Last month our after-sales team was preparing a training session on reviewing bad reviews and complaints, and I dropped thirty-seven chat screenshots straight into the internal group. Ten minutes later legal messaged me privately: the buyer's phone numbers, delivery addresses and tracking numbers were not masked at all. Worse, five of those images had already been forwarded to an industry group. It was only while redoing them that I understood how unreliable masking really is when you just scribble over a screenshot. Whether something is actually covered, and whether the cover can be reversed, are two completely different questions. This is a record of the criteria I worked out afterwards: how much identifying information a screenshot really holds, when to mask part of an image and when to mask all of it, and how coarse a mosaic has to be before it is genuinely safe.
1. Start by seeing how much a screenshot carries
What gets missed in a chat screenshot is rarely the conversation itself but the ring around it. We reviewed one incident where the phone number was covered thoroughly, yet the contact name next to the avatar spelled out the person's full name, so the cover was pointless. The table below is the checklist I later stuck to my desk; I go through it every time before sending anything out.
| Item | Where it usually appears | What I do |
|---|---|---|
| Phone number (11 digits) | Order card, delivery details, contact profile | Mask the whole block, not even the first or last digits left visible |
| ID number (18 digits) | Support verification messages, refund receipts | Cover the entire line, block size at maximum |
| Delivery address | Second paragraph of the order details | Mask the house number too, keep nothing beyond the city |
| Avatar plus nickname | Left of the bubbles, member list | Frame the whole avatar square, with 10 pixels to spare outside it |
| Amount and last four card digits | Transfer receipts, payment success pages | Mask the whole receipt, leaving only the timestamp |
| QR code | WeCom business card, payment code | Cover the whole block; even one corner can be scanned |
| Licence plate | Logistics photos, on-site shots | Mask one extra cell beyond the width of the 7 characters |
A phone screenshot is usually 1080 or 1170 pixels wide, and a line with a phone number in it takes about 300 pixels, which looks easy to cover. The real trouble starts when you step back: at 20% zoom, the ring of leftover pixels around the selection edge becomes readable again.
2. Four mistakes I have made
- Covering only half of it: the selection covered the main digits but the last one stuck out. My habit now is to box it and then drag a dozen pixels further right.
- Block size too fine: my first version used the smallest setting and looked noisy enough on screen. Then someone enlarged the image to 400% and inferred digits from the centre colour of each block, and still recovered the odd or even value of the middle four digits of a phone number. My rule of thumb: the block must be three to four times the height of the characters being hidden, and running text such as names and addresses needs a step coarser than that.
- Using a semi-transparent block as a mask: lay a black box at 80% opacity over it, export as JPG, pull the curves, and the text underneath floats back up. Covering has to actually break the pixels apart, not paint colour over them.
- Fixing the main image and forgetting the thumbnail: once a platform generates a thumbnail, the long address bar is squeezed into a narrow strip on the left and remains readable.
3. My complete process now
- On my own computer, circle everything that needs masking with a pencil first, especially corner badges, the status bar and floating order numbers. This marked-up copy stays on file and never goes out.
- Drop the screenshot into Image Mosaic in the browser, drag out a rectangle with the mouse, one region per item, repeating as needed. If a box is off, undo the last region and redo it instead of starting the whole image again.
- Set the block size: for everyday screenshots 1080 wide I start at 24px; for ID cards, bank cards and QR codes I push straight past 60px, preferring a mush over leftovers.
- Export as PNG. A second JPG pass creates blocky colour artefacts along the mosaic edges that look like sloppy masking. It is only encoding noise, but whoever reviews the image will not feel safe about it.
- Final check at 100%, 200% and 400%, then shrink the window to its smallest to look at the whole thing and confirm nothing is left.
4. Partial masking or the whole image
| Approach | When to use it | The cost |
|---|---|---|
| Selecting regions | Training and case reviews, where the logic of the conversation must survive | Missed items are entirely on you, so the checklist has to be followed item by item |
| Masking the whole image | Placeholders that only say an image used to be here, such as a flowchart illustration | All information is lost; useless as evidence |
My choice is clear: anything someone needs to learn from gets partial masking, and an image that is only a placeholder gets blurred out entirely, so there is nothing to miss.
5. Different material needs much more different handling than you would expect
The same select-and-adjust routine applied to different material carries completely different risks. Our team now works in five categories:
| Material | Easiest thing to miss | Block size and approach I use |
|---|---|---|
| Chat screenshots | Contact name beside the avatar, group name at the top | 24-32px, with the avatar boxed separately |
| Order and payment pages | The amount actually paid at the bottom, coupon codes | From 40px, covering the whole amount row in one go |
| ID-style documents | Name and ID number boxed separately, address line left out | 60px and up, line by line, no expecting to get it in one pass |
| On-site photographs | Passers-by in the background, plates at the doorway, sticky notes on the desk | 32-48px, scanning the background at 200% |
| A frame from a screen recording | The browser address bar, file names on the taskbar | 40px, masking the address bar as one horizontal strip |
The most expensive lesson was the second category: amounts. Once I masked an order number and forgot the payment amount below it, and that image was later used for price comparison, which led to a formal internal notice. Now amounts and order numbers count as one mandatory group for me; if I box one, I have to box the other.
6. Questions I keep getting asked
Q: Can a mosaic be reversed?
Not when the blocks are large enough, because the pixels have been replaced. But with the finest setting, where each block is a few pixels across, scaling up plus image restoration can infer structural information such as digit count and length. So err on the coarse side and do not leave detail in for the sake of looks.
Q: Can I not just use the pen tool in a screenshot utility?
You can, but it is slow and unreliable. Freehand strokes leave a semi-transparent transition at the edge; one stroke on its own looks fine, but across the whole image they form readable gaps. A selected region at least guarantees a boundary calculated in one pass.
Q: Can a masked image still serve as complaint evidence?
Not advisable. Masking is irreversible, destructive processing. Use the original for evidence and the masked version only for sharing. Our new rule is to archive the original separately and save masked versions in their own folder with a _mask suffix.
Q: Dozens of screens of chat history. Is there a faster way?
I split it into two stages: screenshot everything first without touching it, then go through the mandatory-items checklist and box each image. Masking as you read is actually the slowest, because you have to re-decide what to cover on every screen.
Q: Text records in tickets need de-identifying too. Is there a shortcut?
Do not edit plain text by hand; edit ten copies and you will miss one. A colleague now uses Privacy Masking, which recognises phone numbers, email addresses, ID numbers and bank cards automatically and outputs forms like 138****1234 that keep the first and last digits. Far faster than copying and pasting.
Q: What if I am worried a case image will be lifted by a competitor?
Masking and anti-theft are two different jobs. Once the private details are covered, add a layer of Image Watermark; the tiled version is harder to crop out than a single corner mark.
Q: Do complaint screenshots go to a server?
No, this kind of work happens locally in the browser and the image never leaves the machine. For complaint material that is a hard requirement; if anything had to go to a server, I would first have to ask the security team whether we were allowed to use it.
The training went ahead in the end, using the thirty-seven reworked images. The biggest gain from that round was not learning to mask images but a rule the team settled on: every screenshot goes through the checklist before it leaves, and whoever draws the box owns the result, rather than settling for looking blurred enough.