Chapter 6. AI Regulation & Standards: EU AI Act, ISO/IEC 42001, C2PA

The customer's business goal

Avoid fines, pass/maintain compliance (ISO/IEC 42001), and meet transparency requirements. Promises to the business:

1. Every AI system is classified by risk, and the class determines its gates.
2. Compliance evidence (documentation, logs, evaluations, risk records) is generated automatically, not scrambled together.
3. Generated content is machine-readably labeled as AI (Art. 50).

Driver: regulator (current as of September 2026)

• Fines (Art. 99), three tiers: up to €35M / 7% of turnover (prohibited practices under Art. 5, in force since 02 Aug 2025); up to €15M / 3% (high-risk violations); up to €7.5M / 1% (incorrect information to a regulator). For SMEs — whichever figure is lower.
• Dates: Art. 5 (prohibitions) + GPAI — since 02 Aug 2025; the AI Office's supervisory powers + Art. 50 (transparency) — from 02 Aug 2026; stand-alone high-risk (Art. 6(2), Annex III) pushed by the "Digital/AI Omnibus" to 02 Dec 2027; embedded high-risk (Annex I) — to 02 Aug 2028. A delay isn't a cancellation: an RMS and documentation take years to build.
• ISO/IEC 42001 (AIMS, published Dec 2023): a de facto requirement in EU B2B RFPs; not harmonized under the AI Act → doesn't grant presumption of conformity, but does provide certifiable evidence of governance.
• NIST AI RMF: voluntary, non-certifiable; a bridge to ISO 42001.

Architectural pattern

Compliance-by-Design Automated Pipeline — compliance built into CI/CD: risk classification, the RMS, and labeling are pipeline steps, not a manual audit. Governance is expressed as policy-as-code; evidence is auto-generated artifacts.

Engineering stack & providers

• Provenance/watermarking: C2PA Content Credentials + Durable Content Credentials (soft binding: invisible watermark + perceptual fingerprint), SynthID; the draft ISO 22144.
• Policy-as-code: OPA (Rego) — gates in CI/CD.
• Compliance automation: an AI system registry (ch. 10), technical documentation templates, RMS trackers.

Engineering implementation

### Step 1. Auto-classifying risk in CI/CD

At registration/release, a system is tagged with an AI Act category (unacceptable / high-risk / limited / minimal). The class determines which gates are mandatory. For Kovcheg: the credit module → high-risk → the full gate set; the text assistant → limited → Art. 50.

### Step 2. A Risk Management System for high-risk

Regular logging of risks and mitigations (Art. 9), tied to the audit trail (ch. 4) and eval reports (ch. 7). The RMS is a living process, not a one-off document.

### Step 3. C2PA watermarking for Art. 50

Generated content (text/image/audio) carries cryptographic Content Credentials plus provenance metadata. Against metadata stripping — Durable Credentials (invisible watermark + perceptual fingerprint, the SynthID idea). The Commission's draft Code of Practice on Transparency cites C2PA as an example meeting all four Art. 50 criteria.

### Step 4. Technical documentation as an artifact

The system dossier (data, model, evaluations, risks, controls) is assembled automatically from the artifacts of chapters 1–5 and 7–10 at release time — not written after the fact.

### Step 5. A single compliance matrix

One table — "requirement → control (chapter) → evidence (artifact)" — links AI Act articles ↔ ISO 42001 controls ↔ implemented planes. This is also the core of the audit in ch. 10.

Where it breaks

• Watermarks get stripped. Paraphrasing text, screenshotting, or re-encoding kills the mark; Durable Credentials reduce this but don't eliminate it — especially weak for text. Don't promise an "unremovable" mark.
• Risk classification isn't binary. Borderline cases require legal interpretation; the automation sets a draft class, a lawyer confirms it. Auto-classification speeds things up, it doesn't replace judgment.
• ISO 42001 ≠ presumption of conformity. Certification helps a lot in RFPs and as evidence, but doesn't automatically satisfy the AI Act (it isn't harmonized).
• Delayed deadlines breed complacency. High-risk is pushed to Dec 2027 — tempting to defer preparation; but an RMS, documentation, and evals take years to build.
• Compliance theater returns. The matrix exists, but the controls don't work in production → the same gap as ch. 0. Fixed by generating evidence automatically from live planes, not screenshots.

Standards and mapping

• EU AI Act: Art. 5, 6, Annex III; Art. 9 (RMS), Art. 11–12 (documentation, logs), Art. 14 (oversight), Art. 50 (transparency), Art. 99 (fines).
• ISO/IEC 42001: AIMS, Annex A controls.
• NIST AI RMF: Govern/Map/Measure/Manage.
• C2PA / ISO 22144 (draft): provenance.

Lab and artifact

Classify both faces of Kovcheg under the AI Act; build the "requirement → control → evidence" matrix, referencing the artifacts of chapters 1–5; attach C2PA credentials to generated content and test their survivability under paraphrasing/screenshotting. Artifact: risk classification + compliance matrix + a provenance demo with an honest report on where the mark gets stripped.

Maturity checklist

• L1: systems are classified, a compliance owner is assigned.
• L2: an RMS + technical documentation as release artifacts, C2PA on output content.
• L3: policy-as-code gates in CI/CD, an AI Act↔ISO 42001 matrix with live (auto-generated) evidence, Durable Credentials, AIMS certification readiness (ch. 10).

Sources

• [EU AI Act — implementation timeline](https://artificialintelligenceact.eu/implementation-timeline/)
• [Article 99: Penalties](https://artificialintelligenceact.eu/article/99/)
• [EU AI Act and C2PA: what Article 50 requires](https://c2paviewer.com/articles/eu-ai-act-content-credentials)
• [ISO 42001 vs NIST AI RMF vs EU AI Act (EC-Council)](https://www.eccouncil.org/cybersecurity-exchange/responsible-ai-governance/eu-ai-act-nist-ai-rmf-and-iso-iec-42001-a-plain-english-comparison/)

Read next

• Next: Quality — The model neither degrades nor discriminates
• All course chapters →
• Connect the expert as an MCP →