Extract Text From Images With ChatGPT

Sometimes you have a picture of words. A photo of a page from a book. A screenshot of a chat you cannot select. A snap of a whiteboard from a meeting. You do not want ChatGPT to explain the words. You want the words themselves, in a form you can paste, edit, or search.

By default, ChatGPT tends to summarize. It reads the image and gives you the gist. That is helpful for some tasks and useless when you need the exact text. The fix is to ask for a transcript in the same message where you upload the image.

What text extraction from an image really means

You are asking ChatGPT to read the visible letters and give them back as text you can use elsewhere. You are not asking for a smoother version or a cleaner rewrite. If the source has a typo, the transcript should keep the typo. If a line is unreadable, the transcript should say so instead of guessing.

When to extract text instead of summarizing

• You need to search a scanned page for a specific phrase
• You are quoting a chat or email in a report and cannot select it
• You want to paste code from a photo into your editor
• You are translating a menu, sign, or product label
• You are archiving handwritten notes into a searchable file

Prompt for a scanned page

Image: a scanned page from a printed report.
Task: return the text exactly as it appears.
Keep paragraph breaks and bullet markers.
Wrap unreadable words in [illegible].
Do not fix spelling, grammar, or numbers.

Why this prompt works

It tells ChatGPT the goal is a copy, not a polish. It names how to handle the gaps you cannot avoid with a scanned page. Without those two lines, ChatGPT will quietly correct things that were not really wrong and hide the parts it could not read.

Prompt for a screenshot of a chat

Screenshot: a chat window with two people.
Task: return the messages in order.
Format each line as: Name: message text.
Use the exact names as they appear.
Do not summarize, translate, or reword anything.

Prompt for a photo of handwritten notes

Photo: a page of handwritten meeting notes.
Task: transcribe the words as text.
Keep the original order and rough grouping.
Mark unreadable words as [?].
If a word could be two things, list both like [maybe: quota or quote].

What to include in your message

• The source type, such as scan, screenshot, or photo
• A clear ask for a verbatim transcript, not a summary
• How to mark parts that are unreadable or ambiguous
• Formatting rules such as keep line breaks or keep bullets
• Any language or script hint if the text is not in English

How to refine the transcript

If a section came back wrong, point at that section rather than reuploading. You can say something like line four should be five words, please look again and mark unclear words. A targeted fix usually works better than a full retry, especially on long images.

Common mistakes

• Asking for a summary when you actually need the exact words
• Sending a photo taken at an angle that warps the letters
• Trusting numbers or dates from a blurry image without checking them
• Letting ChatGPT silently correct typos or grammar in the source
• Skipping the mark unreadable rule so guesses look like facts

How to judge the result

Read the transcript with the image next to it. Check the first line, the last line, and any number or name in the middle. Those are the spots where errors matter most. If you find any drift, you know to review the rest before you trust it.

Takeaway

A transcript is only useful if it is honest about what the source really says. Ask for the exact words, mark the gaps, and check the parts that matter. Verbatim beats polished every time.