AI Redaction: What Automatic Detection Catches, and What It Misses

Automatic face redaction is genuinely useful and genuinely incomplete. A detector finds the pattern it was trained on, covers it, and stops. Everything else in the frame that could identify someone is still there, still readable, and still your responsibility. Knowing the boundary is the difference between a photo that is safe to publish and one that only looks safe.

Step-by-step: see our guide to blurring faces and text in photos for the exact workflow. This article is about judgement: when to trust the detector, and when to reach for a manual tool.

How the automatic pass works

Our AI Redact tool runs a face detection model called BlazeFace through TensorFlow.js. The model and the runtime are fetched from a CDN, then everything after that happens on your device: your photo is drawn into a canvas, the model scans that canvas, and it returns bounding boxes for the faces it is confident about.

Each box is padded outward by ten percent so hairlines and chins are not left exposed, and the region inside is pixelated rather than blurred - the tool samples a colour every few pixels and paints solid blocks across the area. That distinction matters. Blur is a reversible-looking operation to the naked eye and, at low strengths, some detail can be recovered. Replacing a region with large flat blocks destroys the underlying values outright.

The result downloads as a PNG containing only the censored pixels. The original file on your disk is untouched, so keep it if you need the unredacted version later.

What the detector reliably finds

• Front-facing and near-front-facing faces at a reasonable size in the frame.
• Multiple people at once - group shots, classrooms, event photos.
• Faces under ordinary indoor and outdoor lighting.

For the common case, a photo you want to post from a party or a conference, one click does most of the work in a second or two.

Where it quietly fails

Detection models are trained on a distribution of images, and they degrade outside it. The failure mode is silent: you get a redacted image back and no warning that something was skipped. Check the result yourself for these cases.

Faces the model will not see

Profiles turned sharply away, faces at extreme angles, heads that are very small in a wide shot, heavy backlighting or deep shadow, motion blur, and partially occluded faces behind a hand, a mask, or a crowd. Very large images can also push faces into odd scales relative to what the model expects.

Everything that is not a face

This is the bigger gap. Faces are one identifier among many, and a face detector by definition catches only faces. In a typical photo that might also include:

• Licence plates, house numbers, and street signs that place someone.
• Name badges, lanyards, ID cards, and delivery labels.
• Screens showing email addresses, dashboards, or open documents.
• Tattoos, distinctive jewellery, and uniforms with employer names.
• Reflections in windows, mirrors, and glasses.
• Whiteboards and monitors in the background of a work photo.

For any of these, use the manual redaction tool, where you drag a box over the region and choose blur, pixelation, or a solid block. Solid blocks are the strongest option and the right one for anything genuinely sensitive - the pixels underneath are simply gone.

A practical review routine

Run the automatic pass first, then take twenty seconds to review before sharing. Zoom to full size rather than judging a thumbnail; a face that is thirty pixels tall in the preview may be perfectly identifiable in the file. Scan the edges of the frame, where people wander into shot. Read every piece of text in the image, including anything on screens. Then ask what the background says about location.

One more step that is easy to forget: metadata. A redacted photo can still carry GPS coordinates, the capture timestamp, and the device serial in its EXIF block. Redacting the pixels and leaving the coordinates is a common and avoidable mistake, so strip metadata as a separate step before publishing.

Why local processing is the point

Uploading a photo to a redaction service means the unredacted version reaches a third party before it is censored. For a photo of children, a medical document, or an internal screenshot, that undermines the entire exercise. Here the model runs in your browser tab, the image is decoded on your machine, and the only thing crossing the network is the model file coming down to you.

If you also need to annotate an image - circle a detail, add an arrow - our annotation notes cover that, and resizing before you share is a sensible last step, since a smaller image also carries less recoverable detail.