Why is my real photo labelled as AI?

A real photo gets an AI label when the file carries a signal saying AI was involved, and some editing tools write that signal even for a small generative edit. Adobe’s generative tools and Google Photos edits such as Magic Eraser both attach AI credentials to the file, and platforms such as Meta say they read industry-standard signals in the file when deciding on a label. You can check your file for those signals before you post, and remove the metadata ones from your copy with removal confirmed by a re-scan. That only changes future uploads. It does not touch a label on a post that is already live, and it cannot touch invisible watermarks, whatever visual classifiers a platform runs, or a platform’s own disclosure rules.

Our prediction: medium confidence. Take a real photo whose only AI signal is the credential an editing tool wrote, such as a Generative Fill or Magic Eraser edit. We expect a copy cleaned with NoWatermark using Everything removable, with the re-scan showing Content Credentials, XMP and IPTC as removed, not to get the AI label that metadata triggers. The confidence is medium overall because platforms differ. LinkedIn and Meta describe labels driven by metadata, while Pinterest also runs classifiers and X has not said what it reads. For an unedited camera original, don’t clean it at all: see the camera label section below. This is a prediction from platforms’ and vendors’ own documentation, not yet our own test result. We are running that test and will publish the result on this page.

Try it: check and clean your image before you post. It runs in your browser, and your file is never uploaded to us.

What triggers the label on Meta

Meta’s Help Centre says some content made or modified with AI tools “may have an industry-standard signal”, and that “these signals are read by Meta’s systems to determine if a label is applied.”

Meta has also acknowledged that real photos were caught by this. In its July 2024 update, Meta said some content with minor modifications using AI, such as retouching tools, carried industry-standard indicators and was labelled “Made with AI”, so it renamed the label “AI info”.

Two dated reports show what photographers saw:

• PetaPixel, June 2024. Removing a speck of dust with Photoshop’s Generative Fill got the photo a “Made with AI” tag on Instagram. Neural Filters, Sky Replacement, AI-powered noise reduction and Super Resolution did not trigger it in PetaPixel’s tests. PetaPixel linked the tag to the C2PA information in the file; that is its inference from the file’s contents, not a Meta statement. The test predates Meta’s September 2024 change and Adobe’s August 2026 change described below, so do not read it as a current list of safe tools.
• Adobe Community, June 2024. A photographer reported that Instagram had begun labelling images they had used Generative Fill on, based on the metadata Photoshop had added. A moderator in the thread noted that Meta adds the tag, not Photoshop.

Which edits write AI credentials

Adobe. Adobe’s labelling FAQ (last updated 25 August 2026) says that when qualifying generative AI tools are used to create or modify content in Adobe apps, C2PA metadata is automatically attached to the file. Asked whether it can be turned off, Adobe answers: “No. Under the new laws, this metadata cannot be disabled for qualifying content.” Adobe says this applies to generative AI tools, not to traditional AI-powered features, and that it is rolling the support out gradually across Creative Cloud, Firefly and other products through August 2026. The metadata may record the app, the generative model, whether the content was wholly generated or modified, and the date and time.

An older Photoshop export help page (February 2026) lists a Content Credentials option called “None”. Adobe’s newer FAQ is the one that governs generative edits, so do not assume you can opt out. Adobe’s pages do not list which products or versions are covered, so check your own file.

Google Photos. Google’s help page says that if a photo has no Content Credentials, Google Photos adds them when an AI edit is saved, and that photos edited with generative AI such as Magic Eraser, where new pixels are created, show “Edited with AI tools”. A 2024 Google post said edits made with Magic Editor, Magic Eraser and Zoom Enhance already carried IPTC metadata marking generative AI editing, and a September 2025 post says Photos attaches credentials to images edited with AI tools. Content Credentials are not available on Google Photos Web. Google’s pages do not say what Magic Eraser writes on non-Pixel Android or on iOS, so we do not either.

Samsung. Samsung said in February 2025 that the Galaxy S25 series supports C2PA Content Credentials for content created and edited with generative AI. Its release does not say which editing features write them, so we cannot tell you which Galaxy edits will.

Deleting the Generative Fill layer: what users report

Users on Adobe’s community forums report that undoing or deleting the AI layer does not clear the credential.

• In October 2023 a user wrote that after using Generative AI on a file and then undoing or deleting the AI layer, the Content Credential saying the work was generated with AI was not removed.
• In an August 2024 bug report (Photoshop 25.11.0, Windows) another user said credentials stating “at least one was generated with an AI tool” were still added to exported files after removing all the AI-generated content, even after closing and reopening the document. The same report names a workaround: copy the layers into a new document.

These are user reports. We found no Adobe statement confirming or fixing the behaviour, Adobe’s 2026 FAQ does not address it, and we do not know whether it still happens after the August 2026 change. We have not reproduced it. The safe habit is to scan the exported file, not the layer stack.

Edited or generated: where the label shows

Meta describes two placements. Since September 2024, content it detects as only modified or edited by AI has the “AI info” label in the post’s menu instead of on the post, while content detected as generated by an AI tool still shows the label on the post (TechCrunch reported the same change on 12 September 2024). Meta’s current Help Centre repeats it: content with signals indicating it was created with AI is labelled, and content with signals indicating it was modified “will not have the AI info label directly on the post”; the details are under Options. Instagram’s own help adds that information about AI used to create or modify content can be found in the Options menu whatever the type of signal.

Two limits. Meta has not said which signal value decides between “created” and “modified”. Adobe’s metadata can record wholly generated versus modified, and IPTC has separate values for each, but no source ties Meta’s placement to either. And NoWatermark cannot help you predict it: our scanner does not read whether a C2PA credential says generated or edited. Meta also says its labelling methodology “is still evolving”.

Check your file before you post

Run the file through the C2PA Checker or the AI Watermark Checker. The scanner runs on your device and finds C2PA manifests (with the manifest size and claim generator), XMP, IPTC and AI generator metadata. For an XMP DigitalSourceType it shows “Created by a generative AI model” or “Composite including AI-generated elements”.

Its limits are stated plainly. It does not read the DigitalSourceType inside a C2PA manifest. It reports a manifest as detected, never as valid. And SynthID is always shown as Unable to verify: the scanner cannot detect it and cannot say it is absent.

What cleaning changes, verified by re-scan

The Remove AI Label Metadata tool scans the file, removes what you choose, and scans the result again. A signal is reported as removed only when that second scan cannot find it. The image data is copied byte for byte, not re-encoded. Three presets:

• Everything removable: EXIF, IPTC, XMP, Content Credentials and embedded text. By default one EXIF rotation field is kept so the photo stays upright.
• Privacy-safe: location and device data. Provenance is kept.
• Provenance-light: Content Credentials, XMP and embedded text. EXIF and IPTC are kept.

The report lists each signal as removed, kept or remaining. Read it before you post.

The trade-off: the camera capture label

Instagram now has a positive label. Its help page says “Captured with a camera” confirms a photo originated from a camera, using secure metadata that certain cameras attach at capture, and that if this information is “missing, incomplete or unverified” the label is not added. Google’s September 2025 post says the Pixel 10 camera supports C2PA Content Credentials.

So stripping Content Credentials from an unedited camera original can cost you a label you might want. And keeping the credential is not quite enough. Under the C2PA specification, a credential’s hard binding is a hash over the file’s bytes, and the only bytes it may leave out are the manifest itself and padding. EXIF and IPTC are inside that hash, so any cleaning at all, including Privacy-safe, which keeps the manifest and removes EXIF and IPTC, changes bytes the credential vouches for. The credential then no longer matches the file, and Instagram says it will not add the label when the information is “unverified”. If you want the camera label, upload the untouched original. That also means you cannot strip location data and keep a verifiable capture credential at the same time: pick one. Instagram’s page also says a photo created or edited with AI will not get the camera label and may get an AI label instead. We have not tested how Instagram treats a cleaned file.

What cleaning cannot change

• Invisible watermarks. Unable to verify, always. See can you remove SynthID?.
• Classifiers. Meta said in February 2024 it was working on classifiers to detect AI images that lack markers. Its current pages do not confirm they are live for labels, so we cannot say whether one is in play.
• Live posts. Cleaning only changes what you upload next.
• Disclosure rules. These are the platform’s, not the file’s.

Adobe’s stated reason, and what this page is not

Adobe ties automatic attachment to “the California AI Transparency Act and the EU AI Act”. That is Adobe’s statement. This page is not legal advice, and we make no claim about whether removing such metadata is lawful or unlawful in your situation. Read the platform’s terms and any rules that apply to you first.

Meta’s disclosure rules

Instagram’s help says Meta does not require you to label images created or altered with AI, though images still receive a label if Meta’s systems detect they were AI-generated. Meta’s Help Centre says images may also be labelled if they were modified using AI. Disclosure rules differ for photorealistic video and realistic audio; see Instagram, Facebook and Threads: AI info for the wording and the discrepancy between Meta’s pages.

If the post is already labelled

• Adobe’s FAQ says that for content showing an unexpected label you should visit the platform’s own support resources.
• On Instagram, the toggle Instagram’s help describes is used when creating a post. We did not find an official page describing how to change a label on an existing post, or any appeal process for a label Meta applied. If you find one, tell us.
• A cleaned file is a new upload, and we have not tested what any platform does with one. Removing the file’s metadata does not change an existing post.
• Other platforms differ. Pinterest’s help says to contact support to appeal. TikTok’s help says an automatic label cannot be removed from the post. Roboin reported in March 2026 that an automatic label on X was greyed out for one post it tested. We found no documented removal route for LinkedIn. See the guides for Pinterest, TikTok, X and LinkedIn.

Try it on your own image

Drop your image into Remove AI Label Metadata. It lists every signal the file carries, removes the ones you choose, and scans the cleaned copy again so you can see what is gone and what remains, including anything it can only mark Unable to verify. It all runs in your browser: your file never leaves your device. Scan the exported file, not the Photoshop document: the export is what the platform sees.