Compare two text versions with ChatGPT

To compare two text versions with ChatGPT, clearly label which is old and which is new, then request a change table with exact quotes from both sources. Separate added, removed, changed, and unchanged items. Mark movement separately and check the final report against the originals.

This method is useful for reviewing an updated work instruction, learning material, or short guide. You should finish able to see what each version says and which text supports each conclusion.

Preserve both sources and state the comparison direction

Keep both originals. In a separate working note, identify which one you call the old version and which one the new version. Do not ask the assistant to guess their order from similar file names alone.

Paste short material in two labeled sections. Give paragraphs or items IDs, such as O1 for the old text and N1 for the new. An ID helps locate a source; matching numbers do not yet establish that two items correspond.

If the source is a PDF, first check that the working text includes its tables, notes, and final page. The guide to checking PDF text and tracing answers to their source can help with that stage. A good comparison prompt cannot restore an omitted paragraph.

OpenAI’s lesson-material revision example also emphasizes preserving previous versions and verifying changes (source 2). Here, the task is narrower: compare two texts without editing either source.

Example: two versions of upload instructions

The following instructions describe a fictional program. They are not the terms of an actual product. The IDs mark each item’s position in the text.

Old version:

O1: Save the file in CSV format.
O2: The maximum upload size is 5 MB.
O3: Select the date column before uploading.
O4: After the preview, click the Import button.
O5: When processing finishes, download the processing report.

New version:

N1: The maximum upload size is 10 MB.
N2: Save the file in CSV format.
N3: If the file has a date column, select it before uploading.
N4: After the preview, click the Import button.
N5: If an error occurs, download the error file.

Do not pair O1 directly with N1: one concerns format and the other size. First connect corresponding requirements, then check their wording and meaning.

What the comparison table should distinguish

In this exercise, each pair receives one primary status:

• Changed: the same requirement appears in both texts, but a number, condition, or action changes.
• Added: a separate requirement appears in the new version.
• Removed: an old requirement is no longer present in the new version.
• Unchanged: the corresponding text stays the same.

Position changes are a separate field. An item may be unchanged in content but move elsewhere, or both its content and position may change. If execution order matters, check the effect of movement too.

For this example, we compare the processing report and the error file as separate requirements: both the artifact and the download condition differ. The shared word “download” is insufficient to establish that one replaced the other.

Prompt: show the difference and link it to its source

Give ChatGPT both labeled texts and ask:

Compare the old and new texts. Produce a comparison report only;
do not edit the sources or execute the actions described in them.

Match requirements by content, not just by item number.
For each row, provide:
old ID and exact quote | new ID and exact quote |
primary status | position change | short explanation.

Use these primary statuses: changed, added, removed, or unchanged.
Mark movement separately. Do not invent an old source for an added item
or a new source for a removed item. Write "not present" on that side.

Cover every O and N item once, including unchanged items.
Separately check numbers, units, "if" conditions, and action timing.
If a pairing is uncertain, explicitly flag the uncertainty.
Finish with counts for the primary statuses.

Describing the goal, sources, output, and boundaries follows OpenAI’s prompting guidance (source 1). The key checkable requirement here is that each conclusion relies on text found in both versions whenever both sides exist.

Reference comparison: two changed and two unchanged pairs

The following table is a manually prepared example for this article, not a transcript of an actual ChatGPT response. The IDs stay alongside the quotes so you can easily locate each full sentence.

The primary statuses are two Changed, two Unchanged, one Removed, and one Added. Six rows cover ten source items: four pairs have sources on both sides, while the remaining two rows belong to one side each.

Position changes occur in two pairs. That is a separate flag; do not add those flags to the primary counts as though they were two more independent items.

Do not attach unsupported conclusions to a changed number

O2 and N1 change the maximum size from 5 MB to 10 MB. The text says nothing about upload speed. “Uploads are twice as fast” therefore does not follow from this comparison.

O3 requires selecting the date column unconditionally. N3 adds a condition: the action is required if the file has a date column. This does not establish whether the program detects the column automatically.

O4 and N4 are identical: clicking Import happens after the preview in both versions. Check “after,” “before,” “only,” and negative wording separately, because a small replacement can change an instruction.

Second check: account for every item

Extract the old IDs from the table: O1, O2, O3, O4, O5. The new-side list should be N1, N2, N3, N4, N5. Each ID appears once in this reference.

Then compare the quotations word for word against the source text. That establishes whether a quote is actually present. Review meaning separately: even an exact quote may have been matched to the wrong item.

If the assistant finds only the changed number, request a separate check for removed and added requirements. If one paragraph has been split into several, this simple one-to-one matching rule no longer fits: allow several items to map to one requirement and check coverage again.

When to compare manually and how to use the report

For five short items, manual comparison may be simpler. ChatGPT is useful for drafting an initial table and explanation when you check the result against both originals.

For long texts, preserve section headings and item IDs. If comparing in batches, track which sections you have reviewed and look for moved items in other sections too. Absence from one batch does not yet establish removal from the whole document.

Use the checked change list to plan an instruction update. For example, separately flag the size limit, the new date-column condition, and the difference between downloadable artifacts. Finding a described change does not establish that it is correct or approved for use; check that with the person responsible for the source.

Common questions about comparing texts

Does moving an item always change its meaning?

Check wording and position separately. In this example, the CSV requirement moved with identical wording. If execution order matters for the document’s purpose, consider the effect of sequence as well.

Is a summary of changes enough?

Include source IDs and quotations with the summary. Without them, it is hard to determine what the report missed or which sentences the assistant paired incorrectly.

Full source addresses

To open a source, copy its full address into your browser.

1. OpenAI, Prompting: https://learn.chatgpt.com/docs/prompting
2. OpenAI, Revise a lesson package: https://learn.chatgpt.com/use-cases/revise-lesson-package