Can you trust an AI teaching assistant's sources?

A citation is a place to start checking, not a reason to stop. Before using an AI teaching assistant's answer, open the source and find the passage that supports the specific claim. Then check whether its date, scope, and conditions fit the teaching task.

This matters when an assistant drafts a lesson explanation, suggests feedback, or answers a question about a rubric. A polished answer can include a real reference and still say something the reference does not support. Northwestern University Libraries' guidance on evaluating AI-generated content explicitly warns that responses citing actual sources can still contain inaccuracies.

This guide is for educators and teams choosing or building teacher-facing assistants. It focuses on checking evidence, not detecting whether a learner used AI. For a broader review before release, start with testing an AI teaching assistant.

Key Takeaways

• A working link proves that a page exists, not that it supports the answer.
• Check individual claims against the original passage, including its limits.
• A credible source can still be the wrong source for a class, course, or decision.
• When evidence is missing or conflicting, keep the answer in review rather than inventing support.

Why a real citation can still support a wrong answer

Source credibility and claim support are separate checks. A university may publish reliable research, but that does not mean every statement attributed to its paper is correct. The assistant might have changed the population, removed a condition, or made the conclusion stronger than the authors did.

The ALCE research paper, published at EMNLP in 2023, evaluates generated answers through separate dimensions of fluency, correctness, and citation quality. That distinction is useful here: an answer that reads well is not necessarily accurate, and an accurate-sounding claim is not necessarily supported by its citation. The paper is research on defined evaluation tasks, not a current error-rate estimate for classroom tools.

Watch for these different failures:

• The reference may be invented. Its title, author, or publication cannot yet be confirmed. Failure to find it is a warning, not proof of fabrication.
• The reference is real but mismatched. The link opens a different paper or document.
• The source is relevant but insufficient. It discusses the topic without supporting the particular conclusion.
• The answer drops a condition. A statement about one age group, setting, or task becomes a claim about everyone.
• The citation covers only part of the sentence. One supported fact gives an unsupported recommendation the appearance of evidence.

These are checking categories, not claims about how frequently each failure occurs. They help you decide what to fix instead of treating every citation problem as the same problem.

Step 1: Separate the answer into checkable claims

Start with the statements that affect what you will teach, recommend, or send. Checking a paragraph as a whole makes it easy to overlook an unsupported clause.

Suppose an assistant says that a method improves understanding, works for every age group, and should replace a current lesson activity. That is more than one claim. Evidence for improved understanding in a particular study would not, by itself, establish the other conclusions.

Copy each important claim into a review note. Beside it, record the source the assistant attaches. If a sentence mixes evidence with a recommendation, separate them. A teacher's proposed use of a finding should be visible as a decision, not disguised as the finding itself.

You do not need a citation for every ordinary connective sentence. Focus on the facts, interpretations, and recommendations that the teaching decision depends on. Ask: if this statement is wrong, would the lesson, feedback, or decision change?

At the end of this step, you should have a short set of claims you can inspect one at a time.

Step 2: Open the original source and find the passage

Read the source itself rather than relying on the assistant's summary or a search-result snippet. Match the title, publisher, authors where relevant, and publication or revision date. For a school document, check its owner and version too.

Search within the source for the subject of the claim. Read the surrounding paragraph, not just the highlighted phrase. Look for words such as “may,” “under these conditions,” or “in this sample.” Those words can change what the evidence allows you to say.

For a paper, check the methods and limitations before applying a finding to a different group. For a rubric, read the whole relevant criterion and any linked instructions. For guidance, check which education system or institution it addresses.

If the link is broken or the paper is behind a paywall, record that you have not verified it. A broken link does not prove a source is fabricated. Search the exact title through the publisher or your library, and use a legitimate accessible copy where available. An abstract may establish the broad result without supporting the detailed claim you need.

Ask the assistant to identify a passage if that helps you locate it. Then verify the passage yourself. Its second answer is still an output to check.

Step 3: Check scope, date, and teaching context

The source must support the claim in the context where you intend to use it. A reliable document can be irrelevant to the learner, course, or decision in front of you.

Check the following before accepting the answer:

Newer is not automatically better. A historical text may be exactly the source a lesson needs. A current assessment instruction, on the other hand, should not be replaced by an older version simply because it is easier to retrieve.

For internal materials, agree which document takes precedence. If two approved documents disagree and no precedence rule resolves them, send the conflict to their owner. Do not ask the assistant to quietly choose whichever one sounds more convincing.

A worked example: the right source, the wrong conclusion

Consider this invented assistant answer, written to illustrate a checking failure. It is not an observed product output or a quotation from the Department for Education:

“Teachers can use AI-generated lesson materials without checking them, provided the materials include citations.”

Imagine it links to England's Department for Education guidance, Generative artificial intelligence (AI) in education.

The link is real. The publisher is relevant. The conclusion is still wrong.

In the section on opportunities and challenges, the guidance says: “Any content produced requires critical judgement to check for appropriateness and accuracy.” The page was updated on 12 August 2025 and applies to England. It does not provide the citation-based exemption in our invented answer.

The review outcome is contradicted, not merely “needs a better link.” A supported paraphrase would be: the guidance expects professional judgement when checking generated material. Including a reference does not remove that checking step.

The example shows why confirming the publisher and opening the link are necessary but incomplete. You must compare the actual claim with the actual passage. It also shows why a policy's jurisdiction belongs in the explanation instead of being silently turned into a worldwide rule.

Step 4: Give unsupported answers a visible next step

Use review outcomes that explain the problem and what happens next. A generic green “sources included” label cannot distinguish missing evidence from a direct contradiction.

A practical review process can use these outcomes:

• Supported: the passage supports the claim as written, within the intended scope.
• Partly supported: a narrower statement is supported; remove or rewrite the unsupported part.
• Contradicted: the source says something incompatible; reject or correct the claim.
• Not verified: you cannot locate or inspect enough evidence; request the source or keep the claim out of use.
• Conflicting: relevant sources disagree; route the issue to the appropriate educator or document owner.

These are proposed workflow labels, not a certification scheme or a guarantee that the whole answer is correct. A supported citation does not prove that no important evidence has been omitted. For consequential recommendations, look for relevant counterevidence and use subject expertise too.

The assistant should state the gap plainly. For example: “I found the previous rubric, but I cannot verify the criterion for the current assignment. Please select the approved version.” That is more useful than attaching an old document to a confident answer.

Keep checking separate from permission to act. An assistant that finds a supporting passage should not automatically send feedback or change a learner record. Those boundaries belong in the teaching assistant's permissions.

Make citation checks part of testing

Before relying on an assistant, test whether it handles evidence failures as well as straightforward questions. NIST's Generative Artificial Intelligence Profile, published in July 2024, warns that fabricated citations can encourage misplaced trust. Its action MS-2.5-003 calls for reviewing sources and citations during pre-deployment measurement and ongoing monitoring.

For a teacher-facing workflow, turn that principle into a small educator-reviewed test set. Include a relevant source that does not support the claim, an outdated rubric, a missing passage, and conflicting approved instructions. Write the expected response before running the assistant. For the outdated-rubric case, success might mean requesting the current version rather than answering from the old one.

Record the question, claim, source version, passage location, review outcome, and correction. Store only what is needed for the review in an approved location; do not copy identifiable learner work into a public testing tool. Re-run the cases when the model, prompt, or source collection changes.

This is one part of a broader production readiness review. Counting citations or working links is not a substitute for checking whether the evidence supports what users will act on.

An educator's citation-checking checklist

Before using the answer, confirm that you can say:

• I can identify the specific claims that matter to this teaching task.
• I have opened the original source, not only the assistant's summary.
• I can locate the passage supporting each important claim.
• The answer preserves the source's qualifications and does not overstate the conclusion.
• The source's date, version, population, and scope fit this use.
• Missing or conflicting evidence has a named next step.
• Any learner-facing action still has the required educator approval.

If you cannot complete a check, record the gap rather than filling it with confidence. Narrow the claim, find better evidence, or keep it in review.

Start with one answer that could change a lesson or a learner's feedback. Record its important claim, supporting passage, review outcome, and next step. Use that record to decide whether the answer is ready to use.

About the author

Mustapha Liaichi writes about practical AI systems at Coznix. This AI-assisted article draws on the linked research and guidance; the worked example and review labels are illustrative, not classroom trial results. For questions or corrections, use the contact page; publisher information is on the about page.