Decoder. plain-English AI glossary

Watermarking

▲ Rising

Also called Watermarking

Embedding a hidden, detectable signal in AI output so it can later be identified as machine-made.

Think of it like

The faint watermark in a banknote — invisible in normal use, but proof of origin when you hold it to the light.

Example

A text generator subtly biases its word choices in a statistically detectable pattern, so a checking tool can later flag the passage as likely AI-written.

How it actually works

Watermarking aims at provenance: proving what a model produced, to fight deepfakes, misinformation, and undisclosed AI content. In text it nudges token probabilities in a secret pattern; in images it embeds signals in the pixels. The tension is robustness versus imperceptibility — a watermark strong enough to survive edits and paraphrasing tends to be easier to notice or strip. No scheme is yet both invisible and truly tamper-proof.

For product teams

A provenance tool for a world drowning in synthetic media — promising, but not yet a guarantee you can lean on.

For engineers

Embed a detectable statistical signal (token biasing in text, pixel signals in images); robustness to edits trades against imperceptibility.

Related

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open DecoderAdd to Chrome