Skip to main content
The bias check post-hook screens LLM responses for discriminatory language across all 15 protected characteristics defined in the EU Charter of Fundamental Rights, Article 21.

​
15 Protected Characteristics

CharacteristicExample PatternsSeverity Range
sexGender stereotypes, role enforcement, selection biasMEDIUM–HIGH
raceRacial generalizations, supremacy languageMEDIUM–CRITICAL
colourSkin colour capability biasHIGH
ethnic_originEthnic generalizations, xenophobic directivesMEDIUM–HIGH
social_originClass-based bias, preferential treatmentMEDIUM
genetic_featuresGenetic determinism, genetic screening for accessHIGH–CRITICAL
languageLanguage-based competence biasLOW–MEDIUM
religionReligious stereotypes, discrimination directivesMEDIUM–HIGH
political_opinionPolitical stereotyping, discriminationMEDIUM–HIGH
national_minorityMinority exclusion, group slursHIGH–CRITICAL
propertyWealth-based discrimination, access denialMEDIUM–HIGH
birthBirth-based discrimination, origin prejudiceMEDIUM
disabilityCapability stereotypes, exclusion, slursHIGH–CRITICAL
ageAge-based capability bias (both older and younger)MEDIUM–HIGH
sexual_orientationPathologization, discriminationHIGH–CRITICAL
nationalityNationality stereotypes, discriminationHIGH

​
Severity Weights

Each pattern has a severity level that determines its contribution to the aggregate score:
SeverityWeightDescription
LOW0.1Stereotyping, mild generalizations
MEDIUM0.3Differential treatment language
HIGH0.6Discriminatory outcome language
CRITICAL1.0Explicit slurs, hate speech, supremacy

​
Scoring Formula

pattern weight × domain multiplier → per-finding score
sum(per-finding scores) → aggregate score
aggregate score > threshold → action (warn or block)

​
Domain Profiles

Each domain profile sets a default threshold and weight multipliers for characteristics that are especially sensitive in that context.

​
General (default)

  • Threshold: 0.3
  • Multipliers: None (all at 1.0×)

​
HR / Recruitment

The strictest profile — HR is classified as high-risk under EU AI Act Art.6(2).
  • Threshold: 0.15
  • 2.0× multipliers: sex, race, age, disability, ethnic_origin
  • 1.5× multipliers: religion, sexual_orientation, nationality, political_opinion
import { complior } from '@complior/sdk';
import OpenAI from 'openai';

const client = complior(new OpenAI(), {
  domain: 'hr',
  biasAction: 'block',
  // Threshold auto-set to 0.15, multipliers applied
});

​
Finance

Strict on property and social origin for credit/loan decisions.
  • Threshold: 0.2
  • 2.0× multipliers: property, social_origin
  • 1.5× multipliers: race, nationality, birth

​
Healthcare

Strict on disability and genetic features for medical decisions.
  • Threshold: 0.2
  • 2.0× multipliers: disability, genetic_features
  • 1.5× multipliers: age, race, sex

​
Education

Strict on age, disability, and social origin for admissions/grading.
  • Threshold: 0.2
  • 2.0× multipliers: age, disability, social_origin
  • 1.5× multipliers: language, ethnic_origin, birth

​
Configuration

import { complior } from '@complior/sdk';
import OpenAI from 'openai';

const client = complior(new OpenAI(), {
  domain: 'finance',       // Activates finance profile
  biasThreshold: 0.15,     // Override profile default (0.2)
  biasAction: 'block',     // Throw BiasDetectedError
});
FieldTypeDefaultDescription
domainDomainundefinedActivates domain profile (sets threshold + multipliers)
biasThresholdnumberProfile defaultOverride threshold (0–1, lower = stricter)
biasAction'warn' | 'block''warn'Action on threshold exceeded

​
BiasEvidence

When a finding is detected, it produces a BiasEvidence object: 
interface BiasEvidence {
  readonly characteristic: string; // e.g., 'sex', 'race'
  readonly severity: string;       // 'LOW' | 'MEDIUM' | 'HIGH' | 'CRITICAL'
  readonly evidence: string;       // Matched text
  readonly score: number;          // weight × domain multiplier
}
In warn mode, findings are available in response metadata. In block mode, they are on the BiasDetectedError.

Error Handling

BiasDetectedError reference.

Safety Filter

Content safety patterns and HITL gate.