Explain like I'm five · AI Security

What is an Adversarial Example?

Adversarial examples, told as a story about a photo that looks the same to a human eye but is read entirely differently by the chef.

this isn't a salad — it's a pass-through formto a human eye, it's just an ordinary salad photo

To a human eye it's the same salad — but the chef reads it differently.The photo looks unchanged, but the chef reads something else entirely.

reads pixel numberssees the whole picturea person sees the whole picture; the chef adds up each tiny dot's number — nudge those numbers, and the reading flips entirely

The chef's eye reads each dot's number; a human eye takes in the whole picture.Nudge those numbers slightly, and the reading flips completely — even though a person feels nothing changed.

what a person seesjust a saladwhat the chef readsthis is a pass-through formthe picture that fools the eye: normal to a person, something else entirely to the chef

Normal to a person, something else entirely to the chef.The exact same photo — only the reading splits completely apart.

Add a noise patternscatter dots too faint to notice
Nudge the edgespush the shape's edges a little
Stick on a patterneven works stuck onto a real object
Search by trialrepeatedly tune the perturbation
original photoit's a salad ✓a subtly perturbed photoit's a pass form ×identical to a human eyeadd just a handful of tiny dots, and the reading flips completely

Add just a handful of dots, and the reading flips.Identical to a human eye — but the chef's reading splits completely apart.

hides a text command inside the picture→ smuggling a commandperturbs the photo itself, ever so slightly→ making it misread entirelyone hides a command; the other has no hidden command at all — it just breaks the seeing itself

Hiding a command is different from making something misread.Image injection has a hidden command. An adversarial example has none — it just breaks the seeing itself.

In one breath

Adversarial example = a photo that looks the same to a human eye but is subtly perturbed so the chef reads it entirely differently.

An adversarial example is an input with a perturbation (noise, a pattern) so small a human can't notice it, crafted to make a model's recognition or classification fail in a chosen way. This applies to images as well as audio and text, and can be crafted with knowledge of the model's weights (white-box) or without it, through repeated trial (black-box). Unlike prompt injection, which hides an instruction, an adversarial example contains no hidden instruction at all — it simply breaks the model's perception itself.

When grown-ups say it

Adversarial example
A photo that looks the same but is read differently by the chef. Not a hidden command — the seeing itself is wrong.
Image injection
Hiding a text command inside a picture. Covered on the page right before this one. → letters hidden in a menu photo
Perturbation
A small change too faint to notice. A slight nudge to dots or edges.
White-box · black-box
Whether you can see the recipe book, or only try by trial. Knowing the recipe book lets the perturbation be far more precise.
Classifier
The chef's eye, which looks at a photo and tags what it is. Fooling this eye is exactly what an adversarial example aims for.
Misclassification
Tagging something with the wrong label. The outcome an adversarial example is built to cause.
Guardrail
The taster who checks a note's words. Since the seeing itself is wrong — not the words — even this taster can't catch it. → What is a Guardrail?
Red team
People who test the eye with deliberately perturbed photos. AI red teams hunt for this too. → the security world's red team
Next story
Copying the Whole Recipe Book →
Want another term as a picture book? Request →