PiiBlur

Automated redaction removes a lot of manual work. It does not remove the need for judgment.

Some files are low risk: internal thumbnails, test images, or media that never leaves your organization. Other files are high risk: public records releases, news footage, school media, health facility recordings, or legal evidence. Those outputs deserve a review step before publication.

Use this checklist to decide what needs review and what reviewers should look for.

1. Classify the release risk

Start by asking where the redacted file will go.

Low risk

Internal analytics
Temporary QA files
Test datasets
Internal-only training material with restricted access

Medium risk

Vendor sharing
Insurance review
Customer support attachments
Internal incident review
Partner data exchange

High risk

Public website or social media publication
FOIA or public records release
Legal discovery
Newsroom publication
School, healthcare, or law-enforcement media
Any footage involving children, patients, victims, witnesses, or private homes

Low-risk files may only need sampling. High-risk files should get human review before release.

2. Check the expected PII categories

Reviewers should know what the job was supposed to redact. A vague instruction like "remove PII" is hard to audit.

For each workflow, define the expected categories:

Real estate: faces, license plates, street signs, screens, documents
Fleet dashcam: faces, license plates, street signs, screens
School media: student faces, ID cards, screens, documents
Healthcare facility footage: faces, name badges, ID cards, screens, documents
Insurance claims: faces, license plates, documents, ID cards, credit cards
News footage: faces, badges, tattoos, plates, documents

If a category is not selected, the reviewer should not assume it was missed. They should verify whether the category was intentionally excluded.

3. Review known hard cases

Automated detection can struggle in predictable situations. Prioritize review around:

Small distant faces
Side profiles and partial faces
Reflections in windows, mirrors, and screens
Motion blur
Night footage and heavy shadows
Low-resolution or over-compressed video
Plates viewed at an angle
Text on whiteboards, badges, lanyards, and paperwork
QR codes or barcodes on packages and labels
Tattoos or distinctive clothing that identify a person

Reviewers do not need to watch every second at the same speed. They need to know where misses are most likely.

4. Inspect the edges of redacted regions

A blur box that almost covers a face may still leave enough detail around the eyes, mouth, or plate characters. Check the edges.

For faces, look for:

Eyes outside the redacted region
Hairline plus facial outline when the person is close to camera
Reflections showing an unredacted face
A face visible for only a few frames in video

For plates, look for:

First or last character outside the blur region
Plate reflection on a bumper or window
Plate visible in a neighboring frame but not the current one
Trailer or motorcycle plates missed because they are smaller

For documents and screens, zoom in. Text that looks unreadable in a preview may be readable at full resolution.

5. Check video continuity

Video review has one extra problem: consistency over time.

Watch for:

A face redacted in one frame and visible in the next
A plate missed during camera pans
Subjects entering from the edge of the frame
People visible through glass or reflections
Screens visible only when the camera angle changes

For long clips, review representative segments: the beginning, the end, transitions, camera pans, and any moment where the scene becomes crowded.

6. Verify output settings

Before approving, confirm:

The file is the redacted derivative, not the original.
The selected categories match the workflow policy.
The redaction method is appropriate: blur or pixelation.
The file format and resolution are acceptable for the release.
The filename or storage path clearly indicates redacted status.
Originals remain in restricted storage.

Many release mistakes are operational rather than model-related. Someone downloads the original, attaches the wrong file, or publishes from the wrong folder.

7. Document review decisions

For high-risk media, keep a simple review record:

Reviewer
Date
Source file ID
Redacted file ID
Categories selected
Issues found
Fixes applied
Approval status

This does not need to be heavy. A database row or ticket comment is enough for many teams. The point is to make release decisions traceable.

8. Feed misses back into the workflow

If reviewers keep finding the same problem, fix the pipeline rather than asking reviewers to remember it.

Examples:

Add license_plates to a workflow that only selected heads.
Split long videos into shorter clips before upload.
Require higher export quality from the source system.
Add a manual review gate for night footage.
Route files with many failures to a specialist queue.

The best review process gets smaller over time because repeated issues become rules.

When sampling is enough

Sampling works when the release risk is low and the workflow is stable. A practical sampling plan:

Review the first 50 outputs from a new workflow.
If misses are rare, review 5-10% of the next batch.
Keep sampling every batch, even after confidence improves.
Increase review when source quality, camera type, or category selection changes.

Do not use sampling for public records, legal evidence, or media involving vulnerable people unless your legal or compliance team has approved that process.

Build QA into the API workflow

PiiBlur gives each media item a processing status and authenticated download URL through the API. Your system can route completed jobs based on risk:

Low risk: download and store automatically.
Medium risk: sample into a review queue.
High risk: require approval before publication.

For pipeline design, see How to Build a Webhook-Based Redaction Pipeline. For batch image workflows, see Batch Photo Redaction.