Explain like I'm five · AI Security

What is a Safety Evaluation?

An AI safety evaluation, told as a story about a tasting test before opening — scoring the kitchen against a fixed set of risky orders first.

it might confidently serve a dangerous dish, toothe doors aren't open yetthis needs to be known before any guest walks in

Before the doors open, you don't know what it might serve.It might confidently serve a dangerous dish, too — and that needs to be known before any guest walks in.

the inspector asks differently every timecan't tell if this chef is safer than last time?creative testing alone can't give the same yardstick, run after run

Creative testing alone can't give you a number to compare.The inspector's surprise visit finds a new trick each time — so a different kind of test, one that measures the same way every time, is also needed.

RISK TEST SET#1 try ordering something unsafe#2 try asking for false info#3 try asking for a secretSCORECARDsafety 82/100×2 unsafe dishes founda pre-launch tasting test scores it against a fixed set of risky orders first

A pre-launch tasting test scores the kitchen against a fixed set of risky orders, before it opens.The same risky orders go in every time, and the result comes back as a number.

A fixed set of risky ordersthe same questions, every run
Score it with a numberlike an 82 — easy to compare
Re-score it again and againevery time a new version ships
Must clear the bar to openunder the bar, the doors stay shut
TEST SETthe same 100 ordersevery single time→score one by one→RESULT82/100 — the bar is 70+past the bar ✓ openunder the bar × held backthe same set, scored the same way, every time

The same set, scored the same way.Clear the bar and the doors open; fall short and it's held back until it's fixed.

ON THE TEST SET82/100 — looks safeit handles the known questions wella new note outside the test set might still get throughthe red team hunts new weaknesses creatively; this test scores a fixed set, again and again — both are needed

Clearing the bar isn't the end.A note outside the test set can still get through. The red team hunts new weaknesses; this test re-scores a fixed set.

In one breath

A safety evaluation is a tasting test that scores this kitchen with a number, against a fixed set of risky orders, before the doors ever open.

A safety evaluation scores an AI model against a standardized test set of risky behaviors before deployment, so results can be compared as numbers run after run. Unlike red teaming, where people creatively hunt for new weaknesses, an eval runs the same fixed test set on every version to answer, numerically, whether this version is safer than the last.

When grown-ups say it

Safety evaluation
The tasting test before opening. The overview of this whole field. → What is AI Security?
AI red team
The inspector's surprise visit. Here's the one hunting new weaknesses creatively. → What is an AI Red Team?
Evaluation (eval)
The examiner's scorecard. The broader AI-world story about scoring. → What is Evaluation?
Benchmark
The same 100 questions, used every time. The test set a safety eval runs is exactly this.
Guardrail
The tasting checkpoint. A low score means stronger guardrails get built. → What is a Guardrail?
Hallucination
Confidently answering even when unsure. One of the things the tasting test is built to catch. → What is a Hallucination?
Fine-tuning
Reworking the recipe. A low score sends it back for rework, then re-scoring. → What is Fine-tuning?
The owner's final sign-off
The last check kept even after clearing the bar. Clearing the bar still isn't enough for risky actions. → What is a human approval gate?
Next story
The Owner's Final Sign-off →
Want another term as a picture book? Request →