A name you can see on the screen is not always text. In a scanned contract, a photographed letter or a screenshot of an email pasted into a report, the name is a picture. Find and Replace does not see it. An AI tool may.
Where personal data hides in images
- Scanned PDFs. Each page is an image. All the text, including names and numbers, is pixels.
- Photos of documents. A phone picture of a letter, an invoice or a form.
- Screenshots in Word or PowerPoint. An email or chat pasted as an image to "show what was said".
- Signature blocks and stamps. Often images, with a name beside them.
- Letterheads and logos. The company name and address in a banner.
- Business cards, whiteboards and forms photographed at a meeting.
- Photo metadata. Pictures can carry the date, device and sometimes the location they were taken.
Why it matters for AI tools
AI tools can read images. If you upload a scan of a client's contract, the tool reads the names in it just as it would in typed text. Removing names from the typed parts of a file and leaving the pictures leaves the client exposed.
The reverse also happens. A file that looks like clean text can contain a scanned appendix with everything in it. Open every page, not only the first.
A quick test: can you select the text?
Open the PDF and try to select a line with the cursor. If you can highlight individual words, the page has a text layer. If the selection grabs the whole page as a block, or nothing, the page is an image. Mixed files are common: typed pages with a scanned signature page or appendix at the back. Test the last pages, not only the first.
The same test works inside Word: click a screenshot, and if it selects as a single picture, its contents are not searchable text.
Screenshots of chats and emails
These deserve their own mention because they feel harmless. A screenshot of a text-message thread shows a phone number or contact name at the top, the other person's name in each bubble, and often a time and place in the messages. A screenshot of an email shows the full header with both addresses. If you need the content, copy the words, replace the names, and leave the picture out. If a screenshot is the evidence, treat it as a document in its own right, and do not drop it into the middle of a report you plan to share with an AI tool.
Dealing with it by hand
- Find the images. Scroll through the document and note every scan, screenshot and picture.
- Decide if you need them. Often the answer is no, or the text in them is a line or two you can retype.
- If you need the content, retype or copy the text and put the placeholders in, then leave the picture out.
- Do not blur. A blurred name is often still readable, and a blur can sometimes be reversed. A solid shape that is flattened into the image is safer. Better still, leave the image out.
- Strip photo metadata if you must send the picture.
What a tool can do
Scanned PDFs are read as text, page by page. For Word, PowerPoint, Excel, email and similar files, the text inside pictures is recognized and placed where the picture was, in a block marked "Text from image". From there it goes through the same steps as the rest of the document: names and numbers are found, you review them, and they become placeholders. The pictures themselves do not go into the safe copy. Small icons and decoration are skipped.
In ClientSafe, the strip above the safe copy shows how many images were read as text. English, Spanish, French, German and Arabic are read.
A realistic case
A client sends a folder: a typed agreement, a scanned bank letter, three photos of a handwritten note, and an email thread with screenshots of text messages. The agreement is simple. The rest is where names hide. A tool that works on typed text alone will miss most of them, and a quick search of the typed agreement will say "clean". Doing a text search is not a check on scans. The review step has to include the pages that were read from images.
Where recognition fails
Reading text from pictures is not exact. Handwriting, low-resolution photos, skewed pages, stamps over text and unusual fonts can all produce errors. A misread name can slip past detection because it no longer looks like a name. If the reading is uncertain, ClientSafe marks the document as Needs review. Treat those documents as the ones to read twice, comparing the safe copy with the original page.
A signature drawn as a picture has no readable text. A name typed beside it, as in "Signed: Maria Lopez", is text and is found.
Frequently asked questions
Can I leave the images in and just hope?
No. An image of a document is the document. If it holds names, an AI tool that reads images sees them.
Is a scanned PDF treated like a Word file?
It is read as text first. After that it goes through the same steps.
Does the safe copy include my photos?
No. Pictures stay out. Their text, where it could be read, is included and made safe.
What about handwritten notes?
Expect misses. Handwriting is the hardest case for any text recognition, so review those pages with extra care or retype them.
ClientSafe reads scanned PDFs and the text inside pictures in documents, then finds and replaces names and numbers like it does for typed text. The review step shows what was found, and a document it is unsure about is marked Needs review.