Dennis G Perry, PhD, MBA
2 May 2026

To stop LLMs from corrupting documents, the safest approach is to separate content generation from file manipulation. Let the LLM draft, revise, summarize, or restructure text, but do not let it freely rewrite the underlying .docx, .pptx, .xlsx, or PDF package unless there is a controlled process around it.

The core problem

Most “corruption” occurs because complex document files are not really single-text files. A Word document, for example, is a structured ZIP package containing XML, relationships, media, styles, numbering, headers, footers, metadata, comments, and embedded objects. If an LLM or automation process rewrites that structure too broadly, it can break links, numbering, styles, tables, images, cross-references, or the document package itself.

Best practices

1. Never let the LLM directly rewrite the whole binary document

Avoid workflows where the LLM is asked to “edit this Word file” by regenerating the entire file structure from scratch. That is where the risk of corruption is highest.

Better workflow:

Extract the text.

Have the LLM revise the text.

Reinsert the revised text into the document using a deterministic tool or controlled script.

Preserve the original styles, sections, headers, footers, tables, images, and numbering.

2. Use a copy, never the original

Before using an LLM-assisted workflow, create:

Original document

Working copy

Output copy

Keep the original untouched. Every AI edit should be reversible.

3. Use tracked changes or patch-based editing

The safest model is not:

“Rewrite this whole document.”

The safer model is:

“Change paragraph 4 to this.”

“Replace this heading with that.”

“Insert this section after Section 2.”

“Do not modify tables, figures, captions, styles, headers, footers, or references.”

Patch-based edits reduce the chance of unintended structural damage.

4. Lock down document structure

For Word documents, protect or preserve:

Styles

Section breaks

Headers and footers

Footnotes and endnotes

Tables

Captions

Cross-references

Table of contents fields

Numbered lists

Images and anchors

Bibliography/reference fields

If the LLM is only revising prose, it should not touch those elements.

5. Avoid round-tripping through Markdown unless layout is simple

Markdown is useful for drafting, but it can destroy document fidelity when converted back to Word or PDF. It often loses:

Precise spacing

Custom styles

Complex tables

Image placement

Footnotes

Numbered heading hierarchy

Captions

Cross-references

Page breaks

Use Markdown only for simple drafts, not for polished legal, technical, proposal, patent, or formatted business documents.

6. Use templates

Create a clean Word template with approved:

Fonts

Margins

Heading styles

Caption styles

Table styles

Reference style

Footer/header layout

Then insert AI-generated text into that template. Do not let the AI invent formatting each time.

7. Validate the file after generation

After an AI-assisted document is produced, open and inspect it in the native application.

For Word, check:

Does the file open without repair warnings?

Are headings correct?

Are tables intact?

Are images still anchored properly?

Does the table of contents update?

Are references still present?

Are page breaks preserved?

Are numbered lists still sequential?

Then save it once manually from Word. This often normalizes the file.

8. Use PDF only as final output

Do not use PDF as the editing source unless there is no alternative. PDFs are display-oriented, not editing-oriented. Extracting and regenerating content from PDF often causes layout errors.

Preferred sequence:

Word source → controlled edits → Word review → PDF export

Not:

PDF → AI extraction → regenerated Word → PDF

9. Give the LLM strict boundaries

Use instructions like this:

“Revise only the body text. Do not change headings, numbering, tables, figures, captions, citations, page breaks, headers, footers, or styles. Return only the replacement text for the specified section.”

Or:

“Identify recommended edits, but do not rewrite the document. Provide a change list with section names, original text, revised text, and rationale.”

10. Use structured intermediate formats

For high-value documents, use a controlled editing format such as:

JSON change list

CSV edit register

XML patch

Tracked-change table

Section-by-section replacement text

This gives you auditability and prevents the AI from making invisible structural changes.

A reliable workflow

Use this process:

1. Keep the original document untouched.
2. Extract only the target text.
3. Ask the LLM to revise that text.
4. Review the revised text.
5. Insert it into a copy of the document.
6. Preserve the original formatting and styles.
7. Validate the file in Word, PowerPoint, Excel, or Acrobat.
8. Export final PDF only after review.

For your kind of documents

For formal Word documents with IEEE references, headings, tables, and graphics, the safest instruction is:

“Use the existing Word document as a template. Preserve all styles, headings, numbering, tables, captions, references, headers, footers, and page layout. Replace only the specified prose sections. Do not regenerate the entire document from scratch.”

That one instruction prevents many failures.

Bottom line

LLMs corrupt documents when they are allowed to behave as file editors rather than content editors. The solution is to make the LLM produce controlled textual changes, while deterministic document tools handle the actual file structure.