Take the personal details out before you paste.

What happens

1. Paste or open some text, or paste a PNG or JPEG image of a printed letter. An image is read into text on your computer, and you correct that text against the image.
2. Review the suggestions. A small model and a set of rules mark details to remove. You can keep or change each one, and select anything else to remove, everywhere it appears.
3. Check the copy. Read the whole copy, tick that you have checked it, then copy it or save it as a text file. Your original is not changed.

The result is a copy you have checked, not a guarantee that it is safe to share. The model and rules miss things, so read the parts they did not mark as well.

Models used

Ettin 68M PII

Suggests names, addresses, contact details and identifiers. About 137 MB, MIT licence, revision 500262a2aaf913825ef750ef255c3fe437cd8e64. It reads at most 12,000 characters at a time; longer text is refused rather than cut short, and can use the rules alone.

Tesseract English (tessdata_fast 4.1.0)

Reads printed English text from an image, run by Tesseract 5.5.2. About 4 MB.

Rules, no model

Email addresses, UK phone numbers and labelled references. Also secrets pasted by mistake: a value after a label such as “Password:” or “API key:”, common API key formats, private keys and card numbers. A password written in a sentence with no label is not found.

Both models run on your computer’s CPU and download nothing in normal use. On a Mac, the model runs in a separate process with network access denied. Dates, times and job titles are not suggested.

What has been checked

The main check is one invented housing repairs letter, with the details to remove written down before the model ran. The model and rules made eleven selections: ten fully covered details that should go, one caught only part of a repeated flat number, and a description that would identify the home was missed entirely. With those added by hand, the copy matched the expected one exactly. Without them, it would have kept details that should have gone.

That is one example, not a measure of how often it finds everything. I tried two other small models on the same letter and chose this one; none of them found everything unaided.

I have used Share Copy in my own work since 24 September 2026. That is one person’s use.

Availability