it is attached to (AF-5).

Before you start

This is Module 5, Lesson 2 of 3. It builds directly on Lesson 1's fact/inference/recommendation/unknown split and its source ladder — this lesson assumes you already know a fact needs an external check, and asks a narrower question: when an answer already comes with a citation or a quoted source attached, is that citation actually doing the job it looks like it's doing?

A citation invites trust — and that trust has to be earned each time

A citation is easy to treat as settled the moment it appears: there's a name, there's a link, the sentence looks backed up. OpenAI's own guidance on building citation systems states plainly why a citation exists at all: "reliable citations build trust and help readers verify the accuracy of responses." Read that carefully: citations exist so a reader can verify a response, not so the reader doesn't have to. A citation that is present but never actually checked is not doing the job the word "reliable" in that sentence is doing all the work of.

Four ways a citation can look right and still fail

A citation does not have to be missing to be worthless. It can be attached, formatted correctly, and pointing at a real, existing source — and still not actually support the sentence it is attached to. Four specific patterns are worth checking for by name:

OpenAI's guidance names the underlying standard directly: "each citation must accurately reflect the source content. Selective interpretation of the source content is not allowed." The second sentence is doing real work — it rules out the truncation and altered-wording patterns above specifically, not just outright fabrication. The Open Worldwide Application Security Project's definition of hallucination describes exactly the same shape of failure, one level up: content that "seems accurate but is fabricated." A citation that fails one of the four ways above is a fabrication wearing the specific costume of a citation — formatted correctly, pointing at something real, and still not actually true to what it claims to support.

• **Wrong attribution.** The quote or fact is real, but it is credited to the wrong person, organisation, document or, inside a longer report with several sections, the wrong section. A true statement attributed to the wrong source is still a citation failure, because a reader who goes to check it will not find it where they were told to look.
• **Altered wording.** The quoted text has been changed from the original — a word swapped, a phrase paraphrased but presented as a direct quote. Even a small change can shift what was actually said.
• **A quote truncated to broaden its scope.** The source did say the quoted words, but the full sentence had a qualifier, a scope limit or a condition that the truncation cut off, and without it the claim now covers more than the source actually supports.
• **A source that is accurate but doesn't actually address the claim.** The citation is real, the quote is word for word correct, and it still does not support the specific sentence it has been attached to — it is adjacent to the topic, or answers a related but different question.

Checking a citation properly

Checking a citation is not reading the sentence it's attached to and deciding it sounds plausible. It is: open the actual source, find the specific passage, and compare it, word for word if the sentence is presented as a direct quote, against exactly what the AI's sentence claims it says. If the source's own wording covers less than the claim, less confidently than the claim, or something adjacent to the claim rather than the claim itself, the citation does not support the sentence, whatever else about it looks correct. This checking step is the same kind of activity NIST's AI Risk Management Framework names as its own distinct function: a "measure function" that assesses and monitors output, separate from the process that generated it. Checking a citation is that measuring step applied to one sentence at a time.

Nothing about attaching a citation changes what the AI system fundamentally is doing, either. The US National Institute of Standards and Technology frames it as "an engineered or machine-based system that can, for a given set of objectives, generate outputs such as predictions, recommendations, or decisions" — a citation is one more part of that generated output, produced by the same mechanism as the sentence it's attached to, not a separate fact-checking pass layered on top. Checking whether it actually earns its place is exactly as external a step as checking the sentence itself.

A worked example: a truncated quote that broadens a claim

An AI-drafted summary of a product review states: "Reviewers found the tool 'easy to use' and reliable for everyday work," with a link to the original review. Checking the original review directly, the actual sentence reads: "Reviewers found the tool 'easy to use for basic tasks,' though several noted it struggled with anything more complex." The link is real, and part of the quoted phrase is accurate word for word — but the summary cut "for basic tasks" and the second half of the sentence entirely, leaving a claim that reads as unqualified praise where the source was actually a mixed, qualified assessment. This is exactly the truncated-quote pattern from the list above: the source did use those exact words, and the resulting claim is still broader than what the source actually supports.

Accessibility notes

This lesson is text-first, with no images, audio, video or downloadable artifacts. The practice exercise's model answer sits behind a native disclosure control that is reachable and operable by keyboard and correctly announced by screen readers. The knowledge check uses native radio-button inputs with a visible question and options, and posts its result to a live status region so assistive technology announces the outcome without a page reload.

Sources and limits

This lesson synthesises the sources below into a practical learning model. It is not a security standard, legal advice or a guarantee that any particular agent design is safe.

1. Citation Formatting — OpenAI. States that reliable citations build trust and help readers verify accuracy, and that each citation must accurately reflect the source content, with selective interpretation not allowed.
2. LLM09:2025 Misinformation — OWASP Gen AI Security Project. Defines hallucination as content that seems accurate but is fabricated — the same shape of failure this lesson describes in a citation that looks correct but does not hold up.
3. Artificial Intelligence Risk Management Framework 1.0 — NIST AI Resource Center. Frames an AI system as an engineered system that generates outputs such as predictions, recommendations or decisions — a citation is one more part of that generated output, not a separate fact-checking pass.
4. AI RMF Core — NIST AI Resource Center. Describes the AI RMF's measure function as a distinct activity from generating output — checking a citation is this lesson's version of that same distinct, external measuring step.