Extract Text From a Scanned PDF With AI

The problem

You open a PDF and try to copy a paragraph. Nothing happens. That often means the PDF is made from scanned page images rather than real selectable text. You could retype everything manually, or you can ask AI to read the scanned pages and return editable text. This is sometimes called OCR, which means turning text inside an image into editable text. The important rule: AI must clearly mark anything it cannot read confidently instead of guessing.

What you need

• The scanned PDF, for example scanned letters, forms, old documents, printed reports, receipts, notes or archived paperwork
• ChatGPT or Claude
• No extra software is needed

Safety first

Scanned PDFs can contain sensitive information. Before uploading, check for names, addresses, signatures, ID numbers, account numbers, financial information, medical or legal information, and confidential business information. Only upload documents you are authorised to use with the AI service. If necessary, use a copy with sensitive pages or details removed. The goal is transcription, not rewriting: AI should preserve what the document actually says. It must not rewrite sentences, "improve" grammar, silently correct spelling, invent words that are blurry, or fill missing text from context. Text it cannot read should be marked [UNREADABLE], and text it is unsure about should be marked [UNCLEAR: possible text].

Optional follow-up prompts

Steps

1. Upload the scanned PDF to ChatGPT or Claude. For a very long PDF, you can process it in smaller sections if needed, such as pages 1–10, 11–20 and 21–30. This makes it easier to review the result and spot extraction errors. Normal short PDFs do not need this.
2. Paste the supplied prompt.
3. Let AI extract the text page by page. For example, a scanned page saying "Invoice received 18 September 2026. Delivery expected Friday. Contact James if there are any delays." should come back as --- PAGE 1 --- followed by exactly those three lines, unchanged.
4. Check the REVIEW NEEDED section, and compare those pages with the original PDF before using the extracted text. For example, if a blurry scan appears to say "Delivery expected Fr?day.", AI must NOT quietly decide it says Friday. It should write "Delivery expected [UNCLEAR: Friday]." so the uncertainty stays visible.
5. Take extra care with tables and handwriting. Scanned tables are harder than ordinary paragraphs: AI may be able to rebuild Date | Description | Amount, but if column boundaries or values are unclear it should mark TABLE NEEDS REVIEW rather than inventing a clean table. Handwriting can be much less reliable than printed text, so unclear handwritten words should be marked UNCLEAR or UNREADABLE, not guessed from context.

What success looks like

Before: you have a scanned PDF whose text cannot easily be copied. After: you have editable text with a page-by-page structure, headings and lists where possible, visible markers for unclear text, and a short review list showing which pages need checking. You did not have to manually retype the whole document. AI text extraction can save a lot of time, but scans are not always perfectly readable. Errors are more likely with blurry pages, small print, handwriting, unusual fonts, skewed scans, poor contrast and complex tables. For important documents, compare critical information against the original PDF. The original PDF remains the source of truth.

If it goes wrong

If AI summarizes or rewrites instead of transcribing, send this: Please start again. I do NOT want a summary. I want a faithful transcription of the scanned PDF. Process it page by page. Use: [UNREADABLE] for text you cannot read. Use: [UNCLEAR: possible text] when you are uncertain. Do not improve the writing. Do not fix spelling silently. Do not guess missing words. At the end, list every page that needs manual review.