Responsible AI Risk Taxonomy v2.0

This repository publishes an interactive Responsible AI risk taxonomy for Physical AI Risks: embodied AI, humanoids, robots, drones, autonomous vehicles, cyber-physical systems, and other AI systems that perceive, decide, and act in the physical world.

Live page:

• https://deep1003.github.io/Physical-AI-Risk-Taxonomy/

Contents

• index.html — master public GitHub Pages entry point for the interactive taxonomy.
• docs/pai_risk_taxonomy_bilingual_v2.0.html — versioned copy synced from index.html with docs-relative asset paths.
• data/l4_cards.csv and data/l4_cards.json — extracted L4 card metadata.
• data/l4_references.csv and data/l4_references.json — card-level reference and justification table.
• data/taxonomy_migrations.json — human-approved hierarchy migration notes.
• data/three_h_one_r_primary_exceptions_20260628.* — cards that retain three 3H1R Primary axes with explicit reasons.
• data/reference_addition_rounds/ — candidate selection, access checks, applied additions, and removed candidates from the reference-enrichment rounds.
• data/reference_quality_audit/ — low-relevance reference audit and cleanup logs.
• scripts/ — scripts used to add corpus references and audit reference relevance.
• methodology/methodology.md — concise methodology and curation notes.
• expert-survey/ — frozen baseline expert survey, taxonomy version 2.1.
• expert-survey-revised/ — parallel survey using revised L4 wording and robust EM-supported L3 assignments, taxonomy version 2.2.

Current Snapshot

• L4 cards: 182
• Evidence/justification entries: 360
• Linked references: 353
• Maximum references per L4 card: 5
• L2 card counts: P2 System Safety=92, I2 Interaction Safety=60, S2 Societal Safety=30

The 2026-07-31 wording audit revised 84 L4 labels or definitions. A constrained EM sensitivity analysis using three embedding models, bilingual and English-only inputs, four L3 seed weights, and 1,000 bootstrap replicates supported two L3 moves: PHYSBENCH-REF-0065 from I3.7 to P3.2, and PHYSBENCH-REF-0107 from I3.6 to S3.6. All 24 L3 categories remain populated.

Taxonomy IDs

The current public slice is organized as L1 Physical AI Risks > L2 Categories > L3 Sub-categories > L4 Risk Cards. The three L2 children use short stable IDs: P2 for System Safety, I2 for Interaction Safety, and S2 for Societal Safety.

L3 categories use the same domain prefix and the L3 level number, followed by a local sequence number. For example, P3.1 means the first L3 category under System Safety. L4 risks keep their original card IDs, such as PHYSBENCH-REF-0017.

Human decision on 2026-06-28 moved S3.10 Lack of Robustness in Unseen Environments to P3.9. The category was moved only; absorption into P3.2 or redistribution remains a future human decision.

Human decision on 2026-07-04 realigned the L2/L3 hierarchy to the master structure: P2 System Safety has 5 L3 categories, I2 Interaction Safety has 10 L3 categories, and S2 Societal Safety has 9 L3 categories. The affected L3 groups were moved as groups, and their L4 cards followed their L3 parent.

Scope

The taxonomy focuses on Physical AI risks where model behavior can propagate into physical action, sensor-mediated decisions, human-robot interaction, embodied surveillance, safety control, certification, and lifecycle governance.

The repository is separate from AI_Topic_Space.github.io; future Physical AI risk taxonomy work should happen here.