Watermarking means four different things

Why this post exists

The query this site is shown for most often is "high-speed watermarking". Almost nobody clicks, and they are right not to, because that phrase belongs to a different industry. "Forensic watermarking" and "artificial watermark on checks" are the same story.

Four unrelated technologies share the word. Here is the map, so you can get to the right one quickly.

1. Forensic watermarking — finding out who leaked it

A studio sends a film to two hundred reviewers. Every copy is bit-for-bit different: each carries an imperceptible identifier tied to one recipient. When a copy appears on a piracy site, the identifier says which of the two hundred it came from.

That is forensic watermarking, and speed is the hard part. For live sport you have to embed a per-viewer mark on a stream going out to millions, in real time, without adding latency or a second encode. The usual trick is to prepare two variants of each segment and have the delivery network choose between them per session, so the identifier is written by which segments a viewer received rather than by re-encoding anything. That is what "high-speed watermarking" is describing.

It is mature technology with real vendors behind it, and it answers exactly one question: which recipient did this copy come from? It says nothing about whether the content is real, edited, or machine-generated. If that is your problem, you want an anti-piracy vendor, not this site.

2. Generative watermarking — marking machine-made content

This is the AI one, and it works on a different principle. Instead of attaching anything, it nudges the content itself: tiny adjustments to pixel values, to an audio waveform, or to which words a language model picks. Google's SynthID is the best-known example, it covers images, audio, video and text, and it is now in OpenAI's images too. Anthropic has been marking Claude's text output since August 2026.

The mark is durable, because there is nothing beside the content to strip. The problem is reading it. Detection needs the model owner's own detector, and those are waitlisted, partner-only, or simply unpublished. So the most robust signal in the field is the one you cannot check, which is an uncomfortable thing to build a policy on.

3. Physical watermarking — paper

The original meaning, and still the largest by volume. A true watermark is made during manufacture by varying the thickness of the paper, so it shows when held to light and cannot be photocopied. Banknotes work this way.

Most cheques do not. Watermarked stock is expensive, so cheques usually carry an artificial watermark: a design printed on the back in ink that is invisible head-on and appears when you tilt the paper to a shallow angle. The instruction printed on the cheque — hold at 45 degrees to view — is the giveaway. It is a security feature against flatbed copying, and the detector is your eyes.

4. The one that is not a watermark at all

Content credentials, the C2PA standard, get called watermarks constantly. They are not. A credential is a signed manifest travelling alongside the pixels, recording what made the file and what has happened to it since — closer to a sealed envelope than to a mark in the paper.

The distinction has consequences. Because the manifest sits beside the content rather than inside it, a screenshot removes it completely, while a SynthID mark in the same image would survive. Calling both "watermarking" is how people end up expecting a credential to survive things it was never built to survive. There is a longer explanation of that split in the explainer on AI watermarks.

Which one do you need

• Someone leaked our content and we want to know who — forensic watermarking. Anti-piracy vendors.
• Was this image or text made by a model? — generative watermarking. Mostly unreadable outside the companies that make the models, and no honest tool will tell you otherwise.
• Is this cheque or certificate genuine? — physical security printing. Tilt it.
• Where did this file come from, and has it been altered? — content credentials and metadata. That one you can check yourself, right now.

What this site does

The last of the four. It reads C2PA content credentials, the full metadata a file carries, hidden Unicode in text, and a SHA-256 of the bytes — all in your browser, and it tells you plainly which signals were present, which were stripped, and what each result does and does not prove.

It cannot read SynthID, because nobody outside Google can. It says so rather than guessing, which is the whole point. Run a file through it and see what yours actually carries.