IBI-000385 - Criticalpublie
Biased RLHF Preference Bias
Biais d'apprentissage par renforcement humain biaisé
IBI Score
9.2
out of 10
Severity5_critique
Frequencysystematique
Scope groups5
Detection difficulty4
🇫🇷 Définition (FR)
Biais introduits par les préférences des annotateurs RLHF, souvent non représentatifs.
🇬🇧 Definition (EN)
Bias introduced by RLHF annotator preferences, often non-representative.
TriAxis Taxonomy
Structural Classification
O
Origin
O3
Human-Cognitive
L
Lifecycle
L5
Learning
S
Specific
S05
Annotation bias
💡
Concrete Example
Les annotateurs RLHF de GPT-4 (majoritairement jeunes, américains) introduisent des biais culturels dans le modèle.
📊 Detection Metrics
Annotation AgreementDemographic Parity Difference
⚙ Alert Thresholds
Annotation Agreement0.8
Demographic Parity Difference0.1
📚 Sources & Scientific References
- •Open Problems and Fundamental Limitations of Reinforcement Learning from Human FeedbackStephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, J�r�my Scheurer, Javier Rando, Rachel Freedman et al. (32 auteurs) (2023)
- •
🛡 Impact & Mitigation
⚠ Documented evidence (Studies & Incidents)
- OpenAI rolls back ChatGPT's sycophancy and explains what went wrong : Source : https://venturebeat.com/ai/openai-rolls-back-chatgpts-sycophancy-and-explains-what-went-wrong
Impact:
🎯 Vulnerable groups impacted
Minority language speakersRural areas
🛠 Mitigation methodologies
- AUDIT_TIERS : Third-party Audit
🏢 Sectors concerned
Media & Social Platforms (Risk: /5)
⚖ Applicable regulations
🇪🇺 Europe
- •AI_ACT_10 : AI Act (UE) 2024/1689
- •AI_ACT_15 : AI Act (UE) 2024/1689
- •AI_ACT_9 : AI Act (UE) 2024/1689
🇺🇸 United States
- •US_CO_SB24_205 : Colorado AI Act (SB 24-205)
- •US_NIST_AI_RMF : NIST AI Risk Management Framework 1.0
🌍 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
🌐 autre
- •SG_AI_GOVERNANCE : Singapour - AI Governance Framework
📋 Metadata
Contributor
Mohamadou SIDIBE
Institution
AI BIAIS INITIATIVE
Version
1
Created on
October 14, 2026
Updated on
November 4, 2026
UUID
c0ebe062-1f6b-4d0a-8818-eee3089d6374
Explore other biases
250 documented biases, classified by the TriAxis taxonomy.