Safety icons

The 58 safety icons in AI & LLM, described

Abstainabstain-answer

A speech bubble with a shrug in it — the model saying it does not know.

Adversarial inputadversarial-input

The wide machine with an arrow that turns into a bolt — input built to make the model fail.

Alignment taxalignment-tax

Axes with a line that steps down — what a model gives up in capability to stay safe.

Annotation taskannotation-task

A clipboard with an item, its label and a tick — one thing labelled by a person.

Benchmark suitebenchmark-suite

A clipboard with bars on it — the set of tests a model's score comes from.

Biasbias

Bias — the scales are not even, a skew in data or outcomes that fairness work corrects.

Bias auditbias-audit

A clipboard with two groups' bars under one line — whether the model treats them the same.

Bias checkbias-check

A bias check — does the model treat like cases alike, a fairness audit.

Calibration curvecalibration-curve

Axes with the diagonal and points either side of it — whether confidence matches accuracy.

Canary promptcanary-prompt

A speech bubble with a flag in it — a prompt planted so its answer shows up where it should not.

Chain verificationchain-verify

A clipboard with two checks in a row — each step of the reasoning checked in turn.

Confidence thresholdconfidence-threshold

Axes with a line and points above and below it — answers kept only when the model is sure enough.

Consistency checkconsistency-check

A speech bubble with an equals sign — the same question asked twice, answered the same.

Contamination checkcontamination-check

The flask with a drop that should not be there — test data found in the training set.

Content filtercontent-filter

A content filter — the moderation screen that decides what is not allowed through.

Demographic paritydemographic-parity

Axes with three bars the same height — the same rate for every group.

Error analysiserror-analysis

A clipboard with a lens on it — the wrong answers, looked at one by one.

Escalate to humanescalation-human

The alarm bell with a head in it — the case handed to a person.

Eval costeval-cost

A receipt with a coin on it — what an evaluation run costs to make.

Eval dataseteval-dataset

A folder with bars in it — the examples a score is measured on.

Eval reporteval-report

A clipboard with a line and a check under it — the results, written up.

Eval sliceeval-slice

A clipboard with brackets on it — one part of the eval, looked at on its own.

Eval traceeval-trace

A receipt with a check on it — the log of a run, with its verdict.

Factuality testfactuality-test

The flask with a check in it — the model's claims put to the test.

Failure taxonomyfailure-taxonomy

A clipboard with a tree on it — the ways a model fails, sorted into kinds.

Fairness metricfairness-metric

Axes with two bars the same height — one number for whether groups are treated alike.

Gold setgold-set

A coin landing in a tray — the examples whose answers are known to be right.

Groundedness scoregroundedness-score

A speech bubble with a post standing on the ground — how much of the answer rests on the sources.

Guardrailguardrail

A guardrail — a safety limit the model's output is not allowed to cross.

Input guardrailguardrail-input

Two rails with an arrow entering between them — what comes in, checked before the model sees it.

Output guardrailguardrail-output

Two rails with an arrow leaving between them — what goes out, checked before the user sees it.

Hallucinationhallucination

A hallucination — an answer said with confidence that is true of nothing, invented.

Hallucination detectionhallucination-detect

An eye seeing a wisp — a claim the sources never made, caught.

Human preferencehuman-preference

A scale tipped by one weight — the answer a person liked better.

Human review queuehuman-review-queue

A head waiting over a tray — outputs lined up for a person to look at.

Inter-annotator agreementinter-annotator

A clipboard with two lines that match — how often two labellers say the same.

Jailbreak attemptjailbreak-attempt

An open lock with a bolt in it — a prompt written to get past what the model refuses.

Judge modeljudge-model

A scale come down on one side — a model that weighs two answers and picks one.

Judge rubricjudge-rubric

A clipboard with a small scale on it — what the judge is told to weigh.

Label qualitylabel-quality

A tag with a set stone on it — how good the labels are.

Leakage checkleakage-check

The flask with something seeping below the line — labels that crept into the inputs.

Oversight loopoversight-loop

The rotation loop with an eye in it — a person looking, every time round.

Policy violationpolicy-violation

A shield with a bar broken in two — a rule the output crossed.

Prompt injection blockedprompt-injection-block

The wide machine with a bolt stopped at a line — an instruction smuggled in, and not obeyed.

Refusalrefusal

A refusal — the model declined the request, it will not do that.

Refusal raterefusal-rate

Axes with three short bars — how often the model declines.

Refusalrefusal-response

A speech bubble with a flat line in it — the model declining to answer.

Reward hackingreward-hacking

A machine frame with the reward arrow struck through — a high score reached by cheating.

Rubric scorerubric-score

A clipboard with criteria and a mark against each — a score built up line by line.

Safe completionsafe-completion

A speech bubble with a shield in it — an answer given, kept within the lines.

Safety categorysafety-category

A tag with a shield on it — the kind of harm a piece of content is filed under.

Safety checksafety-check

A safety check — output looked at and screened before it goes out.

Severity labelseverity-label

A tag with two chevrons on it — how bad, on a scale.

Slice analysisslice-analysis

Axes split by a line with a point each side — the score, broken down by group.

Specification gamingspecification-gaming

An open lock with a check in it — the test passed by a door nobody meant to leave open.

Toxicitytoxicity

Toxicity — harmful or abusive content that should not have been said, flagged.

Toxicity scoretoxicity-score

A speech bubble with a hazard diamond in it — how harmful what was said is.

Uncertainty flaguncertainty-flag

A speech bubble with a question mark in it — an answer the model marks as unsure.

In code, each is one import — import { AbstainAnswer } from "@iconmind/react/icons/abstain-answer" — and the same name in Vue, Svelte, Solid, Preact, React Native, Astro, Blade and Flutter.

Narrow it by tag