Rinse · Made with AI labels
Almost always because of something written inside the file, not because of anything in the picture. Here is how to read your own file, how to strip it, and the part that no metadata tool can fix.
Written for people whose real work got labelled. Sources checked 2026-09-01.
The short answer
Image files carry text you never see. Platforms read that text on upload, and some of them turn it into a label under your post.
The main trigger is C2PA Content Credentials, a signed record of how a file was made, stored inside the image itself. Alongside it sit older fields: IPTC, XMP, EXIF. All of them are written by ordinary software doing ordinary things. Adobe apps write Content Credentials. AI generators write them. A growing list of cameras and phones now sign files by default.
So the label does not mean a platform looked at your photograph and decided it looked synthetic. In most cases it means a tool in your workflow left a note in the file, and the platform read the note.
Sometimes it is the pixels, though. Platforms also run their own classifiers over the image, which is a judgement about what the picture looks like. No metadata tool can change that, ours included. That caveat matters enough that it has its own section further down, and it is the reason this page does not promise you a label will disappear.
Two questions, then, and they have different answers: what is inside my file, and what does the platform think of my pixels. You can answer the first one yourself in about thirty seconds.
The evidence
Meta generates AI labels from C2PA, IPTC and XMP metadata written by ordinary tools, including Adobe's. The approach has attached the label to real photographs. The most cited case is a basketball photograph by Pete Souza, the former White House photographer. Meta's Oversight Board has issued a finding on the labelling approach.
Source: fstoppers.com, "Instagram's ham-fisted approach labeling photos AI" (2025).
TikTok uses C2PA Content Credentials to detect and label synthetic media automatically, whether or not the creator declared anything.
Source: storrito.com, "TikTok's 2026 AI labeling rules" (2026).
YouTube applies labels to video it detects as AI-generated. We have seen it reported that C2PA-derived labels there are permanent, but we have not verified that against YouTube's own documentation, so treat the permanence part as unconfirmed as of writing.
Source: storrito.com plus platform documentation, 2026 research pass.
The Leica M11-P, the Nikon Z9 and Z8, newer Canon and Sony bodies, Samsung Galaxy S25 AI edits, and Pixel 10 hardware-key signing all write provenance data. OpenAI attaches C2PA to its generations. Every month, more ordinary files carry credentials nobody chose to add.
Source: eyesift.com, C2PA adoption overview (2026).
These are the sources we hold, dated. Platform behaviour changes often, and Meta in particular has iterated on this repeatedly. Check them yourself before you rely on them. Anything we cannot source, we do not print.
Step one
Before changing anything, find out what is actually in there. The tool for this is
exiftool, by Phil Harvey. It is free, open source, and it is what
a lot of metadata utilities run underneath, ExifCleaner and Rinse included.
# install it (Homebrew), or get the package from exiftool.org
brew install exiftoolexiftool -a -G1 -s photo.jpg
-a shows duplicated tags instead of hiding them, -G1
prints the group each tag came from, -s uses the short tag names you
would write back. Expect more output than you think.
exiftool -G1 -s -a -EXIF:all -XMP:all -IPTC:all -JUMBF:all photo.jpgFour things are worth finding in that output:
[JUMBF] group, at all. JUMBF is the box that
holds C2PA Content Credentials. If it is there, your file is carrying a signed
provenance record.
DigitalSourceType, an IPTC Extension field carried
in the XMP packet. AI tools set it to a value ending in
trainedAlgorithmicMedia. It is about as explicit a declaration as a
file can make.
Software, CreatorTool, HistorySoftwareAgent,
which name the application that touched the file. Generative edit history lives
in the XMP packet.
parameters,
prompt and workflow. Stable Diffusion,
ComfyUI and Automatic1111 write the entire prompt and node graph into the file,
in plain readable text.
If none of those are in your file and the label still appeared, you are probably looking at a classifier decision rather than a metadata one, and nothing below will help. That is worth knowing before you spend an afternoon on it.
Step two, the free way
One command removes every writable tag exiftool can name:
exiftool -all= photo.jpgThat works. It also has sharp edges, and every one of these has bitten someone.
By default exiftool writes the cleaned file in place and keeps the untouched
original next to it as photo.jpg_original. Add
-overwrite_original and that safety net is gone. Copy your files
somewhere else first, and do not let a batch run be the only version you have.
-all= means all. Your copyright notice, your byline, and your EXIF
orientation flag all go, and a stripped orientation flag can leave a phone photo
displaying on its side. Name the tags you want to survive:
exiftool -all= -tagsfromfile @ -Orientation -IPTC:CopyrightNotice \
-overwrite_original photo.jpg
-tagsfromfile @ means "read those tags back from this same file",
so they are restored after the wipe.
This is the step people skip. Run the inspection command again on the output and
confirm the [JUMBF] group is actually gone. Metadata removal has
quiet failure modes: a container the tool does not fully parse, or an encoder
that writes a fresh tag back in. Reading the file afterwards is the only proof
that counts.
exiftool -all= -overwrite_original -ext jpg -ext png -r ./folder
-r descends into every subdirectory below that path. Point it at the
wrong folder and it will happily clean your entire photo library.
Editing a file on your Mac does nothing to a post you have already published. If the label is on something already up, you are deleting it and uploading the cleaned file again.
If you are comfortable in a terminal, stop here. exiftool is free and it is genuinely the right tool. ExifCleaner is a free graphical option too. The rest of this page is for people who would rather not do this by hand every time.
Step two, the $9 way
Rinse is a small Mac utility that does the above on every image you drop into a
folder. A Clean folder sits on your Desktop. Drag one image in, or fifty. Cleaned
copies appear in the Cleaned folder beside it, usually in about two seconds, and
your untouched originals are kept in ~/Rinse/originals/ with a
timestamp on each one.
It shows you what it found. Rinse counts the metadata fields in your original before it cleans, then tells you what it removed ("photo.jpg • 7 tags removed"), records it in its log, and flags any privacy or AI field that survived the pass. That receipt is the point. You are not asked to take the removal on trust, and you can check it with the exiftool command above whenever you want to.
Or read the full product page first.
Rinse re-encodes the pixels into a fresh container as its first pass, at quality 0.92. In normal use the result is visually identical to the original, same dimensions, same colour, and your original is always kept. It is not a byte-for-byte copy of the pixels, and we do not claim it is. If bit-exact pixels matter to you, exiftool leaves them completely untouched, so use exiftool.
Rinse is also not code-signed yet. Our Apple Developer enrolment was submitted in September 2026 and is being verified as of writing, so on macOS 15 Sequoia and later the first launch means dismissing a warning, then clicking "Open Anyway" in System Settings under Privacy & Security. Once, then never again. We would rather tell you now than at the moment you have just paid.
Where this stops
Nobody can promise you that a label will not appear. Anyone who does is selling you something they cannot deliver.
Platforms increasingly pair metadata reading with their own classifier-based detection, TikTok and Pinterest among them. Stripping metadata removes the metadata trigger. It does nothing about a model looking at your image and forming a view. Whatever a platform's classifier decides is beyond the reach of any metadata tool, Rinse included.
Google's SynthID and similar pixel-domain watermarks are woven into the image itself, not into its metadata. Rinse does not remove SynthID and cannot. Neither can exiftool, or any other metadata tool. Only diffusion-based regeneration can weaken a watermark like that, and even then never with certainty. If a page tells you otherwise, you can check it in five minutes and find out that it is false.
It is not built for that and it does not achieve it. Rinse exists because photographers, agencies and studios keep getting real work flagged over fields that Adobe, a camera, or a retouching step wrote without asking. That is the problem it solves, and where it stops is written on the box.
Related questions
Because the trigger is a metadata field, and ordinary tools write those fields. Adobe applications add Content Credentials. Cameras sign files now too (Leica, Nikon, newer Canon and Sony bodies), the Galaxy S25 marks its AI edits, and the Pixel 10 signs with a hardware key. A generative fill on one small corner can leave an edit history in the XMP packet covering the whole file.
It is a known failure mode, not your imagination. A basketball photograph by Pete Souza, the former White House photographer, was labelled this way, and Meta's Oversight Board has issued a finding on the approach.
No. Cleaning a file on your Mac has no effect on something already uploaded. You would need to delete the post and upload the cleaned file again. And even then, see the section above: metadata removal handles the metadata trigger, and a platform's pixel classifier is a separate question that no tool on your Mac can answer.
For most technical users, exiftool is enough. It is free, it is excellent, and the commands are on this page. ExifCleaner is a free graphical alternative. Rinse is $9 for a folder that cleans by itself, bundled binaries so there is nothing to install, C2PA handling, and the count of what it removed. If you already have exiftool aliased in your shell, you probably do not need Rinse, and we would rather say that than pretend otherwise.
The inspection step does. exiftool reads metadata from MP4 and MOV files, so
exiftool -a -G1 -s clip.mov will show you what is inside one.
Rinse itself handles images only in the version you can buy today: JPEG, PNG,
TIFF, HEIC, AVIF and static GIF.
They are what people actually see and search for, and the wording differs by platform and has changed over time. The mechanism underneath is the same one described on this page: provenance metadata read at upload, plus, on some platforms, a classifier looking at the image.
$9. One Mac. Local only. 7-day refund, no questions.
JPEG, PNG, TIFF, HEIC, AVIF and static GIF.