what is AI watermarking?

the term “AI watermark” gets used for two very different technologies. they solve different problems and have different adoption stories.

neural watermarks are invisible signals embedded into a generative model's output at creation time. Google's SynthID nudges token probabilities for text and pixel values for images, video, and audio so a detector with the right key can later recover the signal. Meta's Stable Signature (Fernandez et al., ICCV 2023) fine-tunes a latent-diffusion decoder to embed a hidden binary signature in every image it produces. these work for the model they're baked into and survive moderate edits. they don't survive aggressive paraphrase, recompression, or adversarial fine-tuning; the 2024 paper “Stable Signature is Unstable” demonstrates removal in practice.

C2PA Content Credentials are a cryptographically signed metadata manifest, attached to a file, that records its provenance chain. who or what device captured it, what edits were applied along the way, and whether AI was involved. signatures verify origin and chain of custody. they don't detect synthesis. when present, they tell you where a file came from with high confidence. when absent, they tell you nothing.

adoption as of 2025: Google's Pixel 10 cameras, Leica SL3-S, Canon EOS R1 and R5 Mark II, and Sony PXW-Z300 sign at capture. SynthID is embedded in Google's own Gemini, Imagen, Veo, and Lyria outputs (over 10 billion items signed as of Google I/O 2025). OpenAI signs DALL-E and Sora outputs with C2PA but does not currently ship a neural watermark. Midjourney, Stability AI, Black Forest Labs (Flux), Ideogram, and most open-source generators ship neither. the long tail of generated content in the wild is unwatermarked.

social-media platforms are unkind to provenance metadata. C2PA manifests are commonly stripped during recompression and re-encoding, so a video that was signed at capture often arrives on a user's phone as a bare MP4. watermarks fare somewhat better because they're in the pixels, not the metadata, but screenshots, crops, and quality reductions still degrade them.

presence of a watermark is positive evidence; absence is nothing. a SynthID-positive image is almost certainly Google-AI. an image with no watermark could be made by Midjourney, Stable Diffusion, a human, or a watermarked generator whose signature was stripped on social media. amige. reads provenance when it's there and falls back to model attribution and ensemble detection when it isn't, which is most of the time.

vendor copy tends to describe both technologies as “robust” or “tamper-evident.” peer-reviewed work is more guarded. treat watermarks as a useful positive signal of AI origin when present, never as a negative signal when absent.