Humans-Are-Not-Just-One-Thing

Systematic study of intersectional AI harms in workplace settings

Politecnico di Torino · Engineering AI Systems · 2025/2026

---

🚀 The "MarcChanges" Evolution

We have upgraded the project from a raw dataset to a causal-driven analytical pipeline. The current architecture ensures that every identity marker analyzed is not just a correlation, but a direct cause of the AI harm documented, following Kimberlé Crenshaw’s intersectionality principles.

🛠 Key Technical Upgrades

• Causal Filtering (CQ1 & CQ2): Implemented a rigorous double-check system using LLM-assisted coding to verify the causal link for every identity marker:
  • CQ1: Did this incident happen because the AI Subject was [identity value]?
  • CQ2: Would this incident still have happened if the AI Subject was not [identity value]?
• Causal Outcome Mapping: Every identified harm is now formatted following a strict causal template to ensure logical consistency:
  • Template: Because of [identity], the subject was [harmful outcome].
• Massive Data Processing: Mapped over 58,000 reports to 285 unique workplace incidents, ensuring high-quality deduplication and domain relevance.
• Automated Visualization: A custom suite (quick_analysis.py) that generates 6 distinct research graphs following our scientific framework.

---

📊 Research Framework & Results

Our pipeline answers four core Research Questions (RQs) by filtering out noise and focusing on intersectional causality.

[Phase 0] The Causal Gate (CQ1 + CQ2)

Before any analysis, we filter raw markers using the strict counterfactual questions mentioned above.

• Impact: Only if CQ1=Yes and CQ2=No, the data is retained for analysis.
• See: data/figures/graph0_CQ_filter_impact.png

[RQ1] Category Frequency

Which identities are most frequently harmed?

• Key Finding: Social Class and Race are the leading causal factors in workplace AI incidents.
• See: data/figures/graph1_RQ1_frequency.png

[RQ2] Intersectional Overlap

How do identities amplify harm?

• Key Finding: We identified specific "risk clusters" (e.g., Disability + Chronic Illness) where harm is mathematically amplified through intersectional overlap.
• See: data/figures/graph2_RQ2_heatmap.png

[RQ3] Media Representation Bias

Does the press accurately report these harms?

• RQ3a (Accuracy): 52.3% of harms are Inferred, meaning the media often obscures the specific identity of the victims.
• RQ3b (Power Dynamics): There is a stark disparity in coverage between Privileged and Oppressed groups.
• See: graph3_RQ3a_simplification.png & graph4_RQ3b_representation.png

[RQ4] High-Profile Impact

Do famous AI failures follow these patterns?

• Key Finding: High media reach (e.g., Amazon, ChatGPT cases) often correlates with high identity complexity, even if headlines only mention one factor.
• See: data/figures/graph5_RQ4_high_profile.png

---

📂 Project Structure

src/
├── batch_annotate.py   # LLM engine with causal prompts (CQ1/CQ2)
├── analysis_rq.py      # Computes mathematical scores and metrics
└── quick_analysis.py   # Generates the final research figures (RQ1-RQ4)
data/
├── workplace_reports.csv   # Raw dataset (58k+ entries)
├── annotations_v2.json     # Processed causal dataset with "Because of..." mapping
└── figures/                # Generated research graphs