AI Image Labelling
Label AI images as required: burn a visible badge into the picture and embed a machine-readable note in the PNG or JPEG. All in your browser.
Caption to copy
Producing the image redraws it. Any capture data (EXIF) held by the original is lost in the process.
The image is processed entirely in your browser and is never uploaded.
Guide: AI Image Labelling
It gets harder every year to see that a picture was made by a machine. That is exactly why the European Union's AI Act requires labelling — and it requires it twice over: whoever runs a system that produces synthetic content must mark the output as artificially generated in a machine-readable form, and whoever publishes such a picture must disclose that to the people looking at it. The transparency duties in Article 50 apply from 2 August 2026.
This tool covers both halves. It burns a badge carrying a note such as “AI-generated” into the picture — you choose the corner, the size, the opacity and the style — and it additionally writes a machine-readable note into the file: a text chunk in a PNG, a comment segment in a JPEG. Both happen in the browser, with no upload and no sign-up.
Two things the tool states openly rather than glossing over. WebP cannot carry the machine-readable note; choose that format and you get the visible label only, and the island says so. And producing the image redraws it, which discards any capture data the original held — for a tool that adds metadata, that is a side effect worth knowing about.
Typical use cases
Product images that came out of an image generator
Anyone using generated mood shots or cut-outs in a shop labels them visibly and puts the note into the file as well, where platforms can read it.
Posts on social networks
Several platforms read metadata and add a notice of their own. A picture that already carries the note is more likely to be classified correctly than one that has to be guessed at.
Editorial illustrations on your own site
For a company blog the badge is the simplest honest answer: it sits in the picture and survives even when the picture is reused elsewhere.
Labelling an existing library
Older pictures can be run through one at a time. The settings persist between images, so a series gets the same badge in the same place.
Preparing a caption and alternative text
A ready-made sentence sits underneath the tool for copying. Disclosure in the text complements the badge wherever the picture appears without a caption.
How to use it
Choose an image
A PNG, JPEG or WebP from your own device. The file is read but never transmitted; the preview underneath shows the result at a reduced size, worked out with the same proportions as the final picture.
Set the text and the corner
Pick one of the suggestions or write a note of your own. The corner should sit where little is happening in the picture — a badge across a face reads badly and is the first thing sacrificed when the image is cropped.
Adjust size, opacity and style
The size is a share of the image width, so the badge carries the same weight on a large photograph as on a small one. The style swaps the panel and the lettering; pick whichever stands out clearly against the background at that spot.
Choose a format and embed the note
PNG keeps the image quality and carries the machine-readable note, JPEG does too and is smaller. WebP is the smallest but cannot take the note — the tick box then has no effect, and the line underneath says so.
Produce it and download
The finished picture appears below the tool, together with a statement of whether the note made it into the file. Look at the result once at full size before you publish it.
What happens to your data
Your image never leaves the device. It is read onto a canvas in the browser, given the badge, and reassembled into a file right there; this tool has no server it could send anything to. With pictures that is a different order of magnitude from a snippet of text: a photograph often carries more incidental information than the sender assumes — where it was taken, on what, and when. And it is precisely that capture data the redraw discards, which the tool also points out.
Frequently asked questions
From when do I have to label AI images?
The transparency duties in Article 50 of the AI Act apply from 2 August 2026. Independently of that, labelling obligations can already follow today from competition law or from the rules of individual platforms — so labelling earlier is not wasted effort.
Is the visible badge on its own enough?
For disclosure to the people looking at the picture, a clearly recognisable notice is the core of it. Alongside that, however, the regulation expressly requires the output to be marked in a machine-readable form. That is why this tool also writes a note into the file, and why it says so when the format does not allow it.
Why does WebP get no note in the file?
Because doing it cleanly means more machinery than this tool should carry: PNG has a text chunk and JPEG a comment segment, both lean structures that everything reads. WebP would need an XMP block written into the container. Rather than do that half-way, the island says the note is missing.
Is the original EXIF data kept?
No. The image is redrawn onto a canvas, and capture data such as the camera, the time and the location is lost in the process. If you need those details, keep the original. For a purely synthetic picture the loss means nothing — it never had any.
Can somebody remove the label again?
The note in the file can be deleted with a metadata tool, and the badge can be cropped off or painted over. Tamper-proof provenance is a different thing and needs cryptographically signed data. For disclosure to your own audience, this labelling is nonetheless what is being asked for.