How to Compare Two Documents with AI Without Missing Changes

An AI comparison of two documents can sound convincing while missing a deleted rule or reversing which version changed. Save the original and revised documents separately, compare the actual text first, and require evidence for every reported change. A fluent summary is not a complete change log.

This guide shows a reproducible six-line versus six-line exercise with seven distinct section IDs. It also explains when Microsoft Word or Google Docs is a better first step than asking AI to find differences. Every document and policy in the example is fictional.

Use a document comparison tool before AI when precision matters

If you need to find every wording change, start with the comparison functionality built for that purpose. AI can organize and explain confirmed differences later, but an AI summary alone should not be considered an exhaustive redline.

• Microsoft Word on desktop: Use Review > Compare and identify the original and revised files. In supported Windows desktop editions, the legal blackline workflow normally creates a third comparison document without changing the sources. Check the comparison options before confirming; certain choices can instead place results in the original or revised file. Read Microsoft’s comparison instructions.
• Google Docs on a computer: Open the primary document, choose Tools > Compare documents, select the other file, and open the resulting document. Google describes the differences as suggested edits in a third document. Its comparison does not cover changes in document column layout or table column width. Read Google’s comparison instructions.
• If both revisions belong to one Google Doc: Examine Version history instead of making unnecessary copies. Earlier-version access requires editing permission. See Google Docs version history.

Use the AI workflow below when you want a structured, plain-English inventory of changes. It is a companion to the original documents and a native comparison, not a replacement for source review.

Step 1: Label OLD and NEW before comparing

Call the original OLD (v1) and the revision NEW (v2). Preserve them separately. Do not trust a file name such as “final-final.docx” as evidence of approval or chronology.

If the documents include confidential customer details, business secrets, personal data or other restricted records, use an organization-approved local or enterprise comparison method. Sending the full files to an AI tool is a separate data-handling decision.

Step 2: Copy this fictional before-and-after example

These snippets describe a made-up equipment-lending pilot, not real rental terms. Stable section IDs let you audit the entire result.

OLD (v1)

S1 | Scope | Pilot covers registered neighborhood groups.
S2 | Booking window | Request at least 7 calendar days before pickup.
S3 | Deposit | $100 refundable deposit.
S4 | Late return | $10 per day.
S5 | Replacement | Report missing cables to the coordinator.
S6 | Kit check | Two staff members must check the kit.

NEW (v2)

S1 | Scope | Pilot covers registered neighborhood groups.
S2 | Booking window | Request at least 3 calendar days before pickup.
S3 | Deposit | $150 refundable deposit.
S5 | Replacement | Report missing cables to the coordinator.
S6 | Kit check | One staff member must check the kit.
S7 | Pickup ID | Representative must show photo ID.

Try finding every modification, deletion, addition and unchanged section yourself before looking at the answer. In particular, distinguish “removed from the new document” from “the underlying policy was formally repealed.” Only the former is demonstrated by these texts.

Step 3: Check your answer against the evidence

The exercise has five changes: three modified sections, one section removed from the text, and one added section. Two sections are unchanged.

The unchanged sections are S1 (Scope) and S5 (Replacement). They should not be described as substantive amendments.

The difference between the old and new refundable deposit is $50, but neither version states why the amount changed or when it takes effect. S4 is missing in the revised text, but the documents do not establish whether another policy contains a late-return fee. S7 mentions photo ID, but says nothing about accepted document types or how personal information will be handled.

A precise change report sticks to the evidence and leaves these questions open.

Step 4: Give AI an exact comparison contract

Use the following prompt with the two source texts, or with files whose contents are actually accessible to the selected AI product. OpenAI’s file-upload help lists document comparison as a supported task, but availability and processing vary. Do not assume that every attachment was read correctly.

Compare OLD (v1) against NEW (v2).

Use only the source documents.
Do not invent amendments, approvals, reasons or dates.

FIRST — COVERAGE
- List every section ID found in OLD and NEW.
- Identify any unreadable or missing source content.
- If either document is unavailable, say so.

SECOND — DIFFERENCES
- For every ID, classify it as MODIFIED, ADDED,
  REMOVED FROM TEXT, or UNCHANGED.
- NEW is the revised version.
- Quote a short, exact excerpt from each relevant
  version, or write "not present in this version".
- Do not equate removed text with a formally repealed rule.

THIRD — CHANGE REGISTER
Provide columns:
Section ID | Type | OLD excerpt | NEW excerpt |
Why it may matter | Question for a human reviewer

FOURTH — COMPLETENESS
- Report counts for each change type.
- Account for all distinct section IDs.
- Check numbers, dates, amounts, units and negations.
- Separate observed changes from interpretations.
- Disclose any sections not checked.

OLD (v1):
[PASTE THE ENTIRE OLD DOCUMENT]

NEW (v2):
[PASTE THE ENTIRE NEW DOCUMENT]

The comparison direction matters. A clause appearing only in OLD is removed from text; a clause appearing only in NEW is added. Reversing the versions reverses those labels.

Even if the AI gives the correct total number of changes, you still must review each individual excerpt. Five reported differences are not necessarily the right five differences.

Step 5: Verify the count without AI

The union of section IDs is S1 through S7. The control totals are:

• Modified: S2, S3, S6 — three IDs.
• Removed from text: S4 — one ID.
• Added: S7 — one ID.
• Unchanged: S1, S5 — two IDs.

These groups account for all seven unique IDs.

For this article, we ran a local Python standard-library test of the two exact fictional fixtures. It parsed each ID, compared the corresponding records, and asserted the four expected groups. The test passed. No live AI model, Google Docs comparison, or Microsoft Word comparison was run for this example. It is a checked reference answer, not a claimed AI accuracy result.

When real documents have no stable IDs, use headings and neighboring text to propose matches. Move, split and merged sections may need manual mapping. Never force an unclear paragraph into a neat match just to produce a clean table.

Step 6: Look for changes a text comparison may miss

A paragraph was moved, not rewritten

A moved paragraph may be displayed as a deletion and addition. Check whether its exact wording is present in both files in a different location. Moving a clause can matter, but do not call it a new obligation solely because its position changed.

Table layout changed

A text-only comparison cannot establish that page breaks, borders, column widths or positioning remained intact. Google explicitly excludes document column layout and table column width changes from its comparison results. Inspect the rendered files when presentation affects interpretation.

Numbers or negations are easy to overlook

Manually compare small but consequential tokens such as 7 versus 3, $100 versus $150, two versus one, and must versus must not. A summary that captures the general topic but misses one number is not a reliable review artifact.

Tracked edits are hidden

In Microsoft Word, No Markup hides revision indicators; it does not accept or remove tracked changes. Microsoft’s guidance says you must accept or reject changes to resolve them. Do not mistake a visually clean document for a finalized revision.

Comments are mistaken for current rules

Reviewers may leave a suggestion in a comment that never becomes effective text. Google Docs has an option to include comments in the comparison file, but you should decide whether to review comments separately from the actual document wording.

Step 7: Turn facts into human review questions

For the fictional example, an AI assistant can help prepare sensible follow-ups:

• S2: Can the lending desk support a three-day booking window?
• S3: Who has authorized the $150 refundable deposit?
• S4: Was the late-return clause intentionally removed from this version, and does it remain elsewhere?
• S6: Is a one-person check permitted, or is this an editing mistake?
• S7: What types of photo ID are accepted, and how should staff avoid unnecessarily collecting ID data?

Those questions are not conclusions. The two files provide no approver, effective date, legal interpretation or evidence of actual operations.

For contracts, personnel rules, medical procedures and other high-stakes material, only an authorized reviewer can decide whether the final document is acceptable.

If AI says the documents are identical

Treat “no differences found” as something to verify, not a guarantee.

First, check that both distinct versions were uploaded, read and labeled correctly. Search the originals for known contrasting values. Then compare with Word or Google Docs if those tools are appropriate. If the files contain scans, charts or handwritten annotations, check whether the AI could inspect those regions at all.

Ask the tool to list its inspected sections and missing material, rather than repeating “no changes.” If the tool cannot access one document, supply the actual text or a supported file instead of requesting a more confident guess.

Final document-comparison checklist

• OLD and NEW are clearly labeled and preserved.
• Native or deterministic comparison has been considered for completeness.
• Every changed, added, removed and unchanged section is accounted for.
• Dates, amounts, names, negatives and units were checked against source text.
• Layout, tables, attachments and scanned sections were reviewed separately.
• Proposed comments are not mistaken for accepted wording.
• Interpretations are separated from observable text changes.
• The responsible person reviews the final document before approval.

Official sources and verification scope

• Microsoft Support — Compare document differences using legal blackline. Comparison direction and result-document behavior.
• Microsoft Support — Accept or reject tracked changes in Word. Why the No Markup display does not remove changes.
• Google Docs Editors Help — Compare documents. Creating a comparison file and layout exclusions.
• Google Docs Editors Help — Find what’s changed in a file. Earlier versions and permissions.
• OpenAI Help — Uploading files and audio to ChatGPT. Document comparison as a file-based task and availability caveats.

Verification note: Both versions are original fictional text fixtures. A local Python comparison independently confirmed all five changes and both unchanged sections. We did not run an AI comparison on the files or test live Word/Google Docs UI actions. This guide reports the fixture and the providers’ documented capabilities separately.