Microsoft's page on cross-references says Word inserts them as fields. I read that page this week while building a pre-purchase test for Word humanizers.
That field is why the test cannot stop at visible text.
The words Figure 3 can be ordinary text, or they can be the visible result of a field that points somewhere else in the document. Both versions look the same until the figure moves, the numbering changes, or somebody updates the fields. A screenshot cannot distinguish them.
This is why a polished sample paragraph says little about whether an AI humanizer is suitable for a Word document. The paragraph may read well. The real file may contain citations, footnotes, a table of contents, section breaks, comments, captions, and links between objects. The document is the job. That is the part the test has to exercise.
I ended up with a much smaller buying test than the feature comparisons I had started with. Make a disposable DOCX containing one citation from the reference manager you use, one footnote, one small table with a number and unit, one cross-reference or automatic table of contents, and two ordinary paragraphs. Mark only one paragraph for rewriting.
That file is deliberately unimpressive. It is useful because every item has a clear question attached to it.
Can the service accept the DOCX itself? Does it return an editable DOCX? Did only the selected paragraph change? Does the citation still respond to the reference manager? Does the footnote marker open the same note? Are the number, unit, and table cells intact? Does the cross-reference still update?
A paste box cannot answer most of those questions. It only received text. Even if you carefully paste the result back into Word, the workflow has never carried the document objects through the tool. That may be fine for one paragraph in a plain draft. It is a different job from returning a thesis file ready for continued editing. That difference matters under deadline pressure.
The first check after download is boring: open the result in desktop Word and make sure it is editable. Then use Word's Compare command. In the dialog, assign the untouched test file to Original and the returned version to Revised.
Microsoft's document comparison instructions say the differences can be shown as revision marks while the original remains unchanged. That lets you inspect two things at once. You can see whether the chosen paragraph changed, and whether anything outside the chosen scope changed without a reason.
Scope matters more than it first appears. If three paragraphs need work and the service rewrites thirty pages, every one of those pages becomes review work. A label such as "file upload" does not tell you the smallest unit the tool will change. Before the request runs, you should be able to see its full boundary, especially when selecting half a paragraph causes the tool to include the rest or a larger section.
The next check is easy to skip because the file may already look right. Make the hidden machinery move.
Microsoft documents the Windows sequence as Ctrl+A, then F9. When a table holds a field or formula, repeat the update with that table selected. A working field is better evidence than a field-shaped piece of text. If a cross-reference turns into an error, the document has failed the test even if the page looked perfect a moment earlier.
Do the same kind of functional check for the citation and footnote. Change a harmless piece of source metadata in a test library and refresh the citation, then undo it. Add a footnote before the existing one and see whether numbering follows. The goal is not to torture the file. It is to make each hidden relationship prove that it still exists.
There is one limit to this test: passing it does not certify every Word object. A real thesis may include equations, tracked changes, custom styles, captions, bookmarks, embedded spreadsheets, headers, and section-specific page numbering. Put one example of every feature you would hate to rebuild into the pilot file.
The final check is about meaning, not Word mechanics. Read the changed paragraph beside the original. Confirm every number, proper name, qualifier, technical term, and claim attached to a citation. An editable DOCX proves that Word can open the result. It cannot confirm whether "may" became "does" or whether a source still supports the rewritten sentence.
Start with neither the real paper nor a whole-document run. One disposable file and one selected paragraph reveal the workflow while leaving little cleanup if something fails.
The useful question is no longer "Does this tool support Word?" It is "Can this exact Word document make the round trip, keep its working parts, and show me a change set I can reasonably review?"
The full acceptance checklist is available in this Word-document guide.
