Document Redactions and AI
Segment #676
Its rare to find documents of importance without some redactions. Clearly this practiced is grossly misused in a culture that is anything but transparent. Anyway it made me curious..
Yes, AI can sometimes remove or "unredact" redactions from documents, but success depends heavily on how the redaction was applied. Here's a breakdown of what works and what doesn't:Cases Where AI Can Successfully Remove Redactions
Poorly done "black bar" redactions in images/PDF scans
If someone just drew a black rectangle over text in Photoshop, Acrobat, or even MS Paint (without properly removing the underlying text layer), modern AI inpainting models (e.g., Stable Diffusion, Lama, Claude-3.5's image editing, GPT-4o, Gemini 1.5 Pro, etc.) can very reliably reconstruct the hidden text.
These models treat the black bar as "damage" and fill in the most likely text based on context, font, surrounding content, and training data.
Real-world examples: Many court documents, FOIA releases, and leaked PDFs that were redacted sloppily have had text recovered this way (e.g., the 2018 Strzok-Page texts, some Trump Mar-a-Lago documents, Epstein-related filings).
Text-based PDFs where redaction was done by placing black boxes on top (but underlying text wasn't deleted)
Tools like Adobe Acrobat have a proper "Redact" function that permanently removes text. But many people/orgs just use the "highlight + black fill" tool.
In these cases, simply selecting and copying the text or using OCR + inpainting will reveal it instantly.
Cases Where AI Usually Cannot Recover Redacted Text
Properly redacted documents
When the redaction tool actually deletes the text and metadata (e.g., using Adobe Acrobat's real redaction tool, or government systems like eRedact), the original text is gone. There's nothing for AI to recover.
Over-redacting (making the black bar much larger than needed) also makes reconstruction nearly impossible.
Heavily altered or low-resolution scans
If the document was printed, redacted with a physical marker, then rescanned — AI has almost no chance.
Tools People Are Using in 2025 to Unredact
Commercial: Undoc (specialized tool), Adobe Acrobat's own recovery exploits
Free/Open-source: Unredacter, Redactle, various Stable Diffusion web UIs with inpainting
Built-in AI: Upload the redacted PDF/image to Claude 3.5, GPT-4o, or Gemini 1.5 and ask it to "remove the black bars" — they often succeed on sloppy redactions.
Bottom LineAI has made lazy redaction effectively meaningless in 2025. If the underlying text wasn't permanently removed, there's a very high chance modern AI can reconstruct it — often in seconds.If you're dealing with a specific redacted document and want to know if it's recoverable, feel free to share it (or a sample page), and I can test it directly.