Why a platform labelled your photo as AI, and what in the file it reads
Meta says its systems read an industry-standard signal showing content was made or changed with AI, and that a photo editing tool using AI may leave one on an image. Adobe says its generative tools attach C2PA metadata when they create or change content. So a label can reflect a signal in the file or a poster's disclosure.
Why would a platform label a photo I edited myself?
Meta's Help Centre page "How to identify AI content on Meta products" says content made or changed with AI tools may carry an industry-standard signal, and that Meta's systems read these signals to decide whether to apply a label. Its example is an image edited in a common photo editing tool: if the tool uses AI to change the size or colour, Meta says, the image may carry a signal showing AI editing.
The page describes two outcomes.
- Signals that say the content was created with AI. The content is labelled "AI info".
- Signals that say it was modified with AI. The label is not placed directly on the post. Meta says people can click Options to see how it was modified.
The page adds that people may also be able to label their own content. It says Meta may require a label for realistic video or audio, not for images, but an image may still be labelled if Meta's systems detect that it was generated or modified with AI.
The page does not name the fields or standards Meta reads. In its February 6, 2024 post, Meta said it was building tools to identify the "AI generated" information in the C2PA and IPTC standards. We found no current list, so we do not claim to know what Meta reads today.
What does an editing tool write into the file?
Adobe's Generative AI content labeling FAQ (last updated August 25, 2026) says C2PA metadata is automatically attached when qualifying generative AI tools are used to create or modify content in Adobe apps. Adobe says this applies only to generative AI tools, not to traditional AI-powered features.
Adobe lists what the metadata may include: the application, the generative AI model, whether the content was wholly generated or modified, the date and time, and a unique identifier. It says this metadata cannot be turned off for qualifying content under the new laws, and that Adobe does not automatically add a visible label.
On other sites, Adobe says, each platform decides whether and how to show metadata, so labels can differ. It points people with an unexpected label to that platform's support resources.
What is the IPTC digital source type?
IPTC keeps a Digital Source Type vocabulary, a list of terms for how an image was made. A file can carry one of these terms in its metadata. Four of its terms bear on an edited photo:
| Term | What IPTC says it covers |
|---|---|
compositeWithTrainedAlgorithmicMedia | Named "Edited using Generative AI". Augmentation, correction or enhancement with a generative model, such as inpainting or outpainting. |
trainedAlgorithmicMedia | Named "Created using Generative AI". Media created by an AI model trained on captured content. |
algorithmicallyEnhanced | Named "Algorithmically-altered media". Changes by algorithm, started or set up by a person, that leave the main content as it was, such as sharpening or noise reduction. |
computationalCapture | Several frames from a real scene merged automatically, using signal processing or non-generative AI. IPTC says this includes the HDR processing common in phone cameras. |
The list does not say which app writes which term, or which term a platform acts on.
What did we find in real files?
A file can carry the AI note in a signed record or in unsigned metadata. We measured 31 real files filed as AI-made, all but the C2PA test image of our 32 real files (29 from Wikimedia Commons categories named "Images generated by ...", plus two vendor samples). The run was on October 30, 2026. Our accuracy page lists what the checker showed for each file; it does not list the IPTC terms below.
| Where it sat | Files | Terms |
|---|---|---|
| Inside a signed record | 8 | 7 trainedAlgorithmicMedia, 1 compositeWithTrainedAlgorithmicMedia |
| In unsigned XMP | 1 | 1 trainedAlgorithmicMedia |
| Neither | 22 | None |
In two of the eight signed files, older ChatGPT files, the term sits in the record of an earlier version inside the file. One of our three Adobe Firefly files declares compositeWithTrainedAlgorithmicMedia, the term IPTC names "Edited using Generative AI". The checker reads all three Firefly files as "Content Credentials found; legacy trust-list match". The Gemini PNG with the unsigned declaration reads "Metadata declares AI use".
Third parties uploaded the Commons files, and we cannot know how each was handled first. We sent no file through any platform, so this counts what files carry, not any platform's labels.
Why might a label appear when I did not mean to use AI?
These are causes the sources above allow. We cannot see which one applied to your post.
- A tool in your workflow used a generative feature. Adobe says qualifying generative tools attach C2PA metadata automatically.
- An earlier version carried it. Adobe's Content Credentials overview says new credentials can be added at each editing stage, building a version history.
- A disclosure. Meta says people may be able to label their own content.
- The platform's own systems. YouTube's help page says it may apply a label automatically for content with C2PA metadata, content made with its own tools, and content its internal systems detect as generated or altered with AI.
If a label looks wrong, use that platform's own route. YouTube says creators can usually change the disclosure, but not for its own tools, for content with C2PA metadata, or after manual review.
What this means when you check an image
Open the checker and read the headline. Three are relevant here.
- "Trusted Content Credentials declare AI use". A signed record is present, its signer is on the current C2PA Trust List, and the record says AI was used. It does not say how much of the image.
- "Content Credentials found; legacy trust-list match". A signed record is present and its signer is on the Interim Trust List (frozen January 1, 2026), not the current C2PA Trust List. Read the record's own statements as the signer's claims.
- "Metadata declares AI use". The file carries an unsigned label. Anyone can add or remove it.
If you see "No supported provenance signals found", we found none of the signals we read. That does not show what a platform read, and it does not mean the image is real or human-made. Every result carries our limits line: "These checks cannot determine whether an image is AI-generated." We do not check Meta's watermark.
Guides on this topic
- Visible AI watermarks is the main page on this topic.
- Meta's AI info label covers its triggers.
- IPTC digital source type values explained goes through the list.
- Where AI declarations live in an image file shows the places.
- What LinkedIn's Content Credentials label means covers another platform.
Frequently asked questions
- Does an AI info label mean the whole photo was made by AI?
- Meta's page does not say how much of a photo changed. It says signals of creation get the label on the post, while signals of modification are shown under Options.
- Does every AI-powered edit add a signal?
- Adobe says only generative tools, not traditional AI-powered features, attach it. We found no rule covering every app.
- Can your checker show what Meta read?
- No. It reports the signals it reads in the file you give it, and a platform may read more or less. Meta does not list its fields on the page we read.
- Is the label proof about my photo?
- No. A label can come from a signal in the file, a disclosure or, YouTube says, its own systems' detection. We cannot see which applied to your post. An unsigned label in a file can be added by anyone.