AI Redact PDF — the difference between hiding PII and actually removing it
Redacting a PDF is one of those operations where the consequences of doing it wrong are severe and often invisible until weeks later when a journalist or opposing counsel recovers the underlying text. The right tool removes the bytes; the wrong tool draws a black rectangle on top and leaves the original text in the content stream where anyone with a free PDF editor can recover it. Here is exactly what AI Redact does, the eight categories of PII it auto-detects (including India-specific formats), and the precise difference between AI Redact's byte-level removal and the visual-overlay redaction that ships with many free tools.
How AI Redact actually works
AI Redact does three things in sequence. First, it reads every page of the source PDF and runs an AI detection pass to find candidate PII — names, contact info, identifier numbers, addresses. Second, it presents the findings as a confirmation step so you can uncheck specific instances or whole categories before applying. Third, it rasterises the redacted regions: the page bytes in those regions are converted to a pixel-level black rectangle, the underlying text is removed from the content stream entirely, and the output PDF carries no recoverable trace of the original text in those areas.
The rasterise step is the load-bearing difference from cheap-redact tools. A visual overlay (a black rectangle drawn on top) leaves the original text in the PDF's content stream — perfectly preserved, just hidden by the overlay. Anyone with a free PDF editor can remove the overlay and read the text. AI Redact removes the text from the content stream, so even with the rectangle gone, there is nothing to read.
Eight PII categories auto-detected
AI Redact ships with detectors for the most common PII types globally, plus India-specific identifiers because of the platform's high India traffic:
- Personal names. First+last name pairs, single names with title prefixes (Mr., Mrs., Dr.), names in signature blocks.
- Email addresses. Standard RFC-5322 patterns, including unusual TLDs and embedded mailto: links.
- Phone numbers. US (NPA-NXX-XXXX), India (10-digit + country code +91), UK, Germany, Canada — the common international formats.
- Indian PAN. 5 letters + 4 digits + 1 letter pattern. Often masked to last 4 characters.
- Indian Aadhaar. 12 digits, usually formatted XXXX-XXXX-XXXX. Masked to last 4 digits per UIDAI guidance.
- Bank account numbers. 8-18 digit strings in financial-document context. The context check prevents false positives on random long numbers.
- Addresses. Street + city + zip patterns, India PIN codes, UK postcodes, US ZIP+4.
- GSTIN (Indian GST identifier). 15-character format. Important for business documents where the GSTIN might be redactable depending on disclosure rules.
The confirmation step is the safety net
Before any redaction is applied, the tool surfaces every detected PII instance with checkboxes. This serves three purposes:
- Reviewing false positives. AI detectors are imperfect. A name in a quoted citation might not actually be PII you want to redact. The confirmation step lets you uncheck those.
- Reviewing for missed PII. If a particular piece of PII appears in an unusual format that the detector missed, it shows as undetected. You can either run a second pass after manual correction or use the free Redact tool to draw rectangles by hand.
- Selecting what to keep vs what to remove. Sometimes you want to redact employee names but keep the company name visible. Or redact phone numbers but keep email addresses. Per-category and per-instance checkboxes let you do this granularly.
AI Redact vs Free Redact
The two redaction tools serve different needs:
- AI Redact — automatic detection + byte-level removal. Costs credits per page. Auto-detects 8 categories of PII. Output is cryptographically sanitised — the original text is GONE. Use for compliance-grade redaction (legal production, regulatory submission, public-facing release of internal documents).
- Free Redact — manual rectangle drawing + visual overlay. Free. You draw rectangles by hand over what to redact. The visual overlay HIDES the text but does not remove it. Use for casual redaction (sending a draft with a confidential price hidden, sharing a screenshot with email addresses hidden) where the recipient is unlikely to be hostile.
Three patterns to verify
Habits that catch missed PII before it leaks:
- Run PDF Inspector on the output. After redacting, inspect the output to confirm no extractable text remains in the redacted areas. PDF Inspector shows extractable text per page; redacted regions should be empty.
- Search for the specific PII in the output. Ctrl-F in the output PDF for the redacted name / email / etc. should return zero matches. If a match comes up, the redaction didn't catch that instance.
- Test copy-paste from the redacted region. Select text across a redacted rectangle. The copied text should not include the redacted content. If it does, the redaction didn't remove the underlying text.
When to also strip metadata after redacting
Redacting page content is necessary but sometimes not sufficient. PDFs carry metadata that can leak information you also wanted to hide:
- /Info dictionary author + title. Often pre-filled with the document's original author name and creation timestamp. Run Remove Metadata after AI Redact for a fully-sanitised file.
- XMP metadata stream. Can carry version history, contributor lists, custom properties. Remove Metadata clears this too.
- Annotation author fields. If your document has annotations, each carries an author. Flatten before redacting, or strip annotation authors specifically.
Limits and pricing
AI Redact charges 2 credits per page of source PDF. The tool handles PDFs up to 100 MB. Processing happens on our servers; the file is in memory only during processing and is never persisted. Output is byte-compatible with every PDF reader and carries the redacted state forward through merge / split / compress / any further operation.
Common pairings: AI Redact → Remove Metadata → Flatten for a fully-sanitised compliance-grade deliverable. AI Redact → Compress for the smallest sanitised final file.