IBI-002314 - Severepublie
Mechanistic Interpretability Bias
Biais d'interprétabilité mécaniste
IBI Score
8.3
out of 10
Severity4_grave
Frequencysystematique
Scope groups4
Detection difficulty5
🇫🇷 Définition (FR)
Difficulté à identifier les circuits neuronaux responsables des biais dans les grands modèles.
🇬🇧 Definition (EN)
Difficulty identifying neural circuits responsible for bias in large models.
TriAxis Taxonomy
Structural Classification
O
Origin
O2
Computational & Statistical
L
Lifecycle
L6
Evaluation
S
Specific
S27
Emergent bias
💡
Concrete Example
Les chercheurs ne peuvent pas localiser ni corriger précisément un biais dans un LLM de 1T paramètres.
📊 Detection Metrics
Demographic Parity DifferenceDisparate Impact RatioEqualized Odds Difference
⚙ Alert Thresholds
Disparate Impact Ratio0.8
Equalized Odds Difference0.1
Demographic Parity Difference0.1
📚 Sources & Scientific References
- •Towards Monosemanticity: Decomposing Language Models With Dictionary LearningTrenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nicholas Turner, Cem Anil, Carson Denison, Amanda Askell, Chris Olah, et al. (2023)
- •The Quest for the Right Mediator: Surveying Mechanistic Interpretability for NLP Through the Lens of Causal Mediation AnalysisAaron Mueller, Jannik Brinkmann, Millicent Li, Samuel Marks, Koyena Pal, Nikhil Prakash, Can Rager, Aruna Sankaranarayanan, Arnab Sen Sharma, Jiuding Sun, Eric Todd, David Bau, Yonatan Belinkov (2026)
🛡 Impact & Mitigation
🎯 Vulnerable groups impacted
No specific groups.
🛠 Mitigation methodologies
- AUDIT_TIERS : Third-party Audit
🏢 Sectors concerned
Media & Social Platforms (Risk: /5)
⚖ Applicable regulations
🌍 International
- •ISO_42001 : ISO/IEC 42001:2023
- •UNESCO_2021 : UNESCO - Recommandation sur l'éthique de l'IA
- •OECD_AI_PRINCIPLES : OCDE - Principes IA
- •COE_AI_CONVENTION_INT : Conseil de l'Europe - Convention IA
📋 Metadata
Contributor
Mohamadou SIDIBE
Institution
AI BIAIS INITIATIVE
Version
1
Created on
October 14, 2026
Updated on
November 4, 2026
UUID
12e6f645-3412-488d-88ed-107be38f50a7
Explore other biases
250 documented biases, classified by the TriAxis taxonomy.