China AI Copilot Visits: Trace One Answer to Its Source
A citation is useful only if its source supports the answer. Trace one material sentence into an accessible document and check version, scope and permissions.
Direct answer. Select one material claim and open the cited passage using the demonstrated user's authorized access. Verify that the passage supports the claim and belongs to the applicable version of the document. A link alone does not establish either point.
Choose a consequential sentence
During a China enterprise-copilot visit, use an approved demonstration knowledge base. Pick a sentence that would change a user's action: a return period, approval requirement or product limitation. Avoid asking the host to expose private customer material.
Read the answer and source side by side. Mark whether the answer quotes, summarizes or infers. An inference can be useful when labeled and reviewed; presenting it as an explicit policy is a different claim. Record the document's effective scope as well as its date. A newer document for another region may still be the wrong source.
Keep four checks separate
| Check | What passing it establishes |
|---|---|
| Link opens | The cited resource is available to this user |
| Passage matches | The text concerns the answer's subject |
| Claim is supported | The passage actually justifies the statement |
| Scope is applicable | The version, region and audience fit the question |
Ask the host to demonstrate an unavailable or conflicting source. Does the copilot explain the limitation, invent a combined rule or route the question for review? A clean single-document example does not expose that behavior.
The trace uses knowledge-limit and interpretability themes in NIST AI RMF Core. It is our visit worksheet, not a NIST citation-quality certification.
Bring home a bounded conclusion
State which claims were traced, which documents were used and which permission context was tested. Do not conclude that every answer is grounded because one citation worked. A pilot needs a representative question set and a reviewer who can judge support, not merely link availability.
See the deployment evidence guide, or plan a copilot-focused visit around your source and access requirements.