ComplianceDataLab/eu-ai-act-compliance-scenarios

EU AI Act Compliance Scenarios Dataset

⚠️ EU AI Act high-risk obligations become enforceable on 2 August 2026. Companies building compliance tools have zero labeled training data. This dataset fills that gap.

🔓 This is the FREE sample (75 records)

This sample contains 75 scenarios from the Employment/HR AI vertical. Each record includes a realistic deployment scenario, article-level compliance analysis, GDPR intersections, and hidden compliance traps.

Need more? The full dataset is available:

	Free Sample	⭐ Starter Pack
Records	75	500
Sectors covered	27	30+
AI system types	14	14
License	CC BY-NC 4.0 (non-commercial)	CC BY 4.0 (full commercial)
Use in commercial products	❌	✅
Fine-tune commercial models	❌	✅
Price	Free	$97
	↓ Scroll to explore	Buy Starter Pack →

---

What's inside each record?

Every scenario is a structured JSON object with ground truth labels — directly usable for supervised learning:

{
  "scenario": {
    "title": "Repurposed Surveillance AI for Automated Job Candidate Rejection",
    "description": "A consulting firm repurposed an employee monitoring AI...",
    "system_type": "automated_rejection",
    "sector": "consulting",
    "geography": "Italy",
    "scale": "SME (120 employees)",
    "data_processed": ["emotion_detection", "ethnicity_inferred", ...]
  },
  "compliance_analysis": {
    "classification": { "is_high_risk": true, "legal_basis": "Annex III Area 4(a)..." },
    "obligations": [
      {
        "obligation": "risk_management_system",
        "article": "Article 9",
        "status": "non_compliant",
        "finding": "No risk assessment conducted when repurposing system...",
        "severity": "critical"
      }
    ],
    "gdpr_intersections": [...],
    "penalty_exposure": { "ai_act": "Up to EUR 15M or 3%...", "gdpr": "Up to EUR 20M or 4%..." }
  },
  "complexity_score": 8,
  "traps": ["System repurposing without new conformity assessment creates cascading failures..."]
}

Use Cases

For ML Engineers: Fine-tune compliance classifiers, build risk scoring models, benchmark NLP systems against the traps field for reasoning capability.

For RegTech Companies: Train your product's AI engine with labeled compliance examples. Use as demo data showing how your tool handles complex cases.

For Compliance Teams: Run training exercises — present scenarios to your team, have them identify violations before checking the ground truth.

Dataset Statistics

Metric	Free Sample	Starter Pack
Total records	75	500
Sectors covered	27	30+
AI system types	14	14
Avg obligations/record	4.72	4.78
Avg GDPR intersections/record	3.07	3.06
Avg traps/record	3.04	3.01
Quality score (adversarial eval)	8.57/10	8.57/10

Domain Coverage

AI System Types: CV screening, video interview analysis, psychometric assessment, social media screening, employee monitoring, performance scoring, attrition prediction, salary recommendation, interview scheduling, workforce planning, bias detection, automated rejection, internal mobility, contractor classification.

Industries: 27+ sectors including banking, insurance, healthcare, automotive, aerospace, legal services, e-commerce, fintech, government, university, and more.

Compliance Areas: Articles 9-15 of the EU AI Act (risk management, data governance, technical documentation, record-keeping, transparency, human oversight, accuracy & robustness), plus GDPR intersections (Articles 9, 22, 35, 5(1)(c), 13/14, 25).

Generation Methodology

Generated using an adversarial pipeline with quality assurance:

1. Parameter Randomization — unique combination of sector, company size, geography, AI type, compliance level per scenario
2. Generation — Claude API with EU AI Act domain knowledge
3. Structural Validation — Pydantic schema enforcement
4. Independent Quality Evaluation — second AI pass scoring legal accuracy, realism, complexity, completeness, uniqueness
5. Quality Gate — only scenarios scoring ≥ 7.0/10 accepted

Get the Full Dataset

The free sample gives you a taste. The Starter Pack gives you enough data to build real products.

→ Buy Starter Pack (500 records, commercial license) — $97

Coming soon:

• Credit Scoring vertical (Annex III Area 5) — Q2 2026
• Biometrics vertical (Annex III Area 1) — Q3 2026
• Enterprise Pack (5,000+ records, all verticals) — Q3 2026

Limitations

This dataset is synthetically generated. Important caveats:

• Compliance determinations are plausible interpretations, not validated legal opinions. The EU AI Act lacks established case law as of 2026.
• The dataset is designed for ML training and benchmarking, not as legal guidance for real compliance decisions.
• Generated scenarios may exhibit linguistic patterns typical of LLM output. We recommend using this data for structural training rather than stylistic learning.
• Severity ratings and penalty calculations are reasonable estimates, not authoritative legal assessments.

For real compliance decisions, always consult qualified legal professionals.

License

• Free sample (this dataset): CC BY-NC 4.0 — free for research, evaluation, benchmarking
• Starter Pack: CC BY 4.0 — full commercial use rights included

Citation

@dataset{eu_ai_act_compliance_scenarios_2026,
  title={EU AI Act Compliance Scenarios Dataset},
  year={2026},
  publisher={Hugging Face},
  license={CC-BY-NC-4.0}
}