EU AI Act · Article 50 · applies 2 August 2026
Everyone explains the EU's AI transparency rules. This is the only guide that compiles.
The regulation
"Providers of AI systems […] generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated."
Regulation (EU) 2024/1689, Article 50(2) · fines up to €15M or 3% of global turnover
The implementation — real output, verified July 2026
$ node verify.js signed.jpg claim_generator: pixelsmith/1.0.0 c2pa-rs/0.49.2 action: c2pa.created digitalSourceType: …/trainedAlgorithmicMedia signed_at: 2026-07-24T17:14:58+00:00 validation_status: [] # clean ✓
Left: what the law demands. Right: this book's own test suite satisfying it. Every snippet in the book ran before it was printed.
Written for the developer who got the "can you handle this?" email. Each chapter: the law in one page → a decision tree → working code → edge cases → the evidence that proves compliance.
Early bird · pre-order
€29 €89
Ships September 2026. Price rises to €89 (bundle) / €39 (ebook only) at launch.
Pre-order for €29Payments handled by Polar (merchant of record — VAT sorted, invoice included). Not happy at delivery? Full refund, keep the templates.
Writing code before printing it finds the things blog posts miss. Two examples from the book's own test runs:
The C2PA Node SDK's tsaUrl reads as optional. Omit it and signing dies with failed to downcast any to string. The book documents it — and ships a local timestamp server so your CI never depends on someone else's uptime.
A single quality-90 re-encode strips a signed Content Credential completely — verified, output printed in Chapter 5. Which is exactly why the book builds the layered marking the law's "robust" actually requires.
The high-risk regime was delayed (to December 2027 / August 2028). Article 50's transparency duties were not — they apply 2 August 2026, with one narrow grace period: machine-readable marking for generative systems already on the market before that date gets until 2 December 2026. Chapter 1 sorts your exact exposure in 30 minutes.
Four possible: AI-interaction disclosure (chatbots), machine-readable marking of generated media/text, emotion-recognition/biometric disclosure, and deepfake/public-interest-text labeling. Provider vs deployer matters. The free triage excerpt answers it — subscribe below and it lands in your inbox.
No — it's an implementation guide written by an engineer, mapped to the Act, the Commission's final July 2026 Guidelines, and the Code of Practice, with sources cited at article level. It makes you dangerous in meetings, not admissible in court. For bet-the-company calls, pair it with counsel.
They will — watermarking is a moving target. Every buyer gets 12 months of free updates and a versioned repo. The book is maintained like software, because it is one.
September 2026. Chapters 1 and 5 are already written (excerpts on the Substack). If it slips or disappoints, full refund, no forms, keep the templates.
The 30-minute obligation check, straight to your inbox — plus one build-log email per week until launch. Unsubscribe anytime, obviously.