4 November 2026 · 6 min read
Before You Paste a Screenshot Into ChatGPT or Claude: What to Hide
Screenshots show more than you think. What to hide before pasting one into an AI chatbot, why a black box beats blur for text, and which settings to check.
Before you paste a screenshot into ChatGPT, Claude or any other AI assistant, cover anything the assistant does not need to answer your question: names, email addresses, phone numbers, card and ID numbers, passwords and tokens, faces, home addresses, and the browser tabs or URLs around the edge. Use a solid black box for text, not blur, and save a new copy of the image. Then check that service's own data settings, because the rules differ between products and plans. It takes about a minute and the assistant can almost always still help.
Why is a screenshot riskier than typing your question?
When you type a question, you choose every word. A screenshot brings along everything else that happened to be on the screen: the notification that slid in, the customer's name in the sidebar, the account number in the corner, the other tabs you had open. Modern assistants can read text in images, so assume that anything a person could read in the screenshot can be read by the model too, even if it is small or off to one side.
This is not a claim about what any particular company does with your data. That depends on the service, your plan and your settings. It is simply that you cannot un-share something, so it makes sense to share only what the task needs.
What should you hide in a screenshot before sharing it with AI?
| Look for | Where it tends to hide | What to do |
|---|---|---|
| Names and usernames | Chat headers, profile pictures, 'signed in as', document titles | Black box, or replace with a role when you describe it ('the client') |
| Email addresses and phone numbers | Message threads, invoices, CC lines, signatures | Black box |
| Card, bank and ID numbers | Receipts, statements, forms, photos of documents | Black box, every digit. Partial numbers can still identify you |
| Passwords, API keys, tokens | Terminal windows, .env files, settings pages, URLs with 'token=' in them | Black box, and change the secret if it has already been shared |
| Faces | Video calls, team photos, profile pictures, family photos | Black box, or crop them out |
| Home and delivery addresses | Order pages, maps, shipping labels, delivery apps | Black box |
| Tabs, bookmarks and URLs | The top of the browser window | Crop to the relevant part of the page |
| Notifications and sidebars | Corners of the screen, message previews, calendar pop-ups | Crop, or retake the screenshot with them closed |
Other people's details deserve extra care. A screenshot of a group chat, a client's invoice or a pupil's work contains information that was not yours to decide about. At work, your employer may also have a policy on which AI tools can see what, so check before you paste.
Is a black box better than blur for hiding text?
Yes. A black box replaces the pixels, so there is nothing left underneath to recover. Blur and pixelation only scramble them, and what remains can still carry information. A 2016 paper, 'Defeating Image Obfuscation with Deep Learning', showed neural networks being trained to recognise faces, objects and handwritten digits through pixelation and blur. How far that goes depends on the image, but a light blur over small text is not something to rely on.
| Method | What it does | Use it for |
|---|---|---|
| Black box | Destroys the pixels underneath | Passwords, tokens, card and ID numbers, names, addresses: anything that must stay hidden |
| Pixelate | Averages areas into blocks; may be partly reversible | Making a low-stakes background less distracting |
| Blur | Smears the pixels; may be partly reversible | Softening a low-stakes background. Not for sensitive text |
| Crop | Removes that part of the image from the exported copy | Tabs, sidebars, and anything at the edges |
Two common traps: a highlighter or semi-transparent shape can leave text readable if someone adjusts the contrast, and a file that keeps layers may let someone remove the box. Always export a flattened image and open the new copy to check it before you paste it.
How do you redact a screenshot in a minute?
- Take the screenshot of just the window or area you need. Close unrelated tabs and notifications first if you can.
- Crop away whatever does not help answer your question.
- Open Skrubly's Hide Parts of an Image tool. The image is processed in your browser and is not uploaded.
- Choose Black box and drag it over each sensitive area. Zoom in and scan the corners.
- Download the new copy. The saved file is re-encoded from the canvas, which also drops EXIF, GPS and camera data.
- Rename the file if its name says something private, like 'jane-smith-payslip.png'.
- Open the saved copy and read it once more before you paste it.
Black out names, numbers and faces before you paste a screenshot anywhere. Free, in your browser, nothing uploaded.
Open the redaction toolWhat about photos and documents with hidden metadata?
A screenshot usually carries little hidden data. A phone photo is different: it can include GPS coordinates, the time it was taken and the device model. Office documents and PDFs can hold author names and timestamps. If you are about to give an assistant a photo or a file, Skrubly's Metadata Cleaner can remove that hidden layer from images, video, PDF, Word and PowerPoint files in your browser. It does not touch what is visible on the page, so visible text still needs covering or editing separately.
Remove GPS, author names and other hidden data from a copy of your file.
Open the Metadata CleanerWhich AI data settings should you check?
Redacting controls what you send. The service's settings and terms control what happens next, and only the service can tell you that. Read its current privacy documentation rather than a summary from a blog, including this one. Menu names change, so look for sections called something like data controls or privacy, and find out:
- Whether your chats and uploads may be used to improve its models, and whether you can change that.
- How long chats and uploaded files are kept, and what happens when you delete them.
- Whether there is a temporary or incognito style of chat, and what it does and does not do.
- Whether people at the company may review conversations, for example for safety or feedback.
- Whether a memory feature carries details from one chat into the next.
- Whether your work or school account has different terms from a personal one. Free, paid, business and API products can differ.
Can you still get a useful answer after redacting?
Usually, yes. To explain an error message, the assistant needs the error, not your username. To reword an awkward email, it needs the text, not the recipient's address. To read a chart, it needs the chart, not the sidebar. If a redaction removes something that matters, describe it instead: 'the blacked-out part is a customer's name' or 'the total is around 400 euros'. For text-only problems, copying the relevant lines and replacing the names is often simpler than a screenshot at all.
Quick answers
- Does deleting the chat undo the upload? Do not assume so. Retention depends on the service, so check its documentation.
- I already pasted a password or token. What now? Change it straight away and sign out of other sessions. For an API key, revoke it and issue a new one.
- Is it fine to blur a face? For a casual photo with low stakes, perhaps. If it matters that the person cannot be recognised, use the black box.
- Can I just draw over it with my phone's markup pen? Use an opaque colour and go over it several times, then check the saved copy. Thin or see-through strokes can leave text legible. A solid box is more dependable.
- Does the Skrubly tool handle PDFs? The redaction tool works on images. For a PDF page, take a screenshot of the page and redact that copy.