How to read AI image provenance

Four different things people collapse

• C2PA / Content Credentials — signed manifests bound to the file. Can survive some edits and fail others. A trusted official result is a strong file-level signal, not a pixel oracle.
• SynthID and similar watermarks — designed to live in pixels. A metadata stripper cannot read or remove them. Official OpenAI verification may report SynthID for supported assets after you choose to upload.
• Ordinary metadata — EXIF/XMP/PNG text, software tags, prompts stuffed in comments. Easy to delete or fake.
• Visual classifiers — “this looks AI”. They are not implemented here as a score, and they are not C2PA.

Trusted positive vs missing

A trusted positive from official OpenAI C2PA/SynthID, after you click verify, is evidence about that API’s view of that file. Local inspection showing a C2PA/JUMBF structure is a clue you can screenshot for yourself.

A miss means this scan did not find the field. The image may still be synthetic, the credential may have been stripped, the watermark may be unsupported, or the file may be a screenshot.

Screenshots and recompression

A screenshot is a new raster. C2PA almost always dies. EXIF is usually gone or replaced by the phone. Pixel watermarks may degrade. Treat screenshot provenance as weak unless an official verifier still reports a supported watermark.

Walkthrough: what survives a screenshot.

AI-generated vs. AI-edited (inpainting and composite)

Not all AI media is 100% synthetic. Provenance standards define distinct creation modes via IPTC DigitalSourceType and C2PA action verbs:

• Pure AI-Generated (trainedAlgorithmicMedia) — the image was synthesized from scratch via prompts (e.g. ChatGPT / DALL·E, Grok Imagine, Midjourney). C2PA action: c2pa.created.
• AI-Edited / Composite (compositeWithTrainedAlgorithmicMedia) — a real camera photograph retouched with generative fill, object removal, or inpainting (e.g. Photoshop Generative Fill, Google Magic Editor). C2PA actions: c2pa.infilled, c2pa.edited, c2pa.placed.
• Algorithmically Enhanced (algorithmicallyEnhanced) — computational photography, HDR, and noise reduction applied by smartphone image signal processors.
• Hardware Camera Capture (digitalCapture) — genuine light hitting a camera sensor with no algorithmic generation (e.g. Leica M11-P, Sony C2PA).

Why platforms mislabel edited photos: Social networks like Instagram, Facebook, and TikTok inspect C2PA manifests. Because early platform bots checked only whether the word AlgorithmicMedia was present, genuine real-world photos retouched with Photoshop spot removal or Samsung/Google Magic Eraser were mistakenly branded with giant “Made with AI” or “AI Info” tags. Inspecting the raw manifest reveals whether your photo was synthesized or merely retouched.

Workflow in this kit

• Local pass in AI Image Origin Detector — no file upload.
• If you need OpenAI’s official C2PA/SynthID result, click verify. That path sends the image through our Worker to OpenAI.
• If you needed a sharing copy without credentials, that is Metadata Cleaner, and it still will not erase SynthID.
• Read ChatGPT image checks and C2PA checks before treating a UI badge as proof.
• Definitions stay on their own pages: what content provenance is, what C2PA is, and what Content Credentials are.

Platform labels are not the file

Instagram’s help page says content created or modified with Meta AI tools can be labeled automatically, including a visible “Imagined with AI” watermark on some images. Files that carry industry-standard AI signals, including third-party tools, can get an “AI info” label. Meta’s cross-product help uses the same idea on Facebook, Instagram, and Threads. The older public wording was “Made with AI.” Meta also says not every AI image contains a signal it can read.

TikTok’s September 2026 newsroom note says it labels AI-generated content with Content Credentials, creator labels, and an invisible watermark it says only TikTok can read. Stripping a manifest from your copy does not describe that watermark.

YouTube requires creators to disclose realistic altered or synthetic content in Studio. Its help page says a valid C2PA credential that the entire video was made with AI can be carried forward into “How this content was made.” In May 2026 YouTube said it would also apply a label when its systems detect significant photorealistic AI and the creator did not disclose.

None of those policies is a promise that cleaning EXIF or C2PA will clear the label. Inspect the file with the Content Credentials checker. The label walkthrough is Instagram AI info. Definitions: C2PA, Content Credentials, SynthID.