If you've ever seen Instagram put an "AI info" label on a photo you know is real, the cause is almost never the pixels. It's metadata. I've spent the last few months building AI Flag Remover, a browser-based tool that strips this metadata, so here's what I learned about where it lives and how to inspect it yourself.
| Signal | Where it lives | Who writes it | Who reads it |
|---|---|---|---|
| C2PA manifest | JUMBF boxes (JPEG APP11, PNG caBX chunk, WebPC2PA chunk) |
Adobe, OpenAI, Google, Microsoft | TikTok, LinkedIn, Meta, Pinterest |
| IPTC Digital Source Type | XMP packet, Iptc4xmpExt:DigitalSourceType |
Meta AI, many generators | Meta (Instagram, Facebook, Threads) |
| Generator text | PNG tEXt /iTXt chunks (parameters ,prompt ,workflow ) |
Stable Diffusion WebUI, ComfyUI, Forge | Anyone who opens the file |
The IPTC value to look for is a URI ending in trainedAlgorithmicMedia (fully generated) or compositeWithTrainedAlgorithmicMedia (real photo edited with AI). That second one is the usual culprit when a real photo gets labelled after a Generative Fill edit.
The quickest check, if you have ExifTool installed:
exiftool -G1 -a -s -DigitalSourceType -JUMBF:all -XMP-iptcExt:all photo.jpg
exiftool -PNG:Parameters -PNG:Prompt -PNG:Workflow image.png
C2PA in JPEGs sits in APP11 segments containing a JUMBF box labelled c2pa. This scans the segment markers:
import sys
def find_c2pa_jpeg(path):
data = open(path, "rb").read()
i = 2 # skip SOI
found = []
while i < len(data) - 4 and data[i] == 0xFF:
marker = data[i + 1]
length = int.from_bytes(data[i + 2:i + 4], "big")
segment = data[i + 4:i + 2 + length]
if marker == 0xEB and b"c2pa" in segment: # APP11 + JUMBF label
found.append(("APP11/C2PA", length))
if marker == 0xE1 and b"DigitalSourceType" in segment: # APP1 XMP
found.append(("XMP/DigitalSourceType", length))
if marker == 0xDA: # start of scan, image data follows
break
i += 2 + length
return found
print(find_c2pa_jpeg(sys.argv[1]) or "No AI metadata markers found")
Stable Diffusion prompts are plain tEXt chunks, so a few lines of JavaScript can read them client-side:
async function readPngText(file) {
const buf = new DataView(await file.arrayBuffer());
const out = {};
let p = 8; // skip PNG signature
while (p < buf.byteLength) {
const len = buf.getUint32(p);
const type = String.fromCharCode(...new Uint8Array(buf.buffer, p + 4, 4));
if (type === "tEXt") {
const bytes = new Uint8Array(buf.buffer, p + 8, len);
const [key, ...val] = new TextDecoder("latin1").decode(bytes).split("\0");
out[key] = val.join("\0");
}
if (type === "caBX") out.c2pa = `${len} bytes`;
p += 12 + len; // length + type + data + CRC
}
return out; // e.g. { parameters: "a cat, 8k ... Steps: 30, Seed: 1234" }
}
Removing a chunk is just skipping it when you rebuild the file. Because you're copying the image data chunks byte for byte, there's no re-encoding and no quality loss.
IFD1 too if privacy matters.
C2PA exists for good reasons: provenance helps people spot fakes. Stripping it makes sense for false positives (real photos with a minor AI edit) and privacy (GPS, serial numbers, leaked prompts). If content is genuinely AI-generated, platforms and laws like the EU AI Act may still require you to disclose it.
If you just need clean files, AI Flag Remover does all of the above in the browser (JPG, PNG, WebP, HEIC, AVIF), shows which AI flags it found, and re-checks the output. There's also a free AI image metadata checker if you only want to inspect.
What other metadata quirks have you run into? Let me know in the comments.