Rinse · What triggers the AI label
Canva's background remover, Apple Photos, Lightroom, generative fill. What is actually reported, what nobody outside Meta knows, and how to see for yourself what your editing app wrote into your file.
Written 2026-09-08. Every source is named and dated below, including the one that argues against us.
The short answer
That is the honest headline, and everything below it is either a named report or a command you can run yourself. If a page tells you with confidence that tool X causes the label, ask them how they know.
Here is what is on the record. On 4 September 2026, The Verge reported that Instagram is applying its "AI Content" label to ordinary photographs that involved no generative AI at all. The reported triggers are small, normal edits made in mainstream tools. Meta has not published an explanation, a fix, or a rollback, and has not disclosed what its labelling actually scans for. As of the date on this page, the issue is unresolved.
Source: reporting by The Verge on 4 September 2026, credited to Jess Weatherbed. A straight word about our own sourcing, because this page asks you to trust it. We have not read that article directly. The Verge blocks the tools we use, so everything above comes from several independent write-ups published between 4 and 8 September 2026 that all credit that piece and agree with each other on the details. We would rather tell you that than imply we read something we did not.
So this page does three things. It lists the tools that have actually been named, and marks clearly which of those we have tested (none of them) and which are reported (all of them). It gives you the commands to find out what your own editor is writing into your files, which is the part you can genuinely settle in two minutes. And it tells you where all of this stops.
Already labelled and looking for what to do about the live post? That is a different job and it has its own page: Instagram put an AI label on your real photo. This page is about which edit did it.
Read this before anything else
In the same coverage, a writer reported the opposite experiment: they posted a fully AI-generated image, carrying both C2PA Content Credentials and Google's SynthID watermark, and Instagram applied no label to it at all.
Real photographs with tiny edits were being flagged. A synthetic image that declared itself twice over was not. Whatever Instagram is keying on in September 2026, it is demonstrably not a simple reading of "this file says it is AI".
Source: dailyguardian.ca, 5 September 2026, described by the author as their own test. One test, one account, not reproduced by us.
We sell a $9 tool that removes provenance metadata from images, so this is the inconvenient fact on our own page and we would rather print it than have you find it somewhere else. We cannot tell you that removing metadata prevents the Instagram AI label, or removes one you already have, and we are not going to imply it. The evidence above points away from that, and no one outside Meta is in a position to promise you otherwise.
What we can tell you is narrower and true: you can find out exactly what your editing tool wrote into your file, and you can decide whether you want to hand that over. The rest of this page is that.
What has actually been named
Two named triggers have come out of the September 2026 reporting. Everything else circulating is people reasoning backwards from their own flagged post, which is understandable and is not evidence.
Reported as a trigger, including for edits as small as removing a single speckle. That detail is the one worth sitting with: the tool is assistive, not generative, it is not making up new subject matter, and it is reported to have been enough.
The named case is About Face, the cosmetics brand founded by Halsey. Their photographs were shot on an iPhone and edited only in Apple Photos, and came back carrying an AI Content label. No generative tool was involved at any point in that chain.
Both of the above: The Verge, 4 September 2026. Status re-checked 2026-09-08: still unresolved, with no fix and no rollback announced.
Now the honest version of the table people actually want, which is the one with a column for "have you tested this".
| Tool or edit | Status | What we can say |
|---|---|---|
| Canva Background Remover | Named in reporting | Reported trigger, including for removing a speckle. Not tested by us. |
| Apple Photos edits | Named in reporting | The About Face photographs were iPhone images edited only here. Not tested by us. |
| Photoshop generative fill | Not named in this wave | Adobe's generative features attach Content Credentials, so the file carries a signed record of the edit. Whether Instagram acts on it, we do not know. |
| Lightroom AI denoise, AI masking | Not named in this wave | Widely suspected in public, with nothing behind it that we can source. Check your own export with the commands below. |
| Spot healing, blemish removal | Adjacent to reporting | Removing a single speckle is reported as enough in Canva. Whether the same holds in another app is unknown. |
| A plain camera or phone export | Not a trigger by itself | But newer cameras and phones write provenance data of their own, so "unedited" does not mean "carrying nothing". Read the file and see. |
"Not tested by us" means exactly that. We have a bench for testing our own software and we do not have an Instagram testing account, so we are not going to report a result we do not have. Anyone who does have that result should publish it, because it does not appear to exist yet.
This is a mechanism, not a claim about Meta. Modern editors record what they did.
Adobe applications write C2PA Content Credentials. Editors of all kinds write
Software, CreatorTool and HistorySoftwareAgent
naming themselves, and XMP edit history describing the operations. C2PA records
an ActionsDigitalSourceType drawn from a fixed IPTC vocabulary, and
the value that names generative AI ends in trainedAlgorithmicMedia.
None of those fields carries a notion of "only a bit". An edit history covering a one-inch corner and an edit history covering the entire frame look structurally the same to anything reading the file. That is the shape of the problem, and it is why the honest instruction is to go and look at your own file rather than to trust a list of app names, ours included.
The general version of all this, covering what C2PA Content Credentials actually are and how TikTok and YouTube handle them, is on why your photo says "Made with AI", and the field-by-field removal detail is on how to remove C2PA Content Credentials.
The part you can actually settle
You cannot see inside Meta's classifier. You can see inside your own file, and
that is the half of this you control completely. The tool is
exiftool by Phil Harvey, free and open source, and it is what most
metadata utilities run underneath, ours included.
# install it (Homebrew), or get the package from exiftool.org
brew install exiftoolexiftool -G1 -s -a -JUMBF:all -XMP:all -IPTC:all -EXIF:Software post.jpgRead the export, not the master. The file Instagram sees is the one that reaches your phone: the JPEG or HEIC you AirDropped, synced through Photos, or pulled out of a shared album. Inspecting your catalogue tells you nothing about it.
This is the command most people are looking for and rarely find. If your file carries a C2PA record, that record names the software that made each edit, in plain text, and you can read it out directly:
exiftool -G1 -s -a -ActionsAction -ActionsSoftwareAgentName \
-ActionsDigitalSourceType -DigitalSourceType \
-Software -CreatorTool -HistorySoftwareAgent post.jpgRun on a real file carrying Content Credentials, that prints something like:
[CBOR] ActionsAction : c2pa.created, c2pa.converted
[CBOR] ActionsSoftwareAgentName : gpt-image
[CBOR] ActionsDigitalSourceType : http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia
Real output, run on our bench on 2026-09-08 with exiftool 13.55 against a JPEG
carrying a genuine C2PA manifest. Your file will name your tool instead. If those
lines come back empty, your file is not carrying a C2PA record at all and the
Software and CreatorTool fields are the ones to read.
Nobody can tell you what a given version of a given app writes on a given day, because it changes. You can find out in about a minute, and the answer is certain, because it is your own file:
# 1. copy an untouched original, twice
cp original.jpg before.jpg
cp original.jpg to-edit.jpg
# 2. open to-edit.jpg in the app, make ONE edit, export it as after.jpg
# 3. read both files completely, then compare
exiftool -G1 -s -a -All before.jpg > before.txt
exiftool -G1 -s -a -All after.jpg > after.txt
diff before.txt after.txtEvery line the diff shows is something that edit added, changed or removed. Run the background remover on its own and diff. Run a spot heal on its own and diff. That is how you find out which step in your workflow is writing what, rather than guessing at a list of app names.
Note the field names that show up, because those are yours to check from now on:
a [JUMBF] group means a C2PA record is present,
DigitalSourceType is the file describing what kind of thing it is,
and Software, CreatorTool and
HistorySoftwareAgent are the file naming your application.
exiftool -r -G1 -s -a -ActionsSoftwareAgentName -ActionsDigitalSourceType \
-DigitalSourceType -Software -CreatorTool \
-ext jpg -ext jpeg -ext heic -ext png ./export
Walks ./export and prints, per file, only the fields worth knowing
about. Files that come back with nothing under them are carrying nothing on this
measure. -r descends into subfolders, so point it at the right one.
If you want it gone
One command removes every writable tag exiftool can name:
exiftool -all= post.jpg
That works, and three things about it have caught people out. -all=
means all, so your copyright notice and byline go with it, and so does the EXIF
orientation flag, which can leave a phone photo displaying on its side. Name what
should survive:
exiftool -all= -tagsfromfile @ -Orientation -IPTC:CopyrightNotice \
-overwrite_original post.jpg
Second, -overwrite_original throws away the safety copy exiftool
would otherwise leave beside the file as post.jpg_original. Work on
copies. Third, and this is the step everyone skips: read the output file
afterwards and confirm. Removal has quiet failure modes, including a container the
tool does not fully parse and an encoder that writes a fresh tag straight back in.
# confirm, with a parser
exiftool -G1 -s -a -JUMBF:all -XMP:all post.jpg
# and again with a raw byte scan, which trusts no parser at all
LC_ALL=C grep -a -c 'c2pa' post.jpg
The grep prints 0 when the string is not present
anywhere in the file. Swap c2pa for jumb, for your
camera model, or for a place name, and check those too.
Comfortable in a terminal? Stop here. exiftool is free and it is genuinely the right tool for this, and ExifCleaner is a free graphical option. Our comparison of all three is on ExifCleaner alternatives, and it includes the part where we tell you to keep the free one.
If it already happened
Nothing you do to a file on your Mac reaches a copy that is already on Instagram's servers. They are two different files, and there is no edit that reaches back.
Users in this wave report being unable to get a false label removed after publication. We do not know how the label is applied internally and we are not going to guess. What we can say is this: anyone selling you a "remove the label from your post" button is selling you something they do not have.
We do not have a sourced, current description of Instagram's appeal route, so we are not going to invent one and send you round a loop. Check Meta's own help pages for the current process, and read the fuller version of this on Instagram put an AI label on your real photo, which covers the live-post problem properly. This page stays on the question of which edit did it.
The $9 way
Rinse is a small Mac utility with no window. Installing it puts two folders on
your Desktop, Clean and Cleaned. Drag a file or
a whole export into Clean and a stripped copy appears in Cleaned, usually in about
two seconds. Your untouched original is copied to
~/Rinse/originals/ with a timestamp first, every time, and a macOS
notification tells you how many tags came out ("photo.jpg · 7 tags
removed"). Everything runs on your Mac. There is no network call anywhere in it.
That is the whole product, and the reason to buy it is narrow: exiftool makes cleaning a task you have to remember before every delivery, and a watched folder makes it a step that happens whether you remember or not. If you clean a few files a month, the free tool is the right answer and we would rather you kept it.
What it does not do is promise you an outcome on Instagram. It removes the provenance metadata from your file, and it shows you the count so you are not asked to take that on trust. What a platform's classifier decides about your image afterwards is that platform's opaque decision, and the evidence at the top of this page is the reason we say so plainly.
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
Cleaning a file changes what the file says. It does not buy you a decision from Instagram, and nobody selling you software is in a position to promise one.
Rinse does not prevent, remove or avoid Instagram's AI label, and we do not claim it does. The reporting this page is built on shows real photographs with tiny edits being flagged while a fully AI-generated image carrying both C2PA and SynthID went unlabelled. That is not the behaviour of a system reading metadata and acting on it, so a metadata tool cannot be sold as the answer to it. If we ever run a controlled test of our own, we will publish the result whichever way it falls.
Platforms pair metadata reading with their own classifier-based detection. Removing metadata removes the metadata. It does nothing about a model looking at your image and forming a view, and whatever that model decides is beyond the reach of any metadata tool, ours 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 it is false.
Meta has iterated on this repeatedly, and the reporting notes this is the second such wave after one in 2024. Everything here is dated on purpose. Check it yourself rather than taking it from us in six months, when it may no longer hold. Anything we cannot source, we do not print.
Rinse is not built for that and does not achieve it. It exists because photographers, agencies and studios keep getting real work flagged over fields that an editor, 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
It is named as a trigger in The Verge's reporting of 4 September 2026,
including for edits as small as removing a speckle. That is a report, not a
proof, and we have not tested it ourselves. What you can settle in a minute is
what the tool wrote into your own file: copy an original, run one background
removal, then diff a full exiftool -G1 -s -a -All
read of the before and after files. Every line in that diff is something the
edit changed.
It has happened. The cosmetics brand About Face, founded by Halsey, had photographs shot on an iPhone and edited only in Apple Photos come back carrying an AI Content label, reported by The Verge on 4 September 2026. No generative tool was in that chain. We have not reproduced it and we do not know what specifically was read.
Neither is named in the September 2026 reporting, so anyone telling you yes is
guessing. What is true and checkable is that Adobe's generative features
attach C2PA Content Credentials to the export, which is a signed record of
what touched the file. Read your own export with
exiftool -G1 -s -a -JUMBF:all -ActionsSoftwareAgentName post.jpg
and you will see whether yours is carrying one.
Diff the file against itself. Copy an untouched original, make exactly one
edit, export it, then run
exiftool -G1 -s -a -All over both files into two text files and
diff them. Repeat one edit at a time. This is the only method
that gives you a certain answer, because it tests your app, your version, your
settings, rather than a list of app names on somebody's website.
We cannot promise that, and we will not imply it. In the same period that real photographs with small edits were being flagged, a writer reported posting a fully AI-generated image carrying both C2PA and SynthID and getting no label at all. Meta has not disclosed what it scans for. Removing metadata removes the metadata, which is a real and checkable thing; it does not buy a decision from a platform.
Not with any local tool. Nothing you run on your Mac reaches a file already sitting on Instagram's servers; they are different files. Users in this wave report being unable to get a false label removed after publication. The fuller version of this is on Instagram put an AI label on your real photo.
Unknown. As of 8 September 2026 Meta had published no explanation, no fix and no rollback. The reporting also notes this is the second such wave, after one in 2024.
No, and it cannot. SynthID and similar pixel-domain watermarks live in the pixels, not in metadata. No metadata tool removes them, ours included. Anything claiming otherwise is claiming something you can disprove in an afternoon.
$9. One Mac. Local only. 7-day refund, no questions.
JPEG, PNG, TIFF, HEIC, AVIF and static GIF.