How to tell whether an AI assistant actually checked your sources

An AI assistant gives you a clear answer and a neat list of links. Before you use it in a team update, you need to know whether those links support what it says.

A source can be real, relevant and still fail to establish the particular claim beside it. The assistant may also have read a short extract while its answer sounds as though it reviewed a whole document.

This guide gives you a practical way to check an answer. You will separate evidence of retrieval from evidence for a claim, then use a small policy example to practise correcting an unsupported conclusion.

Four questions behind a source-backed answer

Treat checking as four separate questions. Combining them into “does it have citations?” makes it easy to miss a problem.

Retrieval: Did the application return the intended document or record?
Support: Does its actual content establish this claim?
Currency: Is that the applicable date, version or status?
Coverage: What was searched, and what might be missing?

For connected tasks, Microsoft’s MCP walkthrough recommends inspecting a real tool call and its result. That is evidence about retrieval. You still need to compare the answer with the returned material.

A worked example: two documents, three claims

For example, imagine you are preparing a support handover for an online learning platform. You provide two short records. P1, the policy dated 12 October 2026, says: “Customers may request a course transfer within 14 days of purchase if the course has not started.” N1, a support note dated 15 October, says: “One transfer request was completed within 48 hours.”

The assistant produces: “Customers can transfer any course within 14 days, and all transfers are completed within 48 hours.” The sentence is easy to read, but it has changed the meaning of both records.

Compare the claim with the evidence
Claim	What the supplied records support
Any course can be transferred	P1 includes a condition: the course must not have started.
The window is 14 days	P1 counts from purchase. Omitting that starting point makes the rule ambiguous.
Every transfer takes at most 48 hours	N1 describes one completed request. It does not establish a service promise.

A defensible rewrite is: “P1 permits transfer requests within 14 days of purchase where the course has not started. N1 records one request completed within 48 hours; it does not specify a standard completion time.” This exercise evaluates the two supplied snippets, rather than a complete customer policy.

Check an answer in five passes

1. Identify the claims that matter

Purpose: Focus effort on what someone will act on.

Underline promises, dates, quantities, causes and recommendations. Split sentences containing more than one claim. A working definition may need less checking than a deadline that a customer will rely on.

For the transfer example, eligibility and turnaround are separate claims. Keeping them separate prevents one valid citation from appearing to support the entire sentence.

2. Open the record itself

Purpose: Read beyond the citation label.

Follow the source link or open the supplied file. Find the section or field the answer relies on and read the surrounding conditions. If the link fails or you do not have access, mark the claim unverified rather than assuming the document says what the assistant reports.

Check whether the application retrieved the whole record, an extract or a search preview. Ask for the relevant section if it is missing. A list of search results is a useful starting point, but its titles are not the underlying evidence.

3. Check dates and authority

Purpose: Use the version relevant to the task.

A recent comment can describe a proposed change while an older approved document remains in force. Look for document status and ownership as well as publication date. If those signals conflict, ask the person responsible for the policy.

In the example, N1 is newer than P1, but it is a case note rather than a replacement policy. Newer does not automatically mean more authoritative.

4. Ask what the search excluded

Purpose: Keep the scope visible.

Record the selected folder, repository, date range, filters and result limit. If a tool returned five records, a conclusion about every record needs additional evidence. If the search returned nothing, distinguish no match from missing access or the wrong query.

For a high-consequence decision, inspect all material claims against the authoritative records. Sampling can reveal problems in a low-stakes draft, but it cannot establish that every unchecked statement is correct.

5. Correct the answer and retain the uncertainty

Purpose: Leave a usable record of the check.

Choose whether to keep, qualify, correct or remove each claim. Ask the assistant to revise the wording using the checked evidence, then inspect the revision. A second confident answer is another draft to evaluate.

For P1 and N1, retain the eligibility conditions and remove the universal turnaround promise. Record the unanswered question: is there an approved transfer-processing target?

A template for one material claim

Claim to check:
Source title and exact link or record ID:
Relevant passage, section or field:
Source date or version:
What the source establishes:
What it does not establish:
Search coverage or missing access:
Decision: keep, qualify, correct or remove:
Checked by and date:

You can complete this in a note or alongside a draft. Keep the exact source location short enough that another person can repeat your check without restarting the research.

Try a ten-minute evidence review

Take one permitted, low-stakes AI summary and choose three claims you might repeat to a colleague. Complete the template for each. If you cannot locate support, change the wording before sharing it.

The useful result is an answer whose statements and limits you can explain. You do not need a longer bibliography when the real problem is a missing condition.

Sources checked on 21 October 2026. The worked example and template are original Coded Female teaching material. Product controls vary by application and organisation.