Guide
One sentence carries this page: an AI chat is the one destination where you need the model to read most of the image, so cover the values and keep the structure, sweep the edges of the frame, and remember that the paste cannot be taken back and the reply can say out loud what the box was hiding.
Why an AI chat is not just another place you send a picture
Four properties separate it from a post, an email or a support ticket, and each one changes what you should cover.
Close reading is the point. When a person glances at a screenshot in a thread, most of the frame is scenery. A model does not have scenery. The tab strip, the account name in the corner, the half-visible notification, the file path in the title bar and the other window behind yours are all content, read at the same resolution as the thing you were asking about. Anything in the capture is in play.
The exchange is written down. A chat is a transcript, and a transcript is a document that can be scrolled back, exported, shared with a colleague, pasted into a ticket or kept by the service. How long any of that lasts, and whether it is used for anything beyond answering you, depends on the particular product, your plan and settings you should check for yourself rather than assume in either direction. The part that does not depend on any of that: once the image is sent, sending it is not an event you can reverse.
It answers back. This is the property with no equivalent anywhere else. Every other destination is a place your image sits. This one generates new text about your image, text you did not write, and that text lands in the same transcript. A redaction that works on the picture does not constrain the prose.
Usefulness is a hard constraint. On a public post you can cover anything inessential and lose nothing. Here, covering the wrong thing covers the question: black out the error text and you have asked about an error you did not show. That is why “cover everything you possibly can” — sound advice on most of this site — is the wrong instinct in this one case.
The restatement channel: how a cover gets undone in words
A cover removes a value. It does not remove the fact that something was there, the size and position of the hole, or the surrounding layout that tells a reader what kind of thing used to fill it. People are mediocre at exploiting that. A model is good at it, because recognising a format from its shape and context is close to the centre of what it does. A sixteen-character gap in a familiar settings panel, a covered field directly under a label that says Routing, a blacked-out strip in a header where a workspace name belongs — all of these narrow the possibilities enormously, and some of them narrow it to one.
The practical consequences are small and specific. Cover generously: run the shape past the field edges so the hole does not publish the exact length of what it replaced. Do not leave a precisely sized void inside a recognisable interface when you can change the frame instead and not capture that panel at all. And when the model needs to know the kind of thing you hid, tell it in words — “this is a 40-character API key” — which is strictly safer than leaving a key-shaped hole for it to identify.
Then read the answer for restatement, which is a step almost everyone skips. If the reply names the provider, the institution, the person or the value you covered, your thread now contains that in text. Nothing has leaked further on its own, but the conversation you were about to forward or screenshot for a colleague is no longer the sanitised thing you thought you had. Decide at that point, not later.
Deciding what the model actually needs to see
Work through the capture in four passes. The order matters, because the last two cover the things you were not thinking about.
1. Keep the question. The error text, the failing line, the stuck control, the number that looks wrong. If you cover this, you have wasted the exchange. Keep it exactly as it appeared, including spelling and punctuation, because an error string is often matched verbatim.
2. Keep the structure. Layout, column headings, field labels, menu names, the shape of the interface. This is what makes an answer accurate rather than generic, and it is almost never sensitive on its own. A spreadsheet with its headings intact and its cell values covered is still a question a model can answer well.
3. Cover the values. Account and card numbers, routing and policy numbers, names, email addresses, postal addresses, phone numbers, keys, tokens, session cookies, order and invoice identifiers, balances, salaries, case or patient references, dates of birth. The test is not “is this secret” but “does the answer require it”, and the answer almost never requires the value.
4. Sweep the frame. Four edges and the window chrome, where nothing you came here for lives: the signed-in account name, the workspace or tenant in the address bar, the tab titles, bookmarks, the taskbar or dock, the clock, a calendar notification that arrived while you were capturing, and anything belonging to someone who is not you — a colleague in a call tile, another person's name in a message list, a customer's order behind the dialog. This pass finds more than the first three combined, and it is the one people run out of patience for.
The clipboard trap
This is the mistake most likely to undo everything above, and it looks exactly like success right up until the image renders in the chat.
The editor on this page listens for a paste anywhere on the document, takes the first item on the clipboard whose type begins with image/, and loads it straight onto the canvas. Capture, then paste here, is the fastest way in — no file, no dialog. What the page cannot do is put anything back. There is no copy-the-result button and no clipboard-writing code in it at all. The only exit is Download PNG, which writes the canvas to a file called hideshot-<timestamp>.png in your downloads folder.
Which means: after you paste a screenshot in here and redact it, your clipboard has not changed. It still holds the original, unredacted capture. Click into the chat box, press paste out of habit, and you send precisely the image you spent five minutes fixing. Your redacted version is sitting in the downloads folder, untouched and unsent.
The fix is to attach the downloaded file, or copy the downloaded PNG from your file manager — which does replace the image on the clipboard — and then paste. Either way, look at the attachment thumbnail in the chat before you press send. The black boxes are visible at thumbnail size, and that preview is the last honest check you get.
A running order that holds up
Decide the question first. Knowing what you are asking tells you what has to stay, which is the only way the keep-or-cover calls get easy.
Set the frame when you capture. Grab a window or a region rather than the whole screen. This editor has no crop, no resize and no rotate — every operation replaces pixels inside a shape you draw, and the export keeps your source's exact pixel dimensions — so if something should not be in the picture, the cheapest fix by far is to not capture it.
Paste or drop it in, then cover the values. Black Box for anything that must not be recoverable, Pixelate when you want to show that a field was there and roughly how much it held. Rectangle for rows and fields, oval for faces and avatars, lasso for anything on a slant.
Sweep the frame, then download. Undo steps back one cover at a time if you overshoot; Clear starts again from the original image.
Attach the file, check the thumbnail, send, then read the reply for restatement. And if the same screen comes up again later in the thread, redact the second capture the same way you redacted the first — two differently covered versions of one screen, sitting in the same conversation, can be lined up against each other.
What the editor above does and does not do here
It is a drawing surface, not an AI tool, and the distinction is the whole reason it is safe to use before the paste. Three modes — Black Box, Blur, Pixelate — each available as a rectangle, an oval or a freehand lasso. Every operation replaces the pixels inside the path you drew and nothing else; there is no whole-image operation of any kind. Black Box fills the path with flat #111111 and has no strength setting. Blur downsamples the selected area by a fixed factor of ten. Pixelate fills each block with the colour of that block's centre pixel, so hard edges survive instead of being averaged away. Undo restores stored pixel snapshots in draw order, and Clear redraws from the original image the page kept.
Two mechanical details worth knowing before you aim at small text. The canvas is created at your image's exact pixel dimensions and displayed scaled to fit, so what you see is usually smaller than what you are editing; and a rectangle drag under about four image pixels, or an oval under six, is discarded silently, so a very small cover on a large capture can simply fail to appear. Check that each box is actually there before you download.
What it does not do: there is no persistence of any kind here. No account, no saved presets, nothing written outside the open tab — the image, the undo snapshots and the retained original live in that tab's memory and go when it closes. And your image is never sent anywhere; the work happens inside the page in front of you, which is exactly why doing it before the paste, rather than describing your situation to the assistant and hoping, is worth the extra minute. Whether mild blur can be worked backwards at all is a separate question, and whether AI can unblur a photo covers it properly.
Common mistakes and misconceptions
“I will just tell it to ignore that part.” An instruction is not a redaction. Whatever was in the frame was sent. The text around the image changes what the model does with it, not what arrived.
“I deleted the message, so it is fine.” Deletion changes what you can see in the thread. Whether the turn persists anywhere else is a property of the service, not of the button, and it is not verifiable from inside the chat. Decide before the send.
“A blur is enough, it is only a machine reading it.” That is backwards. Mild blur is the one mode with a plausible route back, and the reader here is both working at full resolution and good at reconstructing plausible text from a coarse pattern.
“It is tiny on screen, nobody could read that.” The size of the bubble in the chat has nothing to do with the pixels in the file. Your export is the full size of your capture, and the model gets the file.
“I covered the value, so the field is handled.” The hole keeps its size and its position in a layout that may be recognisable. Cover past the edges, or change the frame.
“It is only my own data anyway.” Screenshots of shared tools carry other people by default — a name in a message list, a face in a call tile, a customer behind the dialog. They did not choose the paste.
“I pasted it into the editor, so the redacted one is on my clipboard.” It is not. The clipboard still holds the original. Attach the downloaded file.