Key takeaways
- An imperceptible watermark lives inside the pixels, samples or frames of a file, not in metadata beside it. That is why it survives screenshots, re-uploads and the re-encoding every platform applies.
- It does not replace C2PA. The C2PA manifest carries the detailed provenance record; the watermark lets a verifier find that record again after the manifest has been stripped.
- EU AI Act Article 50 requires AI-generated output to be marked in a machine-readable, detectable way. California SB 942 goes further and names the fields a hidden 'latent disclosure' must carry.
- Judge a watermark on five things: whether it is truly imperceptible, what it survives, how often it gives false matches, what it can carry, and who can check it.
What imperceptible watermarking is
Imperceptible watermarking embeds a machine-readable signal into the content of a media file, so that software can later detect it and read what it says, while a person looking at or listening to the file notices nothing.
“Imperceptible” is the important word. A visible watermark, such as a logo in the corner of a stock photo, works by being seen. An imperceptible one works the other way round: it changes the content by amounts that sit below what human eyes and ears can pick up, and it is only useful because a detector can find it.
The signal is part of the content itself: the pixels of an image, the samples of an audio track, the frames of a video. That is the property everything else follows from. Anything that copies the content copies the watermark with it.
The same is true for audio. The watermark is carried in the waveform itself, not in an ID3 tag or a file-name convention, so it survives the re-encoding, trimming and compression every platform applies on upload.
Watermarks, labels, metadata and fingerprints
Four techniques get called “provenance”, and they fail in different places. The practical question for each is what happens after the file has been uploaded to a platform, screenshotted or re-saved, because that is what happens to most content within minutes of being published.
| Watermark | Label | Metadata / C2PA | Fingerprint | |
|---|---|---|---|---|
| Where it lives | Inside the content | On top of the content | Beside the content, in the file | In a database, not the file |
| Visible to people | No | Yes | No | No |
| Survives upload and screenshot | Yes | Partlyuntil cropped | Nousually stripped | Partlyif near-identical |
| Machine-readable | Yes | No | Yes | Yes |
| Can carry a record | Yes, a compact payload | A short phrase | Yes, a full signed record | No, only a match |
None of the four is enough alone, and the laws below do not treat them as alternatives. A visible label tells a person. Metadata carries the full, signed record. A watermark makes sure that record can still be found once the metadata is gone. A fingerprint helps match copies of content that was never marked at all.
How watermarking works with C2PA
C2PA is the open standard for attaching a signed provenance manifest to a file: who made it, with what tool, and what has been done to it since. It is the most complete record available, and it has one well-known weakness. The manifest sits in the file's metadata, and most social platforms, messaging apps and screenshots throw metadata away.
The C2PA specification answers this with soft binding: an identifier embedded in the content itself, which a verifier can use to look up the original manifest after it has been stripped. An imperceptible watermark is the standard way to provide one. The specification keeps a public list of registered soft-binding algorithms, and Verda's image, audio and video watermarks are on it.
How to evaluate an imperceptible watermark
Every vendor describes its watermark as invisible and robust. These five questions separate the claims that can be checked from the ones that cannot. Ask for results on your own content, not a vendor's demo set.
Is it actually imperceptible on your content?
Ask for side-by-side samples and blind comparisons on material like yours: skin tones, flat colour, quiet passages in audio, fast motion in video. Artefacts tend to show up in exactly those places.
What does it survive, specifically?
“Robust” means nothing without a list. Ask which transformations were tested and at what strength: compression, resizing, cropping, screenshots, platform re-encoding, format changes, and for audio, re-recording through a speaker. Ask where it stops working, too. An honest vendor has an answer.
How often does it claim a match that isn't there?
A false positive, reading a watermark in unmarked content, does more damage than a miss, because it attributes content to someone who did not make it. Ask for the false-positive rate and how it was measured.
What can it carry?
Some watermarks carry a single bit, “AI-generated”. Others carry an identifier that resolves to a full record. California SB 942 needs the second kind: a flag alone cannot convey a provider name, system version, timestamp and unique ID.
Who can check it?
A watermark only the vendor can read is a claim nobody else can verify. SB 942 requires a free public detection tool with an API, and C2PA soft binding assumes third parties can resolve the mark. Check that detection is open, free and documented.
What the law requires
Three regimes now expect AI-generated media to carry a mark that machines can read. They differ on who owes the duty and on how much the mark has to say.
| EU AI Act Art. 50 | California SB 942 | India IT Rules | |
|---|---|---|---|
| Who owes it | Providers of generative AI systems | Covered providers: over 1M monthly users | Platforms and intermediaries |
| What the mark must do | Machine-readable and detectable as AI-generated | Hidden “latent disclosure” with four fields, plus a free detector | Prominent label; permanent metadata or unique ID where feasible |
| Robustness standard | Effective, interoperable, robust and reliable, as far as technically feasible | Permanent or extraordinarily difficult to remove, to the extent technically feasible | Where feasible |
| Key date | 2 Aug 2026 new systems; 2 Dec 2026 existing | Operative since 2 Aug 2026 | In force since 20 Feb 2026 |
The EU text is in Article 50; the detail is on our Article 50 guide. California's is SB 942, amended by AB 853, covered field by field in our SB 942 guide. India's amended intermediary rules are on the India IT Rules page.
What watermarking cannot do
Being clear about limits is part of choosing well, and the laws acknowledge them with their “as far as technically feasible” wording.
- No watermark is impossible to remove. A determined attacker with enough control over a file can degrade the signal, usually at a visible or audible cost to the content. The realistic goal is that ordinary handling never removes it and deliberate removal is costly.
- It proves where content came from, not whether it is true. A watermark records who marked a file and when. It says nothing about whether what the file shows actually happened.
- Absence proves nothing. Most content in circulation carries no watermark at all. Not finding one does not mean a file is authentic, or that it is fake.
- Text is different. Watermarking written text is a separate problem with weaker guarantees. This guide covers images, audio and video.
How Verda approaches it
Verda Inc. develops imperceptible watermarking for AI-generated image, audio and video, embedding durable, machine-readable provenance that is free for anyone to verify.
- Image, audio and video, each registered on the C2PA soft-binding algorithm list, alongside a signed C2PA manifest.
- Free public verification. Anyone can check a file or link at verda.ai/verify, and there is an API for platforms and detectors.
- Provider-resolvable provenance. The mark resolves to a record of which system produced the content and when, which is what SB 942's latent disclosure asks for.
Common questions
What is imperceptible watermarking?
Imperceptible watermarking embeds a machine-readable signal into the content of an image, audio clip or video. People cannot see or hear it, but a detector can read it, and because it is part of the content rather than the metadata it survives uploads, screenshots and re-encoding.
Is an invisible watermark the same as an imperceptible watermark?
In practice, yes. 'Invisible' is used for images and video; 'imperceptible' also covers audio, where the mark must be inaudible. Both mean a signal that people do not notice but software can detect.
Can an imperceptible watermark be removed?
No watermark is impossible to remove. A good one survives ordinary handling such as compression, resizing, cropping, screenshots and platform re-encoding, and deliberate removal degrades the content noticeably. The EU AI Act and SB 942 both set the bar at what is technically feasible.
Does C2PA replace watermarking?
No. C2PA manifests are stored in metadata, which most platforms strip. The C2PA specification uses an embedded watermark as a 'soft binding' so the manifest can be recovered after it is removed. The two are designed to work together.
Does the EU AI Act require watermarking?
Article 50(2) requires providers of generative AI systems to mark output in a machine-readable format that is detectable as AI-generated, and to make the solution effective, interoperable, robust and reliable as far as technically feasible. It does not name one technique, but an imperceptible watermark is the practical way to meet the robustness standard.
Does California SB 942 require an invisible watermark?
SB 942 requires covered providers to embed a latent disclosure, which is a hidden, machine-readable mark, carrying the provider name, the system name and version, the time and date, and a unique identifier. It must be permanent or extraordinarily difficult to remove to the extent technically feasible, and providers must offer a free public detection tool.
Does a watermark prove content is real?
No. A watermark shows who marked a file and when. It does not show whether what the content depicts is true, and the absence of a watermark proves nothing either way.
Check a file, or talk to us about marking yours
Verification is free for anyone. If you run a generative AI system and need to meet Article 50 or SB 942, we can walk you through it.