AI watermarking: flagging content made by a machine

A hidden signature inside the content

Think of the watermark on a banknote, but digital. When the AI creates an image, a sound or a text, it slips in a subtle signal. Verification software can later read it to confirm where the content came from.

There are two main approaches:

• Watermarks built into the content: an invisible pattern written into pixels, audio or word choices. Google DeepMind launched SynthID in August 2023 for generated images; according to its official page, the technology now also covers video, audio and text.
• Provenance metadata: a « label » attached to the file describing how it was made. The C2PA consortium develops an open standard for this kind of information.

Why it matters to you

A realistic photo of an event that never happened, a cloned voice on the phone, a fake video of a politician: deepfakes and AI scams are on the rise. A watermark offers a way to check where content came from, provided the tool that made it added one.

In Europe, the EU AI Act pushes in this direction: providers of generative AI must make their output identifiable, and certain content such as deepfakes must be clearly labeled.

A concrete example: before sharing a shocking image you received in a group chat, you can run it through a verification tool offered by some assistants or AI makers, and look for the original source.

Its limits, plainly stated

• It isn’t foolproof. Google DeepMind itself acknowledges that SynthID can’t withstand every extreme manipulation.
• It isn’t universal. Each company has its own technique; one checker won’t necessarily recognize another company’s watermark.
• Text is the hardest case. Rewritten or translated text can lose its mark. AI text detectors, which work without a watermark, are often wrong.

Common misconceptions

• « No watermark means human-made. » Wrong, a missing mark proves nothing.
• « Watermarks ruin the image. » Modern techniques are specifically designed to stay imperceptible.
• « With watermarks, I no longer need to think critically. » It’s one clue among others, to be weighed against the source and context. Unmarked AI-generated images still circulate.