Use AI to make a navigation aid for a long document: a section map, a list of passages to inspect and questions that remain unanswered. Keep the original open. A summary cannot tell you that an unread section is safe to skip.
Check access and extraction first
Confirm that the exact tool, account and feature are approved for the document and its data. A public document may be suitable for practice; a private contract, medical record or email export is a separate authorization question. Removing a cover page does not remove confidential content elsewhere.
Inspect the product’s file-upload documentation for supported files and extraction behavior. Text retrieval can miss images, scanned pages, tables or footnotes. OpenAI’s documentation, checked on 28 September 2026, distinguishes text-based retrieval from PDF visual retrieval; do not assume every account reads the same file in the same way.
Training controls and temporary mode do not supply permission to upload. The NIST Privacy Framework treats privacy as a data-processing risk, not simply a training switch.
Map before summarizing
Use a low-consequence, permitted source first:
List the document title, date/version and section headings in order. Give page or section references. Identify unreadable regions and referenced appendices that were not supplied. Do not reconstruct missing text or infer that an omitted section is irrelevant.
Compare the map with the original contents and page count yourself. A model’s claim that it read every page is not a coverage test. Keep a ledger of supplied files, sections checked and gaps. If you split the document, retain the original page labels and cross-references.
Extract the passages that answer your question
My reading question is [specific question]. Locate relevant passages in the supplied document. For each, give a short exact quote, original page/section, a tentative paraphrase and any condition or cross-reference that limits it. Mark absent information as not found. Do not decide my legal obligations or recommend an action.
This request produces candidates for inspection. The NIST Generative AI Profile describes confidently stated false content, including divergence from supplied inputs. Providing a file does not remove that risk.
A checked passage, including its exception
Fictional teaching source, page 4, section 2:
Members may book a room for up to two hours. Longer bookings require prior written approval from the coordinator. See Appendix A for room-specific restrictions.
AI-generated teaching extraction, checked against that exact text:
| Candidate answer | Check against source | Result |
|---|---|---|
| Members can book for two hours | Omits the approval route and appendix | Incomplete |
| Longer bookings require the coordinator’s prior written approval | Matches the condition in section 2 | Supported by the excerpt |
| All rooms allow two-hour bookings | Appendix A has not been supplied | Not established |
A useful question list is therefore: “Where is Appendix A?” and “How is prior written approval requested?” It is not “The booking is compliant.” The missing appendix stays visible until someone obtains and reads it.
Turn gaps into questions, not decisions
Ask for unresolved definitions, absent annexes, conflicting dates and passages you cannot interpret. Verify each quote in the original, including punctuation, units and negation. Correct the extraction before using it in a note or discussion.
For contracts or other consequential documents, use the map to prepare for the responsible professional or authority. Do not use an AI ranking of “red flags” as evidence that everything else is acceptable. The contract-literacy guide develops that narrower question-preparation role.
Your finished artifact should identify the source version, checked locations and open questions. The verification checklist can hold those checks. If the document is unreadable or incomplete, the useful result is a precise list of what must be obtained next.



