Find what's hiding in your PDF

Text colored so close to the page background that a human reader sees nothing, while an AI parsing the text layer reads it in full. White text on a white page is almost never legitimate.

Text drawn with PDF render mode 3, which paints no pixels at all yet stays in the machine-readable text layer. Normal authoring tools essentially never use this.

Text set below 4pt — too small to read on screen or in print, but fully extractable by software. The most common source by far is a chart scaled down to fit a page: axis ticks and labels routinely land between 1pt and 4pt, and a single figure can account for hundreds of these. What is worth reading is the text that forms sentences rather than single words and numbers.

Text that sits entirely outside the visible page area: present in the file, extractable by any tool that reads the whole page, and never drawn where a reader would see it. PDF engines normally clip it away before extraction — this scanner deliberately reads the full sheet first, so the text itself is quoted below. Text beyond the sheet of paper (outside the media box) stays out of reach.

The page is cropped smaller than its full media size, which can push content into the trimmed margins. Frequently benign in scanned files — inspect what falls outside the crop.

Optional-content layers that are off by default or explicitly marked hidden. The layer contents stay in the file and remain readable by software even though they are not shown.

One or more files attached inside the PDF. Routine on engineering datasheets, unusual on papers or contracts. The names are listed below — do not open them without checking.

The document carries JavaScript, which can run when the file is opened in some viewers. Expected in interactive forms, out of place in a static paper or contract. It is never executed here.

Comments and sticky notes whose text is invisible in normal rendering but still sits in the file. Often legitimate, but a classic place to hide instructions — always read the content.

Text matching known patterns that try to steer an AI reviewer or summarizer — for example, telling it to ignore its instructions or return a positive verdict. Suggestive, not conclusive: a paper about prompt injection may quote such strings legitimately, so read the surrounding context.

Hidden text instructing an AI reviewer to work fixed phrases into its review. Some conferences add one to every submitted paper, so that a review written by a language model gives itself away when those phrases turn up in it — meaning a finding here says nothing about the authors, who usually have no idea it is there. Where a venue does this, its own guidance is normally that a watermark revealing AI use needs no action, while hidden text that tries to influence the decision does. Matched on the shape of the instruction, so read it as "looks like a reviewing watermark", not as proof of one.