Edited with AI vs made with AI: how generative fill, inpainting and upscaling are labelled
By The NeutronField editorsPublished Updated
Short answer
An AI edit to a real photo is labelled in at least 3 places: the file's content credential, its IPTC source type, and the platform label. Adobe now attaches a content credential automatically when its generative tools modify an image, Google Photos marks photos "Edited with AI tools", and IPTC has a separate source type for edits. Meta moved its "AI info" label for edited images into the post menu in 2024.
Generated against edited
The industry vocabulary draws the line for you. IPTC, which maintains the standard photo metadata terms, uses "trainedAlgorithmicMedia" for images created with generative AI and "compositeWithTrainedAlgorithmicMedia" for "augmentation, correction or enhancement using a Generative AI model, such as with inpainting or outpainting operations" (IPTC Digital Source Type). A third term, "algorithmicallyEnhanced", covers changes such as sharpening or noise reduction that do not alter the main content.
What an edit records
| Who | What an AI edit writes | Visible label |
|---|---|---|
| Adobe apps | C2PA credential: modified with generative AI, app, model, date; cannot be disabled | None added by Adobe |
| Google Photos | IPTC metadata and a credential when a generative edit is saved | "Edited with AI tools" in the photo details |
| Samsung Galaxy AI | A watermark marking generative photo edits | A visible watermark may be overlaid |
| Meta platforms | Reads industry indicators in the file | "AI info", in the post menu for edits |
| IPTC standard | compositeWithTrainedAlgorithmicMedia for edits; trainedAlgorithmicMedia for generated images | None; used by apps and platforms |
Adobe now attaches C2PA metadata automatically "when qualifying generative AI tools are used to create or modify content" in its apps, recording whether the content was wholly generated or modified with generative AI, along with the app, model and date; it says this "cannot be disabled for qualifying content" and that it applies only to generative tools, not traditional AI features (Adobe AI content labeling FAQ). Images that are fully generated in Firefly have long carried credentials automatically (Adobe Firefly).
Google Photos shows "Edited with AI tools" on photos changed with generative edits such as Magic Eraser, "where new pixels are created", and adds a credential when such an edit is saved (Google Photos Help); Google had already written IPTC metadata for Magic Editor, Magic Eraser and Zoom Enhance edits (Google, 2024). Samsung says images changed with Galaxy AI generative photo editing are watermarked, and a visible watermark may be overlaid (Samsung).
Inpainting, fill and upscaling
Generative fill and inpainting replace part of an image with new, generated pixels, so they count as AI edits under every scheme above. Removing an object with a generative eraser counts too, because the gap is filled by a model.
Upscaling is less clear. Google lists its AI zoom feature among the edits that get IPTC metadata, and Adobe's rule covers its generative tools; a traditional upscaler that only sharpens fits the "algorithmically enhanced" term instead. When in doubt, check what the tool wrote with a credential viewer, and keep the original photo. Our method for reading files is the same on every page.
How platforms label edits
Meta began labelling AI images in 2024, renamed its label "AI info" after small retouching edits were being tagged "Made with AI", and in September 2024 said that for content "only modified or edited by AI tools, we are moving the 'AI info' label to the post's menu" (Meta newsroom). Fully generated images keep the label in view.
Labels follow the records in the file, so the same photo can be labelled on one upload and not the next if it was re-saved in between. For fully generated images see ChatGPT images; for realistic edits of people see real people in prompts; for selling edited work see selling AI images.
A missing label does not mean an image is human-made.
Sources
- Adobe: AI content labeling FAQ, updated 2026-08-25 (accessed 2026-11-02)helpx.adobe.com
- Adobe Firefly: Content Credentials overview (accessed 2026-11-02)helpx.adobe.com
- IPTC NewsCodes: Digital Source Type (accessed 2026-11-02)cv.iptc.org
- Meta: Our approach to labeling AI-generated content and manipulated media (accessed 2026-11-02)about.fb.com
- Google: AI editing transparency in Google Photos, 2024-10-24 (accessed 2026-11-02)blog.google
- Google Photos Help: Content Credentials (accessed 2026-11-02)support.google.com
- Samsung: Generative photo editing on the Galaxy S24 (accessed 2026-11-02)samsung.com
Questions
What does the AI info label mean on Instagram?
It is Meta's label for content it detects or users disclose as made or changed with AI. Since September 2024, for images only edited with AI, the label sits in the post menu rather than on the image.
Does Photoshop generative fill add a label?
Adobe says its apps attach C2PA metadata automatically when qualifying generative AI tools modify content, recording that generative AI was used. Adobe does not add a visible label.
Is an edited photo AI-generated?
Not entirely. IPTC calls it compositeWithTrainedAlgorithmicMedia, edited using generative AI, as distinct from an image created entirely by a model.
Does removing an object with AI count as an AI edit?
If the tool fills the gap with generated pixels, yes. Google Photos, for example, marks Magic Eraser edits as edited with AI tools.
Does upscaling count as editing with AI?
It depends on the tool. Generative upscalers create new detail and may be labelled; simple sharpening fits IPTC's algorithmically enhanced term instead. Check the file with a credential viewer.
Related reading
- Made with AIWhat your AI image carries: credentials, watermarks, metadata and the prompt itselfAn AI-generated image file can carry four records: a signed content credential, an invisible watermark in the pixels, ordinary metadata fields, and the prompt saved as text.
- Made with AIWhat ChatGPT puts in your images: content credentials, and when they surviveOpenAI says every image made with ChatGPT, Codex or its API carries two records: a signed C2PA content credential and an invisible SynthID watermark.
- PromptsPhotorealistic prompts: realistic results, and why realistic images need the clearest labelsFor photorealistic results, describe the photograph rather than asking for realism: name the lens, the viewpoint, one light source and a plain finish.
- Made with AISelling AI images: what stock sites and marketplaces require you to declareSome marketplaces take AI images and some refuse them.