RESEARCH · GLOBAL
Fair Outputs, Biased Internals: Latent Bias in LLMs for High-Stakes Decisions
Research reveals that while instruction-tuned large language models (LLMs) can show behavioral fairness in high-stakes decisions like mortgage underwriting, they retain and even amplify biased associations in their internal representations. This disconnect raises concerns about 'latent bias' and its potential causal potency.
WHY IT MATTERS
Undetected latent bias in LLMs poses significant compliance and ethical risks for BFSI, demanding deeper inspection beyond output fairness to ensure equitable financial outcomes.
Source: arXiv · 2026-05-18