Bias Amplification in AI Models
UX research to identify and mitigate amplification of human perceptual bias in generative models.
Project Overview
Role
Lead researcher - study design, audit workflow, synthesis.
Timeline
Proposal to ongoing analysis - audit integrated pre-launch.
Methods
Human-centered AI, prompt ladders, human baseline comparison.
Tools
Gemini, ChatGPT, DALL·E, R, Qualtrics.
Problem
Models amplify perceptual defaults - lowering trust and creating product risk.
Goal
Move from reactive fixes to a pre-launch audit - quantify amplification, diagnose drivers, add guardrails.
Impact metrics
Amplification score
Computed after the comparison dataset is complete.
Mitigation coverage
Audit steps integrated into pre-launch checks.
Decision outcomes
Prompts and copy reviewed before ship.
Process
1. Human baseline
Use prior categorization data to set a measurable baseline.
2. Model prompts
Test vague-to-specific prompts to observe prototype defaulting.
3. Compare and quantify
Compare outputs to the baseline to estimate amplification.
Research insights
Prototype defaulting
Vague prompts aligned with an East-Asian prototype - echoing human defaults.
Prompt specificity
Enumerating sub-groups increased diversity - stronger at higher specificity.
Audit before launch
A lightweight loop flags amplification early so teams choose fairer defaults.
Model results
Deliverables
Bias audit framework
Standardizes prompt ladders, review gates, documentation.
Fairness checklist
Pre-launch decision gates for prompts, data notes, UI copy.
Reproducible workflow
Repeatable documentation for cross-team audits.
Reflection
Bias amplification is a product-quality issue. A light audit loop helps teams ship with clarity.
Ready to audit bias in your AI launch?
Use this framework to compare human and model outputs and reduce risk before release.
Download My Sample Pre-Launch Audit