What leaves your device when an app reads a screenshot
Every screenshot carries context you did not choose to share: a name in a corner, a balance, a message thread, an address in a footer. When an app offers to read the text in that image, the first question worth asking is where the reading happens.
The cloud path
Most OCR features are a thin client over a server. The image is uploaded, recognised on hardware you do not control, and the text is returned. That is not automatically a problem, but it means the picture left your device, and the privacy policy of the service decides how long it stays. Free tiers in particular are often funded by keeping what passes through them.
The local path
Text recognition is a solved problem at small scale. A language model of a few megabytes can find character shapes on a screenshot in under a second on an ordinary laptop, with no network request at all. The trade is accuracy on unusual fonts and handwriting, where larger cloud models still win.
What local recognition cannot leak
An offline tool has no upload step, so the failure modes change shape. There is no request log, no retention window and no third party to subpoena. What remains is your own storage: if the tool keeps a history of recognised text, that history is a file on your disk, and it is worth knowing where it lives and how to clear it.
When the cloud is the right call
Handwriting, rotated photographs, dense scans and non-Latin scripts are still better served by a large model. If you need those, crop the image to the part that matters before uploading, so the surrounding context never travels with it. Cropping is the cheapest privacy control there is.
A practical test
Disconnect from the network and try the feature. If recognition still works, it is local. If it fails or hangs, it was a server call all along, which is worth knowing before you point it at something sensitive.
