Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions

2026-05-18 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, FinTech & Digital Financial Services · Depth: Expert, extended

Summary

A study on instruction-tuned large language models (LLMs) like Gemma-3-12B-IT, Llama-3.1-8B-Instruct, and Qwen2.5-14B-Instruct reveals that while they exhibit behavioral fairness in high-stakes decisions like mortgage underwriting, they retain and amplify biased associations in their internal representations. Using a synthetic dataset of 1500 paired mortgage applications differing only in racially-associated names, the researchers found no output-level bias (e.g., approval rates of 27.27% for White-associated names vs. 27.13% for Black-associated names in Gemma-3-12B-IT). However, internal demographic representations increased monotonically from 0 to ~1200 in Gemma-3's penultimate layer. Through activation steering and cross-layer interventions, this suppressed information was shown to be decision-relevant, causing near-complete decision reversals when reinjected at critical layers. This latent bias is asymmetric, affecting decisions in one demographic direction more than the reverse, and is susceptible to adversarial prompt engineering and parameter-efficient fine-tuning with fewer than 6,000 parameters.

Key takeaway

For CTOs and VPs of Engineering deploying LLMs in high-stakes financial services, relying solely on output-based fairness audits is insufficient and risky. You must integrate mechanistic interpretability techniques, such as activation steering and representational analysis, into your AI governance frameworks. This will help detect hidden, asymmetric biases that could be exploited through adversarial prompting or minimal fine-tuning, ensuring robust and truly fair model deployment.

Key insights

Fair LLM outputs can mask exploitable, causally potent, and asymmetrically biased internal representations.

Principles

• Behavioral audits alone are insufficient for LLM safety.
• Latent bias can be asymmetric and model-dependent.
• Suppression, not elimination, creates vulnerabilities.

Method

The study used a matched-pair design with synthetic mortgage applications, representational analysis, activation steering (including cross-layer), prompt engineering, and parameter-efficient fine-tuning to expose latent bias.

In practice

• Implement dual-layer testing frameworks for AI governance.
• Combine output evaluation with representational analysis.
• Guard against prompt engineering and fine-tuning attacks.

Topics

• Large Language Models
• Algorithmic Fairness
• Latent Bias
• Activation Steering
• Mortgage Underwriting

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, AI Ethicist, AI Security Engineer

Related on AIssential

• Exposing the Unsaid: Visualizing Hidden LLM Bias through Stochastic Path Aggregation — TreeTracer visualizes hidden LLM biases by aggregating stochastic generation paths and comparing semantic contexts through a novel Sankey adaptation.
• The Next Evolution of Fairness: Why LTAE and LAAM are expanding into AI Safety. — AI safety's structural failures, mirroring human oversight issues, necessitate governance frameworks like SLAAM to prevent asymmetric enforcement and anchor drift.
• Evaluating Second-Order Bias of LLMs Through Epistemic Entitlement — LLMs exhibit "second-order bias" in judging social bias, a subtle form missed by current evaluation methods.
• Large language models demonstrably withhold, soften, or suppress true information about specific people and topics, implemented at every layer... — LLMs actively suppress true information through various technical layers, making legitimate and illegitimate censorship indistinguishable to users.
• Contaminated Collaboration: Measuring Gender Bias Transfer in LLM-Assisted Student Writing — Gender bias from LLM writing assistants transfers to human-produced text, particularly suppressing agency in female-targeted writing.
• Adversarial Feeds Steer LLM Agent Decisions Against Their Defaults — External information feeds significantly influence LLM agent decisions, acting as a control surface independent of the core model.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.