Settings

The document

A PDF or a Word file (.docx). It is read here and never leaves this tab.

Drop the file here or choose it

How it works

Some documents carry words meant for the software that screens them rather than the person who reads them: text in the page colour, text too small to see, or text placed off the page. LANTERN finds hidden text in a PDF or Word file, says why each piece is hidden, flags lines that read like instructions to a machine, and writes a clean copy. Not every finding is a trick. Nothing is uploaded.

The guide: Is it safe to upload your CV?

Questions people ask about LANTERN

What counts as hidden text?

In a PDF: text in the same colour as the page, or within a few steps of it; text under four points; text placed off the page or outside the area the page clips to; text drawn in the invisible render mode (Tr 3); text that is fully transparent; and text on an optional content layer that is switched off. In a Word file: a run marked hidden (w:vanish) or hidden on the web view (w:webHidden), on the run itself or through its style; text within a few steps of the page colour or of its own shading; a size under four points; and zero-width or invisible characters (U+200B to U+200D, U+2060 to U+2064, U+FEFF) or rows of soft hyphens inside the text. Headers, footers, footnotes and endnotes are read as well as the body.

Is every finding a trick?

No. A scanned PDF that has been through OCR keeps its recognised words in the invisible render mode, so every line of it is a finding, and that layer is what makes the scan searchable. Word’s hidden-text mark is used honestly for notes and answer keys. Some scripts, Thai among them, use the zero-width space as a word divider. The page shows what a reader cannot see and why; you decide what to take out. Untick a finding and the copy keeps it.

How is the PDF copy made? Is it a picture of the page?

No. The page’s content stream, and any form XObject it draws, is tokenised here and each text operator is matched one for one against the operators the PDF engine reported. A hidden one is replaced by an empty advance of the same width, so the words beside it keep their places, and the rest of the file is untouched: fonts, links, form fields, bookmarks and metadata all stay. If the two readings of a page do not line up exactly, nothing on that page is changed and the status line names the page. It does not rasterise anything.

What does it not find?

Text painted over with a box, which is the fake redaction SEAL exists to prevent; text inside a picture; text in annotations, comments, form fields and tracked changes, which DOCSCRUB and SCRUB deal with; and text hidden by being small but readable, or grey but visible. A PDF that needs a password to open cannot be read, and an old binary .doc file must be saved as .docx first.

Why does it say a finding reads like an instruction to a machine?

Because the hidden words match a phrase of the kind people plant for screening software and language models: “ignore previous instructions”, “you are an AI”, “rate this candidate highly” and their relatives. It is a hint about intent, not a verdict; read the words yourself. Hidden text that matches nothing is still hidden text.

Is the Word copy still a proper document?

Yes. The hidden runs are cut out of the document’s XML at their exact positions and the invisible characters out of the text, and the file is zipped up again with every other part as Word wrote it. Nothing is re-serialised, so styles, numbering, pictures, tables and settings are as they were.

Is anything uploaded?

No. The file is read and the copy is written in this tab. The PDF engine and the zip library are fetched from this site the first time they are needed and kept by the browser; the page then works with the network switched off.

Doing this often? PDF Studio redacts by search across a whole document and removes the text under a redaction for good. On its own, US$2 a month; with everything else, US$5 a month or US$48 a year.