If Instagram put an AI label on a photo you took yourself, look at the last app that saved it. Meta said as much when it renamed the label in July 2024: its system had been labelling content with “minor modifications using AI, such as retouching tools”, because those tools write a note into the file saying AI was involved.
What Meta says it reads
Meta’s February 2024 announcement names the sources. It reads the C2PA and IPTC metadata standards and labels images carrying those markers from Google, OpenAI, Microsoft, Adobe, Midjourney and Shutterstock. Pictures made with Meta AI get “Imagined with AI”, backed by visible markers, invisible watermarks and metadata. The same post says Meta was building classifiers to spot AI content that carries no markers at all, and admits that “there are ways that people can strip out invisible markers.”
Two updates followed. On 1 July 2024 “Made with AI” became “AI info”. On 12 September 2024, content that was only edited with AI tools had its label moved into the post’s menu, while fully generated content kept it on the post, and the label began saying whether it came from industry signals or from the poster’s own disclosure.
How a real photo ends up carrying the signal
The IPTC vocabulary that editing apps write into a file has separate codes for separate kinds of AI involvement. Three of them matter here, quoted from IPTC’s own list.
| Code | IPTC definition | Typical cause |
|---|---|---|
trainedAlgorithmicMedia | Digital media created algorithmically using an Artificial Intelligence model trained on captured content | A picture made from a text prompt |
compositeWithTrainedAlgorithmicMedia | Augmentation, correction or enhancement using a Generative AI model, such as with inpainting or outpainting operations | A real photo with one object removed or the canvas extended by a generative tool |
algorithmicallyEnhanced | Modification or correction by algorithm without changing the main content of the media, initiated or configured by a human, such as sharpening or applying noise reduction | Ordinary processing; no generative model involved |
A holiday photo with a stranger removed by Generative Fill is, in IPTC’s terms, a composite with AI media. Photographers had complained that the old “Made with AI” tag landed on real photos, and TechCrunch noted in July 2024 that photos with objects removed using Adobe’s Generative Fill “might still be tagged” under the new name. Those photographers saw a false positive. We read it differently: the file was telling the truth about the edit, and the label summed it up with too little context, which is what the September 2024 change tried to fix.
Meta has not published which codes send a label to the post and which send it to the menu, and we haven’t run posts through Instagram to find out. Nor do we know how often its classifiers put a label on a real photo that carries no metadata at all. Those two gaps are why nobody, us included, can tell you for certain what a given upload will get.
One more route in: a Content Credentials manifest. Editors that write C2PA record the actions taken on the image, and a generative edit is one of them. The Firefly and Photoshop guide covers when Adobe adds one.
Check the file before anything else
The label comes from something, and the file is the part you can see. Our scanner reads it on your device and reports any C2PA manifest (it names the current manifest’s claim generator, lists any earlier ones, and doesn’t validate the signature) with the declared actions, plus every metadata block it found. It reports that an XMP block is there without printing the IPTC source type inside it; for that line, exiftool -DigitalSourceType photo.jpg does the job. The public Content Credentials inspection site, contentcredentials.org/verify, reads the manifest too. More ways are in checking an image for Content Credentials.
An empty result is informative. If the file carries no C2PA and no IPTC source type, the label came from somewhere else: an invisible watermark in the pixels, Meta’s own detection, or a disclosure toggle switched on when the post was made.
What you can do about it
If the photo is yours and no AI tool touched it, but the file carries a stray tag from an app, cleaning that file removes the file-level signal. ExifTool does that for free and leaves the pixels alone. WipeTheAI strips the same metadata, but it needs a Google sign-in and both of its wipe modes also change the pixels. It does nothing about Meta’s classifiers, and nothing about which markers Meta decides to read next year. Cleaning a file changes what the file says. Instagram’s rules and its detection are Meta’s to change.
If you did edit with a generative tool, say so in the caption. Since September 2024 Meta has shown the label for edit-only content in the post menu rather than on the image, so an honest file costs less than it did.
How the label relates to invisible watermarks and to C2PA, three separate layers, is in C2PA, SynthID and platform labels compared.
Sources
- Meta: Labeling AI-generated images on Facebook, Instagram and Threads (6 February 2024)
- Meta: Our approach to labeling AI-generated content and manipulated media (updated 1 July and 12 September 2024)
- Instagram Help Center: label AI content on Instagram
- IPTC NewsCodes: digital source type
- TechCrunch: Meta changes its label from “Made with AI” to “AI info” (1 July 2024)