HighLanguage Disparity●Realized
Disparate quality across languages/dialects
Fairness & BiasDescription
Markedly worse accuracy/safety in regional languages and dialects excludes or harms non-English-dominant customers.
Example scenario
The assistant handles English well but mis-advises customers in a regional language, an inclusion/SEBI concern.
Real-world evidence●Realized
Amazon's internal AI-based recruiting tool, trained on historical hiring data, learned to penalise resumes from women — a confirmed production incident where historical bias was amplified by the model. The system was used for candidate ranking before being scrapped in 2017.
Primary mitigations
- Multilingual evaluation
- language-specific guardrails
- fallback to human for low-resource languages
- localisation testing.
Detection signals
Per-language accuracy/safety dashboards; complaint analysis by language.
Mitigating controls
7 Non-agentic controls
ZYC-FAIR-001Non-Agentic
Group Fairness
Fairness
ZYC-ETH-001Non-Agentic
Non-Maleficence
Ethical Governance
ZYC-ACCU-001Non-Agentic
Lexical Similarity
Accuracy
ZYC-FAIR-008Non-Agentic
Language/Script Bias
Fairness
ZYC-FAIR-009Non-Agentic
Form Factor Bias
Fairness
ZYC-FAIR-010Non-Agentic
Document Accessibility Bias
Fairness
ZYC-FAIR-011Non-Agentic
Multi‑language Consistency
Fairness