- Amazon's AI recruiting tool discriminated against women for 3+ years before they scrapped it.
This wasn't a bug. It was the AI doing exactly what it was trained to do: replicate history.
Thread on why Black Box AI amplifies bias—and how Knowledge Graphs fix it structurally. 🧵 - #AIBias #HRTech
THE TRAINING DATA TRAP:
Amazon trained on 10 years of resumes. Tech industry = male-dominated.
AI identified pattern: "Being male" correlates with "being hired"
Statistical optimization → systematically penalize female indicators - Source: Reuters investigation 2018
SPECIFIC BIAS DISCOVERED:
✗ Penalized word "women's" (e.g., "Women's Chess Club Captain")
✗ Downgraded graduates from all-women's colleges
✗ Optimized for male-coded language patterns
Amazon engineers tried to fix it. They failed. Why? - THE PROXY VARIABLE PROBLEM:
Even after removing explicit gender terms, the AI found proxies:
- Verb choices ("executed" vs "collaborated")
- Sentence structure patterns
- Types of extracurricular activities
- Linguistic style correlations - Black Box models reconstruct bias through correlations you can't see.
THE REGULATORY RESPONSE:
NYC Local Law 144 (2023): Annual bias audits mandatory
- Selection Rate by demographic
- Impact Ratio threshold: 0.8
- Falls below? Prima facie discrimination - EU AI Act: Recruitment AI = "High-Risk"
- Explainability required
- Human oversight mandated
GDPR Article 22: Right to explanation
Candidates can demand "meaningful information about the logic involved" in automated decisions. - Black Box output: "Neuron 4,502 fired at 0.8 intensity"
Legally useless. Audit fails.
You can't comply with regulations you can't explain.
THE VERIPRAJNA SOLUTION: Knowledge Graphs
Not prediction engines. MEASUREMENT systems. - Architecture:
1️⃣ LLM extracts entities from resumes (reading)
2️⃣ Entities mapped to ontology nodes (normalization)
3️⃣ Demographic nodes EXCLUDED from inference graph (structural masking)
4️⃣ Graph algorithms calculate skill distance
DEMOGRAPHIC MASKING: - "Women's Chess Club" → LLM extracts → "Leadership Role in Strategic Activity" (gender stripped)
The inference graph has NO path from Candidate → Gender → Decision.
It's not "trying not to be biased."
It CAN'T be biased. The variable doesn't exist in the reasoning engine. - SKILL DISTANCE MATHEMATICS:
Instead of "Will they succeed?" (prediction)
We ask "How close are their skills?" (measurement) - Graph embeddings (Node2Vec/GraphSAGE):
- PyTorch vector ≈ Deep Learning vector ≈ AI vector
- Cosine Similarity measures semantic distance
- Jaccard Index measures explicit overlap
Deterministic. Auditable. Fair.
THE BUSINESS CASE: - ✓ Avoid "Amazon Moment" reputational damage
✓ Pass NYC/EU regulatory audits
✓ Expand talent pool (semantic matching finds non-pedigree candidates)
✓ Build recruiter trust (explainable = adoptable)
Black Box = legal liability.
Glass Box = competitive advantage. - 📖 Read the full technical whitepaper here: https://veriprajna.com/whitepapers/glass-box-paradigm-fair-ai-recruitment-knowledge-graphs
📧 [email protected]
🌐 https://veriprajna.com
💬 WhatsApp: +919217059957
#Veriprajna