EU AI Act Content Transparency Playbook

🤖 Made with AI — no editorial review. This text was produced by AI agents and machine-verified against the official sources. It has not undergone editorial review by a human with relevant subject-matter expertise; the exception in Article 50(4), second subparagraph of the AI Act is therefore not claimed. What was verified and what was not: Provenance.

Deutsche Fassung: README.de.md. Language rule: the English set carries what is portable across the EU, the German set leads wherever the enforcement layer is German (KI-MIG, Bundesnetzagentur, UWG). Verbatim quotes from EU documents stay in their official English wording in both.

Decision tree, case catalog and a git-native editorial-evidence gate for the AI Act's content transparency duties (Article 50).

⚠️ Not legal advice. This playbook is a technical and editorial working aid. The European Commission's guidelines (C(2026) 5054 final) are legally non-binding; only the Court of Justice of the European Union can interpret the AI Act with binding effect. There is no case law on Article 50 yet — this playbook builds a defensible position, not a safe harbour. As of: 24 August 2026.

Why this exists

Article 50 of Regulation (EU) 2024/1689 has applied since 2 August 2026 (Art. 113; guidelines para 153). The duties that bite on the publishing side — the deepfake label and the disclosure for published text on matters of public interest, both deployer duties under Art. 50(4) (para 19: "the professional deployer's transparency obligations laid down in Article 50(3) and (4) AI Act") — apply with no grace period. The transitional period until 2 December 2026 covers only the provider marking duty under Art. 50(2) for systems already on the market (para 153; Commission FAQ on Article 50).

For text, outside law-enforcement use authorised by law (para 139), the law offers exactly one way out. Art. 50(4), second subparagraph does not apply where "the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content" (para 130). Two cumulative conditions (para 133), and the bar is substantive: fact-checking is "a minimum requirement" (para 134), while "[s]uperficial, solely formal or procedural checks (e.g. spell-checking or grammatical correction), the mere existence of an editorial policy, automated review processes or cursory editorial approval without substantive engagement" do not qualify (para 135).

The exception is also order-sensitive:

"Any substantive AI intervention occurring after the human review or editorial control process has taken place will therefore cause the exception to become void." — guidelines para 136

The form of the evidence, however, is legally open. The Code of Practice on Transparency of AI-Generated Content states that adhering deployers do not have "to document individual instances of human review or editorial control over individual text publications" (Sec. 2, Commitment 4) — while expressly allowing them to "record additional information on the nature of the review or the type of involvement of the AI system in the published text". Anyone who is not a signatory has to demonstrate compliance "through other adequate means" and is expected to run a gap analysis against the code (para 148).

That open space is what this repository fills. The review becomes a review record bound to the reviewed bytes by SHA-256, the sign-off becomes a server-attested PR approval, and the order rule of para 136 becomes a CI check that turns red the moment anything touches the content afterwards.

Two sentences that carry the whole playbook:

• The editorial exception is text only. For images, audio and video there is none — there only the deepfake test decides, and the art exception merely relaxes the form of disclosure (Art. 50(4); paras 119–123).
• Mnemonic: text asks "who checked it?" — image asks "does it look real?"

How this sits next to other AI Act tooling

There is no shortage of AI Act tooling, and this repository replaces none of it. A snapshot of the official community marketplace for Claude plugins on 24 August 2026 lists 2,282 plugins, several of them on the AI Act — among them eu-ai-act-compliance (a risk classifier under Article 6), gia-eu-ai-act-compliance (classification under Regulation (EU) 2024/1689 and screening against the prohibitions of Article 5), norma-claude-skill (an EU compliance methodology with templates for the AI Act, ISO 42001 and NIS2) and sentinal-stack (agents for DLP, AI Act compliance and audit evidence).

The dividing line is the object, not the quality. Those tools classify systems or supply templates. This repository works one level down, on the content a system produces: which Article 50 duty attaches to one specific text, image, audio or video file, and how the human review behind it is anchored as evidence in Git and CI. Whoever needs to know whether a system is high-risk reaches for the classifiers; whoever needs to know whether this article needs a label and how to show the review happened is in the right place here. Marketplace snapshots age — re-check the picture before relying on it.

What's inside

playbook/en/   the nine chapters, English
playbook/de/   the same nine chapters in German — the originals
plugins/       the Claude Code skill — this repository doubles as its marketplace
gate/          the editorial-evidence gate: schema, check script, PR template, CI workflow
tools/         label-crop-check, the geometric pre-check for labels set inside an image
examples/      the worked example that this repository's own CI runs over

For text the decision tree has two exits, and the playbook builds out both: the exception path — human editorial review, evidenced through the gate (chapter 03, gate/) — and the labelling path without review (chapter 09). This repository itself takes the second; the notice at the top of this file is the example.

Every chapter exists in both languages. The table links the English files; the German originals sit in playbook/de/ and lead where the enforcement layer is German.

Path	What it is
playbook/en/01-decision-tree.md	Decision tree, eight stages (0–7): exposure, chatbot, deepfake test, publication, public interest, editorial exception, form of the label, neighbouring duties
playbook/en/02-case-catalog.md	71 classified cases in seven groups — image (advertising/e-commerce), image (journalistic), audio, video, text, internal content, legacy content — each with a verdict and a source anchor
playbook/en/03-editorial-exception.md	The text exception in depth: both cumulative conditions, what does not qualify (para 135), the order rule (para 136), what documentation is and is not required
playbook/en/04-labelling-form.md	Form, wording and placement per modality; the optional EU icons; a German/English label text library; accessibility; pre-publication checklist
playbook/en/05-agencies-and-contracts.md	Commissioned work: who counts as deployer, who labels, who carries the unfair-competition risk, plus a contract checklist
playbook/en/06-myths-faq.md	Eight widespread misconceptions, corrected with sources
playbook/en/07-provider-marking.md	Provider marking under Art. 50(2): what a watermark proves, what it does not, why it never replaces your own label
playbook/en/08-legal-basis.md	Legal basis: timeline, verbatim norms, German authorities, sanctions, what supervisors expect as evidence, and when this needs re-checking
playbook/en/09-labelling-path.md	The labelling path: publishing without editorial review — when it is the right one, what belongs in the label line and what does not, placement without layout damage, the provenance page, the as-of date
gate/README.md	The editorial-evidence gate: review-record schema, the zero-dependency check script, PR template, ready-made GitHub Actions workflow
plugins/eu-ai-act-content-check/skills/eu-ai-act-content-check/SKILL.md	The Claude Code skill: runs the Article 50 check before publishing and prepares — never attests — the review record
tools/label-crop-check/README.md	Geometric pre-check: does an in-image label survive the 1:1, 4:5, 9:16 and 16:9 platform crops?
examples/blog-artikel.md	A worked example: a fictional local-news article and its review record, used by this repo's own CI

The gate in 30 seconds

Every reviewed content file <name>.md gets a sibling record <name>.review.json. The record says who reviewed the substance with what subject-matter competence (para 134), who holds editorial responsibility and where that is publicly findable (para 138), what was actually changed or rejected, how the content was classified — and it carries the SHA-256 of the released version.

Any later change — a hand edit, a build step, an AI rewriting the teaser — changes that hash, invalidates the record and fails the check. Para 136 stops being a promise in a handbook and becomes a state the pipeline verifies.

Break it yourself (the script needs Node 18 or newer; its console output is German):

node gate/scripts/check-review-record.mjs examples
echo "Exit: $?"   # 0 — the record matches the content

printf '\nThis sentence was added after sign-off.\n' >> examples/blog-artikel.md
node gate/scripts/check-review-record.mjs examples
echo "Exit: $?"   # 1 — "Hash-Abweichung": changed after editorial sign-off

git checkout -- examples/blog-artikel.md   # assumes the file is committed — otherwise remove the appended line by hand

Nobody had to notice that change. The hash noticed. In a real pipeline the merge would be blocked and a fresh review due — before publication, not after.

What the gate does and does not do. It enforces the process and makes it auditable; it does not enforce the quality of the review. A green gate over a review that never really happened is a well-documented breach — para 135 rules out "cursory editorial approval without substantive engagement". The gate is a documentation form that goes beyond the legal minimum and is expressly permitted (Code of Practice, Sec. 2, Commitment 4) — not a safe harbour.

Quickstart

The gate, in five steps (details: gate/README.md):

1. Copy gate/scripts/check-review-record.mjs and gate/review-record.schema.json into your repository, then run node gate/scripts/check-review-record.mjs --self-test once — exit 0 means the copy is intact.
2. Decide which directories hold label-relevant content and add one record per content file: copy gate/templates/review-record.example.json, fill it in, set content_sha256 with shasum -a 256 <file>.
3. Copy gate/github-actions/editorial-gate.yml to .github/workflows/ and point it at your directories.
4. Copy gate/templates/PULL_REQUEST_TEMPLATE.md to .github/.
5. Enable branch protection on the publishing branch: require a pull request and at least one approval, switch on "Dismiss stale pull request approvals when new commits are pushed", and mark the gate's status check as required. Without step 5 the gate is a recommendation, not a gate.

The skill. This repository is also a plugin marketplace. Two commands install the skill:

/plugin marketplace add eduard-wolf/eu-ai-act-content-transparency
/plugin install eu-ai-act-content-check@eu-ai-act-content-transparency

Without the plugin mechanism, copy the folder plugins/eu-ai-act-content-check/skills/eu-ai-act-content-check into your project's .claude/skills/ directory, so that .claude/skills/eu-ai-act-content-check/SKILL.md exists; a local clone of this repository also works as a marketplace source (/plugin marketplace add <path to this repository>). The skill format is documented for Claude Code — and the instructions themselves are plain Markdown, readable without installing anything: SKILL.md.

The skill walks the decision tree before content is published, deployed or merged, and prepares the review record — it never fills in the reviewer, the review date, the pruefung booleans or the final hash. A record produced end to end by an agent is precisely the "automated review process" that para 135 excludes.

The crop check. node tools/label-crop-check/check.mjs --image 1600x1200 --label 1150,40,120x60 tells you whether a label placed inside an image survives the usual platform crops — 1:1, 4:5, 9:16 and 16:9 by default, others via --ratios; exit 1 if one of them cuts it. Zero dependencies, --self-test built in.

Honest scope

• A playbook plus a verification pattern — not a framework. No CMS plugin, no service, no database. Two Node scripts with zero dependencies, a JSON schema, a workflow file and nine chapters. Adopt the parts you need.
• Two languages, one rule. The English set carries what is portable across the EU; the German set leads where the operative enforcement layer is German: the Bundesnetzagentur is the central market surveillance and complaints body under § 2 KI-MIG (bundesnetzagentur.de/ki), and unfair-competition law (UWG) runs alongside Article 50. The scripts' console output is German. Quotes from EU documents are kept in their official English wording.
• Non-binding sources, no case law. Guidelines and FAQ are Commission interpretation, not law; the Commission itself announces a review of the guidelines (para 155). Fines under Article 50 can reach EUR 15 000 000 or 3% of total worldwide annual turnover, whichever is higher — for SMEs whichever is lower (para 152).
• A label is no free pass. "Reliance on the attenuated transparency obligation cannot be a justification for failing to respect the fundamental rights of individuals or rightsholders under Union law on intellectual property or Union data protection law" (para 124, whose footnote 34 extends this to published texts informing the public on matters of public interest). German unfair-competition law (§§ 5, 5a UWG) applies alongside, label or no label — see playbook/en/08-legal-basis.md.
• The machine-readable provider marking never replaces your own perceivable label. "[D]eployers cannot rely on the machine-readable marking embedded in the content by the provider under Article 50(2) AI Act, since those markings are not immediately clear and distinguishable for the natural persons exposed to the deep fake content." (para 117)
• Grey zones stay grey zones. Where the sources do not decide a case, the chapters say so and ask for a documented classification instead of manufacturing certainty. [zu verifizieren: …] marks deliberately open points — no official source covers the case there; nothing is invented.

Status

As of: 24 August 2026 — three weeks after Article 50 became applicable, with no case law and no published administrative practice on it. Every legal statement in this repository carries a dated stamp and a source; the events that would change the conclusions — a first CJEU or BGH ruling, first documented warning letters or Bundesnetzagentur proceedings, a review of the guidelines, an amendment to the KI-MIG — are listed in playbook/en/08-legal-basis.md, section 7.

License

MIT — see LICENSE.