OCR can read added screenshot notes as if they were original text

Why added notes survive OCR

Tools that add text, shapes, drawings or signatures to photos, screenshots and documents write those marks into the image itself. That is what makes an annotation typeable or visible at all. When OCR then converts the image to text, it treats annotated and original characters the same way: both are just pixels. The transcription may be perfectly accurate and still attribute the sharer's words to the page. [Apple: adding text and shapes with Markup]

• Treat every character in a screenshot as coming from the image, not from the page, until you check the page itself.

Distinguish an annotation from the underlying page

Look at how the text sits in the image. Added notes often differ in style or sit outside the page's layout, but style alone is not decisive. The stronger test is comparison: obtain the accessible original page, or a permitted capture that the sharer says has no added notes. Only then can you check whether the words appear in the source at all. If the page is public and reachable, read the relevant passage there and use that as your search input instead of the screenshot text.

• Prefer the live public page, or a fresh permitted capture, as the source of any search string.
• Mark which words you confirmed on the page and which appear only in the screenshot.

A provenance problem, not a recognition typo

It is easy to assume a wrong result means OCR misread a character. Here the recognition may be flawless; the failure is where the words came from. Keep those two failure modes separate in your notes. A typo calls for a corrected transcription. A provenance problem calls for evidence about the text's origin, and often for a better artifact.

Separate observation, limit and decision

Write three short lines before you search. The observation states what the image shows or what OCR returned. The limit states what you could not check, such as an inaccessible page or an unconfirmed annotation. The decision states what you will do with the text: use it, set it aside, or search only the part you confirmed. Those lines protect you from later treating a guess as a finding. This worksheet is an editorial aid, not a platform control.

• Observation: what the pixels or transcription show.
• Limit: what source you could not access or verify.
• Decision: use, hold, or search only confirmed words.

When provenance cannot be settled

If the original page is absent, private, mismatched or ambiguous, write down that precise limit and stop there. Do not fill the gap by assuming the annotation was added maliciously. Do not treat a matching style, a badge or a privacy setting as proof of anything about the sharer. An annotation is common and often helpful; the problem is only that its words may not belong to the page you think you are searching. A labelled uncertainty is a usable result; an invented origin is not.

• Do not call a screenshot forged merely because it has added text.
• Do not access private material or try to identify anyone to resolve the question.

Carry the label into your search notes

If you proceed with any text anyway, attach the provenance label to it. Words marked as added or unknown should not later be quoted as statements from the public page. If new evidence arrives, such as the original page or an unannotated permitted capture, update the label and redo the affected searches rather than keeping earlier results that rested on the annotation. Keep each search result tied to the version of the text it used.

• Re-run affected searches if the source text changes.
• Keep the label visible wherever the string is reused.

A short provenance worksheet

For each screenshot, note the search string, whether it appears on the accessible page, whether annotations are visible, and the resulting label. A hypothetical example: a screenshot shows a product page with an added line about a recall. If the public page cannot be reached, the label is provenance unknown and the recall line is not used as page text. This is simply hypothetical. The worksheet itself is a reader aid; no tool is promised to produce it.

A couple more questions