Can ChatGPT extract text from an image? Yes, with three catches

Short answer: yes. ChatGPT, Claude and Gemini can all read the text in a clear photo or screenshot. They fall short in three situations: when you need an exact, word-for-word copy, when the document is a long scan, and when you need a file at the end and not a chat reply.

What a chatbot does well

Upload a sharp screenshot of an error message, a slide, a menu or a page of a book, ask for the text, and you will usually get it back correctly. A chatbot can also do things a plain text reader cannot: explain what the text means, translate it, or pull out just the phone number. If you have one image and you want the gist or a quick copy, a chatbot is a fine tool and you do not need anything else.

Where it goes wrong

The trouble starts when the job is to copy, not to understand.

• Tidying. A chatbot tends to fix what looks like a mistake. A typo is corrected, an unusual spelling is normalized, punctuation is added, a clumsy sentence is smoothed. The result reads better than the page and is no longer a copy of it.
• Summarizing tables. A long table or a list of line items may come back shortened, or described in a sentence, instead of row by row.
• Guessing. A word that is blurred or cut off at the edge of the image is often replaced by one that fits. Nothing in the reply tells you a guess was made.
• Long scans. A scanned PDF of 60 pages is too much text for one reply. The assistant often summarizes, skips pages or stops partway, and you end up pasting pages in one at a time.
• No file. The text usually arrives in the chat window. Getting 60 pages of it into a document is manual work.

None of this matters for a shopping list. It matters for a contract clause, a quote you plan to cite, a serial number, or an invoice total.

Why this happens

A chat assistant is built to be helpful in a conversation. When you show it an image, it reads the image and then writes a reply, and the reply is shaped by what seems most useful. Fixing a typo seems useful. Shortening a long table seems useful. Filling a gap with a likely word seems more useful than leaving a hole.

Exact copying asks for the opposite habits: change nothing, shorten nothing, and leave a visible gap where the page cannot be read. You can push a chatbot in that direction with a careful prompt, and on a single page it often works. Over many pages the instruction tends to wear off, and you have no easy way to see where.

How to test it yourself

Pick an image where you can check the answer, ideally one with a table, a few numbers and an odd spelling or two.

1. Ask a chatbot to extract the text exactly as written, without correcting anything.
2. Compare the numbers first, digit by digit.
3. Look for the odd spellings. Were they kept or fixed?
4. Count the rows in the table and compare with the image.
5. For a multi-page PDF, check that the last page is there in full.

If everything matches, your document is in the easy category and a chatbot will do. If a number, a spelling or a row is different, you need a tool that copies.

What to use instead

For exact text, long scans or a file you can keep, use a tool set up to copy and nothing else. Our image to text converter reads photos, screenshots and scanned PDFs page by page. Each page is copied verbatim in its original language: reading order is kept, table cells are separated by tabs, and text that cannot be read is marked [illegible] instead of being invented. The text appears next to the original page so you can check and correct it, and you download the whole document as TXT or Word.

It returns the text, not a rebuilt copy of the page: fonts, images and exact layout are not reproduced. It is also meant for printed and typed text. For cursive and old letters, the handwriting to text converter is the better fit.

The first page of every document is free and needs no account, so you can run the same test on it. A whole document is $5, paid once.