AI Content Label: Add an «AI-generated» Mark and Check It

Meta reads the IPTC «Digital Source Type» field and the C2PA signature that editors write. Generative fill, background expansion or a sky replacement record «an image with AI-generated parts» — and the label appears even if AI touched a tiny area. The «Check a file» mode shows exactly which field caused it. This tool cannot remove other people's labels and will not.

It depends on the country and the purpose. In the EU, under Article 50 of the AI Act, whoever professionally publishes a realistic deepfake — a person, place or event that did not exist — must disclose it; personal non-professional use is outside the rule, and for artistic works the label must not hamper the display. In China the CAC rules and GB 45438-2025 require a label; in Korea the duty lies with the AI service. In Russia, under Federal Law 243-FZ, labeling is a right, not a duty. This is general information, not legal advice.

The visible label will: it is in the pixels or the sound. Messengers usually strip metadata and the C2PA signature when they recompress, so for publishing through them the visible label matters more than the metadata record. Sending it «as a file» rather than «as a photo» usually keeps the metadata.

No. A label is a statement by whoever saved the file, and it is easy to lose: a screenshot, recompression, a social network. Many generators set no label at all. The tool only tells you what is written in the file and does not guess from pixels — the «probability of AI» from detectors is often wrong both ways.

The C2PA signature is intact, but the signer's certificate is not on the C2PA trust list: anyone can sign like this with a certificate they issued themselves. Here that case is shown apart from «signer verified» and from «signature mismatch». A mismatch means the file was changed after signing.

SynthID is Google's invisible mark in images, video and audio from Gemini and Imagen. Its detector is closed and available only in the Gemini app. If the file carries Google metadata — «Made with Google AI» or their C2PA signature — you see it here; check SynthID itself in Gemini.

Metadata is added without re-encoding: pixels and sound stay byte for byte. A visible label changes pixels, so the picture is rebuilt: JPEG at the source's quality, PNG losslessly. In WAV and MP3 the disclaimer is inserted before the recording without re-encoding it; M4A is re-encoded once. A video with a visible label is encoded once, at a bitrate taken from the source, so the file does not balloon.

It is the implicit label of the GB 45438-2025 standard: an AIGC field with JSON where Label 1 means «AI-generated», 2 — «possibly», 3 — «suspected», plus the code of the service that made the file and of the platform that spread it. Douyin and WeChat read it on upload. The «China» preset writes it together with the visible «AI生成» text.

It depends on your role under the law. Whoever publishes a deepfake, or AI-written text on a matter of public interest (Art. 50(4), the “deployer”), has to disclose it clearly — that is what the visible label, the spoken disclaimer and the caption do. A provider of a generative system (Art. 50(2)) must mark every output in a machine-readable, detectable way; the Code of Practice names signed, timestamped metadata plus an invisible watermark for that. This tool writes unsigned IPTC/AIGC metadata: it does not sign C2PA or embed a watermark, so a provider cannot rely on its label alone. This is general information, not legal advice.