Explain like I'm five · AI Security

What is a Watermark?

Watermarks, told as a story about an invisible mark baked into a dish so it can still be checked later.

the dish leaves the kitchenis this really from this kitchen?someone else claims they made itlooking at the plate alone, you can't tell which kitchen it came from

Looking at the plate alone, you can't tell which kitchen it came from.Once it's out the door, there's no way to prove who made it.

made by this kitchencopied and replateda totally different kitchen's dishwith no mark on any of them, there's no telling them apartonce it leaves the kitchen, there's no way to trace it back

Once served, they all look like the exact same plate.With no mark, there's no telling the copy from the original.

a tiny mark baked into the dish — invisible to the eyelater, a special tool can still reveal the marka watermark = a mark baked into the dish, checkable even laternot a paper label — it's left in the dish itself

A watermark is a mark baked into the dish, checkable even later.Not a paper label — it's left in the dish itself.

Baked in while cookingnot something added on afterward
Invisible to the eyeno change in taste or look
Revealed with a special toolnot something anyone can just glance and see
Not foolproofheavy enough rework can erase it
① the dish is made② the mark is baked in③ it's served to a guest④ checked later

Make the dish, bake in the mark, serve it, and check it later.The check can happen long after the dish has already been served.

stays inside the dish, even after it's servedthe model card is before serving — the watermark is afterrework it heavily enough and the mark can fadeno mark is perfect — but having one still helps tracing far more than having none

The model card is the label before serving; the watermark is the mark after.A model card is a notice posted before ordering; a watermark is a mark left on the finished output, after the fact. Neither is perfect, but each closes — at a different point in time — the gap that opens because the chef trusts the guest's words by instinct.

In one breath

A watermark = an invisible mark baked into the dish. Checked later with a special tool, it reveals whether a dish really came from this kitchen.

Watermarking is a technique for embedding a mark in AI-generated output (text, images, audio) that's hard for a person to notice but can be confirmed by a dedicated detector. It's inserted at generation time as a statistical pattern (e.g., a bias in token selection) or signal, and later checking for that pattern reveals whether the content was AI-generated. Unlike a model card, which is written before deployment, a watermark is applied after the fact to output that already exists — and it has the limitation that heavy enough transformation of the output can break the mark.

When grown-ups say it

Watermark
An invisible mark baked into the dish. Only a special tool can reveal it.
vs. the dish's label
The label comes before serving — the watermark after. Companions at different points in time. → the dish's label
Detector
The special tool that reads the mark. Used only by inspectors — no one can see it with the naked eye.
Provenance
Proof that this dish really came from this kitchen. This is what the watermark is meant to back up.
Synthetic content
A dish made entirely by the chef, not a person. The mark matters most for this kind of dish. → the parrot world's synthetic beans
Scrubbing attack
Deliberately heavy rework meant to erase the mark. This is the watermark's biggest limitation.
Compliance
A rule that makes marking mandatory. Some jurisdictions now require labeling AI output by law. → the security world's compliance
Audit log
The ledger of when each dish's mark was baked in. Even if a mark fades, the ledger can still help trace it back.
Next story
The Inspector's Surprise Visit →
Want another term as a picture book? Request →