dsgvo-ai-act-labeling

Apply GDPR workflows and EU AI Act Article 50 disclosure decisions for Ottili flows: processing register entries, human-review evidence and AI labeling. Use when an agent must classify a data processi

SKILL.md

GitHub

dsgvo-ai-act-labeling

When to use this skill

Use this skill when an agent must sort a data processing activity into its GDPR bucket, work out whether an AI system triggers an EU AI Act Article 50 disclosure that needs human-review evidence, or keep the GDPR processing register up to date for Ottili flows. Do not use it for general privacy policy text, non-EU AI systems, or member-state-specific law beyond the EU baseline.

Procedure

  1. GDPR gate. Capture controller, purpose, data subjects and categories, recipients, retention and security measures (Art. 30(1)). Even an activity with no personal data gets an entry, marked personal_data: false, so the decision is auditable. Exactly one Art. 6(1) legal basis per activity; the script rejects bases it does not recognise.
  2. Register. Append the entry to the JSONL register. Editing an entry in place breaks GoBD immutability, so never do it. When something changes, append a fresh entry with a new recorded_at and a note that points at the entry it supersedes.
  3. AI Act gate. Run four gates in order: (1) no AI → not-ai, no Art. 50; (2) transparency-only (chatbot, emotion-recognition, deep-fake, biometric-categorization) → Art. 50(1)-(4) disclosure with human review; (3) high-risk / Annex III (Art. 6) → Art. 50(5), human review mandatory together with an evidence file whose digest the script computes; (4) general AI → Art. 95, document only. Unknown ai_type blocks rather than inventing a class.
  4. Human-review evidence. Record reviewer_id, an ISO-8601 reviewed_at and evidence_path. For high-risk the file must exist when the label is produced and the label carries its sha256 digest. A review has to name the system, the risk class and the mitigations checked — a checkbox with no content behind it counts as missing.
  5. Label and store. Emit the machine-readable label next to the system artefact. Re-run classification on every material change and every release cycle; a label emitted for v1 is not valid for v2.

Decision tables

AI Act risk classes

ai_type Risk class Art. 50 Human review
none not-ai not required n/a
chatbot, emotion-recognition, deep-fake, biometric-categorization transparency-only 50(1)-(4) required
high-risk / Annex III (Art. 6) high-risk 50(5) required, evidence file mandatory
anything else general not required not required

GDPR legal bases (Art. 6(1))

Key Label
art6_1a_consent Consent (Art. 6(1)(a))
art6_1b_contract Contractual necessity (Art. 6(1)(b))
art6_1c_legal Legal obligation (Art. 6(1)(c))
art6_1d_vital Vital interests (Art. 6(1)(d))
art6_1e_task Public task (Art. 6(1)(e))
art6_1f_legit Legitimate interests (Art. 6(1)(f))

Pitfalls from research

  • Never auto-approve high-risk. A missing evidence file blocks the label; the script raises and the agent must not bypass it.
  • Transparency-only systems still need human review. A chatbot label with no reviewer is rejected, not warned.
  • Re-run classification on material change. A label emitted for v1 is not valid for v2; re-verify on every release cycle.
  • Inherit GoBD retention from gobd-archive. Label artefacts are business records under HGB/AO; coordinate retention timing with that skill.
  • Member-state additions are unverified. Germany (BDSG), France and the Netherlands may add requirements beyond the EU baseline; flag them until local counsel confirms.

Verification checklist

  • Processing activity classified; personal data present or not
  • Art. 30 entry complete; legal basis valid
  • Register is append-only; never edited in place
  • AI Act gate run; risk class determined
  • High-risk label has reviewer, date and evidence file with checksum
  • Transparency label has reviewer and date
  • Label re-verified on release change; no stale reuse
  • Retention inherited from gobd-archive

Near-miss triggers (stop and re-read)

  • "I'll auto-approve this high-risk label" → BLOCKING; evidence file required.
  • "A chatbot doesn't need a reviewer" → transparency-only still requires one.
  • "The label from v1 is still valid" → re-run on every release cycle.
  • "BDSG doesn't apply to us" → member-state additions are unverified until local counsel confirms.

Related skills

gobd-archive — processing-register entries and human-review evidence are correspondence (6y)

References

Worked examples

Good output — no-AI system, no disclosure

Input: invoice-export with ai_type: "none", a plain export script. scripts/disclosure_label.py --system system.json emits: {"system": "invoice-export", "risk_class": "not-ai", "article_50_required": false, "human_review_status": "n/a", "label": "no-ai-act-scope"} and exits 0. No reviewer, no evidence file. The decision is recorded so the Art. 30 register can show why no label was needed.

Bad output — transparency chatbot with no reviewer

Input: support-bot with ai_type: "chatbot" and an empty human_review block. scripts/disclosure_label.py exits 2 with: transparency system 'support-bot' requires human-review evidence Fix: add reviewer_id and an ISO-8601 reviewed_at. A chatbot is transparency-only under Art. 50(1)-(4) and still needs a human reviewer — the label is not optional just because the risk class is low.

Good output — chatbot label with reviewer

Input: support-bot, ai_type: "chatbot", reviewer ops-1, reviewed_at: "2026-10-09T10:00:00Z", no evidence file. The label carries risk_class: "transparency-only", article_50_paragraph: "1-4", human_review_status: "approved", label: "article-50-disclosure". Exit 0.

Bad output — high-risk auto-approved without evidence

Input: recruiter-ai with ai_type: "high-risk", reviewer set, but no evidence_path and no evidence file. scripts/disclosure_label.py exits 2 with: high-risk system 'recruiter-ai' BLOCKING: human-review evidence missing Fix: produce the evidence file (the documented review record naming the system, the risk class and the mitigations checked) and re-run. A high-risk label with no evidence file is never issued — the script blocks, and the agent must not bypass the block.

Good output — high-risk label with checksummed evidence

Input: recruiter-ai, ai_type: "high-risk", reviewer lawyer-1, reviewed_at: "2026-10-09T11:00:00Z", evidence_path: "review-1.md". The label carries risk_class: "high-risk", article_50_paragraph: "5", evidence_sha256: <digest of review-1.md>, retention_years: 10, human_review_status: "approved". The digest lets an auditor re-derive the file's identity without re-running the classification.

Bad output — unknown ai_type classified as general AI

Input: fraud-detector with ai_type: "fraud-detection", a real system that is neither in the transparency set nor marked high-risk. scripts/disclosure_label.py exits 2 with: unknown ai_type 'fraud-detection' for system 'fraud-detector'; choose from none, ['biometric-categorization', 'chatbot', 'deep-fake', 'emotion-recognition'], high-risk, 'Annex III high-risk (Art. 6)', general Fix: decide the risk class explicitly (it may well be Annex III — fraud screening is a candidate) and set ai_type accordingly. Never let an unknown type fall through to "general AI": that silently documents a decision nobody made.

Good output — general AI documented, not disclosed

Input: report-ml with ai_type: "general", an internal analytics model. The label carries risk_class: "general", article_50_required: false, article_95_documented: true, human_review_status: "not-required", label: "ai-act-general". Exit 0. The decision is documented under Art. 95 even though no disclosure label is emitted.

Bad output — label from v1 reused for v2

Input: the v1 label for recruiter-ai is copied unchanged when the model is retrained to v2. The skill rejects this: a label emitted for v1 is not valid for v2. Re-run scripts/disclosure_label.py on every material change and every release cycle; re-verify the evidence file's checksum, since the review may no longer describe the new model.

Good output — entry superseded, never edited

Input: the retention period for an activity changes from 6y to 10y. The register is append-only: a new entry is appended with a fresh recorded_at and a supersedes note pointing at the earlier entry id. The earlier line is never modified. scripts/process_register.py always sets recorded_at from datetime.now(timezone.utc) — it never copies the caller's timestamp, so a caller-supplied recorded_at cannot fake the audit trail.