🤖 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.
⚠️ 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. With C(2026) 5054 final the Commission approved the content of the draft guidelines; they will be formally adopted once all language versions are available. There is no case law on Article 50 yet — this playbook builds a defensible position, not a safe harbour. As of: 24 August 2026; updated on 1 October 2026.
For text, the decision tree has two exits: the editorial exception and the label (chapter 01, stage 5). The first branch is developed by chapter 03 and the gate; this chapter develops the second — for everyone who publishes with AI agents and cannot or does not want to review the substance themselves. Wording, icons and placement per modality are in chapter 04; this chapter is about the path as a whole. This repository walks it itself — the block above this paragraph is the example.
🔑 Key point of this chapter: The label is a statement of origin, not a quality judgement. It says where the text comes from — not how good it is. That is how it is to be read, how it is to be written, and how it is to be placed.
The path begins with the competence test from chapter 01, stage 5, and chapter 03, section 3. Its question is not "am I an expert?" but "can I review THIS text on the substance?". Para 134 ties human review to persons "possessing relevant knowledge and professional judgement pertaining to the subject matter under scrutiny" and makes fact-checking the minimum. If one of the four test questions from chapter 03, section 3.1, comes out "no" — typically in someone else's field; in practice also when too many texts land on one person to still notice mistakes (test question 3) — or if the exception fails for another reason (AI intervention after sign-off, para 136; automated or cursory review, para 135), it is not available for this text.
Then the wording of the provision applies: Article 50(4), second subparagraph, AI Act orders the disclosure ("shall disclose") and sets the exception beside it ("This obligation shall not apply where …"). Labelling is the standard case; the exception is the exception. Anyone who deliberately forgoes the editorial review — for lack of time, for lack of capacity, or because there is no subject-matter expertise on the topic — is not choosing an inferior branch but the path the law names first. Scope remains the preliminary question: the line is mandatory only for published text on matters of public interest (para 131; chapter 01, stages 3 and 4); outside that, the same line is voluntary (chapter 04, Group B) and does no harm as long as it is true.
⛔ A word of warning: A review record without a substantive review is worse than none. Para 135 expressly rules out "cursory editorial approval without substantive engagement"; a record that logs such an approval as an editorial review documents a false statement — with a date, a name and a review note, supplied by your own side (this playbook's reading, chapter 03, section 3.4). Anyone taking this path therefore creates no review record, but records the decision in one sentence: "labelled because the competence test for this topic comes out negative."
It fulfils the transparency duty. Article 50(4), second subparagraph requires exactly one piece of information: that the text is "artificially generated or manipulated". Para 132 names the form — "clear and perceivable by natural persons (e.g. visible or audible measures) without them needing to rely on any specific technical tools or performing dedicated actions" — and Article 50(5) the timing: at the latest at the first exposure (paras 141, 143). The provision requires nothing more: no wording, no icon, no form to fill in (chapter 04).
It does not fulfil the duty of accuracy. Para 124: reliance on the — there attenuated — transparency obligation "cannot be a justification for failing to respect the fundamental rights of individuals or rightsholders" — and footnote 34 extends that to text: "This equally applies to published texts informing the public on matters of public interest." The prohibition of misleading practices (§§ 5, 5a UWG — the German Act against Unfair Competition), copyright, personality rights and data protection law apply alongside the label, not beneath it (chapter 08, sections 4 and 6).
The AI is a tool, not a third party. The OLG Hamm (the Higher Regional Court of Hamm; judgement of 12 May 2026, I-4 UKl 3/25, ECLI:DE:OLGHAM:2026:0512.4UKL3.25.00, full text: nrwe.justiz.nrw.de; leave to appeal on points of law granted, judgement para 118 — the NRWE case-law database lists the judgement as final (rechtskräftig) as of 24 August 2026) attributed the statements of a website chatbot to the operator as its own commercial practices (geschäftliche Handlungen): the chatbot was „(lediglich) ein technisches Mittel" — "(merely) a technical means" — over which the defendant „hinreichende Steuerungsgewalt besaß" — "had sufficient control" (judgement, para 70), and not a third party (judgement, para 77). The objection that consumers know AI is error-prone is rejected by the judgement: „Vielmehr trifft es zu, dass ein Großteil der angesprochenen Verbraucher in besonderer Weise auf die Richtigkeit der computergenerierten Antwort vertraut, da Maschinen im Allgemeinen als weniger fehleranfällig als der Mensch wahrgenommen werden." — rather, it is true that a large proportion of the consumers addressed place particular trust in the correctness of the computer-generated answer, since machines are generally perceived as less error-prone than humans (judgement, para 104; translated from the German original). Carrying this over to statically published AI texts suggests itself, but has not been adjudicated. Consequence: a label lowers the expectation as to origin, not as to truth.
The label is therefore the reason why no editorial review is needed — not the reason why facts about yourself, your own service and third parties would not need checking. That remaining field of checks is small and needs no subject-matter expertise in the topic, because it concerns your own knowledge:
| Still to be checked — even without editorial review | Anchor |
|---|---|
| Figures and facts about your own company and your own service: reach, number of customers, certifications, results, promises of currency | § 5(2) no. 1 UWG — essential characteristics of the goods or service; on this playbook's reading, promises of currency (Aktualitätszusagen — assurances that content is up to date, such as "always up to date" or "as of 2026") are statements of fact about such a characteristic — no decision specifically on this is documented |
| Statements about your own competence, qualifications, awards | § 5(2) no. 3 UWG — attributes of the trader; the OLG Hamm category of case (invented specialist medical titles) |
| Factual assertions about identifiable third parties — persons as well as companies | § 824 BGB (Bürgerliches Gesetzbuch, the German Civil Code: untrue assertion of fact, negligence suffices) and personality rights (para 124, footnote 34); value judgements are to be distinguished from these |
That is a fraction of the text — not "review everything", but check what nobody else can check. The preliminary question remains whether the publication is a commercial practice (geschäftliche Handlung, § 2(1) no. 2 UWG); for a provider's informational content that showcases its expertise, the BGH (the German Federal Court of Justice) regularly answers yes (judgement of 19 May 2011, I ZR 147/09 — "Coaching-Newsletter").
A label line consists of at most three kinds of statement — only two of them belong in it.
| Type | Examples | Status | Anchor |
|---|---|---|---|
| Mandatory core: the origin statement | "Created with AI" · "This text was generated using artificial intelligence" | Mandatory. Fulfils Article 50(4), second subparagraph; no wording prescribed, EU icon optional | para 132; chapter 04, sections 2 and 3 |
| True process facts | "not editorially reviewed" · "machine-verified against the sources" · "as of DD Mon YYYY" | Voluntary, permissible, in your favour — provided they are true | a declaration of knowledge (Wissenserklärung), not a clause (section 3.1); if untrue, § 5(2) no. 1 UWG |
| Value judgements about yourself | "professionally created" · "by AI experts" · "with high-performance AI" · "high quality" | Not in the line | § 5(2) no. 1 and no. 3 UWG (section 3.2) |
Process facts are declarations of knowledge (Wissenserklärungen — statements that convey facts, as opposed to declarations of intent that produce legal effects), not liability clauses. They shape what a reader may legitimately expect, instead of excluding a legal consequence. On the prevailing reading, without a contract — a freely accessible page, no registration, no login — §§ 305 et seq. BGB (the rules on standard terms) are not applicable, because no contractual terms are incorporated; such a notice operates purely factually, by shaping the objective recipient's perspective (objektiver Empfängerhorizont — the standard of how a reasonable recipient would understand the statement) (so, for freely accessible stock-market information on a bank's website, LG Itzehoe (the Regional Court of Itzehoe), judgement of 25 March 2014, 7 O 163/13 — a first-instance decision; the question is not settled by the highest courts). That is why the line — unlike an exclusion of liability in general terms and conditions (AGB) — cannot be "invalid". The classic "No guarantee of currency, correctness or completeness", by contrast, does not belong in it: as a contractual clause it is regularly invalid (§§ 307, 309 no. 7 BGB) and can itself be the subject of a warning letter (Abmahnung) vis-à-vis consumers; without a contract it achieves nothing that "not editorially reviewed" does not achieve more precisely.
Condition: they must be true.
- "Machine-verified against the sources" may appear only if that verification exists: named sources, an actual comparison, a result that can be shown — which quotes, which source references, which links, with what finding. "The AI had another look at it" is not a verification against sources; a run may only be called what it is.
- The date must come from the actual verification or generation run — not from the build, not from the deploy, not from the commit of a layout change (section 6).
- On this playbook's reading, an untrue process statement is a statement about an essential characteristic of the service and can be challenged as misleading under § 5(2) no. 1 UWG; whether that reference point holds for free content has not been decided, and a judgement specifically on process statements attached to online texts is not documented. A wrong as-of date is worse than none.
"Professionally created", "by AI experts", "with high-performance AI", "high quality" are not origin statements but advertising claims about the trader's competence and the quality of the service — the reference points of § 5(2) no. 3 and no. 1 UWG respectively. They are open to challenge because they have to be provable, and they are especially delicate next to a line that says nobody reviewed the content on the substance: the claim of expertise then stands directly beside the admission that no expertise was applied to this text. Precisely this constellation — claimed expertise that does not exist — is what the OLG Hamm prohibited under § 5(2) no. 3 UWG (chapter 08, section 4). Rule: showing beats claiming. The line stays sober; the evidence goes on the provenance page (section 5).
"Article 50(4) AI Act" does not belong in the line. Para 142 requires information that is "noticeable, easy to understand by and accessible for the natural person concerned"; for textual notices the guidelines recommend "plain-language labels" and "simplified wording, particularly for users with lower digital literacy and children" (para 37, written for Article 50(1)); for the audible disclaimer the Code of Practice requires "plain and simple natural language" (Section 2, Measure 1.1). A legal citation lengthens the line, reads like officialese and adds nothing legally — the duty is fulfilled by the origin statement, not by citing it. The source reference belongs on the provenance page. (The block at the head of this chapter names the provision because it addresses a specialist audience that works with it — an exception, not a model for a website.)
In each case the arrow stands for the link to the provenance page. An EU icon in front of it is optional; icon plus text beats icon alone (chapter 04, section 2). "Machine-verified against the sources" is added only if that verification exists (section 3.1).
| Variant | German | English |
|---|---|---|
| Minimal — only the mandatory core | Mit KI erstellt. | Created with AI. |
| With process facts | Mit KI erstellt · nicht redaktionell gegengelesen · Stand: 24.08.2026 | Created with AI · not editorially reviewed · as of 24 Aug 2026 |
| With provenance link | Mit KI erstellt · nicht redaktionell gegengelesen · Stand: 24.08.2026 · Wie dieser Inhalt entsteht → | Created with AI · not editorially reviewed · as of 24 Aug 2026 · How this content is made → |
Per content item, not per layout element. The line sits at the head of the post: below the title, before the first paragraph. Code of Practice Sub-measure 1.2.2 lit. f: "For published text, Signatories will place the icon or equivalent label, for example above or at the top of the text, near the headline of the text, or in the colophon at the beginning of the text, as long as placement is clear, consistent, and distinguishable for the end-user." Article 50(5) requires the information "at the latest at the time of the first interaction or exposure"; the guidelines' example for text reads "at the start of an AI-generated or manipulated text publication on matters of public interest" (box after para 143).
Not sufficient, under para 142, is whatever is "easily overlooked or missed by natural persons under normal exposure or interaction conditions" — supplemented by the negative list for Article 50(1) (para 38), whose standard carries over to paragraph 4 via paragraph 5:
| Evasive placement | Why it does not hold |
|---|---|
| A sticky banner across the whole website | not specific to the individual content item — para 38: "Generalised disclosures that are not sufficiently specific to the AI system's outputs and interactions", negative example "Services on this website use AI"; para 39 (likewise on paragraph 1) warns of "habituation effects (so-called “banner blindness”)"; not "near the headline of the text" (Code of Practice, lit. f) |
| A blanket notice in the Impressum (the legally required site notice), in the privacy policy, on an info page | "part of terms of use that are often not read by users" (para 142); "Disclosures contained only in terms and conditions, URLs, or documentation" (para 38) — such pages may supplement, not replace |
| The footer of the post or of the page | not at first exposure (Article 50(5)); Code of Practice lit. f names the top and the beginning, not the end |
| Metadata only, structured data, an HTML comment | not perceivable without technical tools (para 132); on the provider marking para 117, chapter 04, section 1 |
Technically: one component, one place. On websites with a shared layout that is exactly
one template component — a partial, a component, a shortcode — rendered in the same place for
every affected content item; no element per section, no manual insertion per post. One field per
content item (for instance ai: true and as_of: 2026-08-24 in the front matter) controls
display and date. Labelling only the affected parts is permissible (Code of Practice lit. f:
"Signatories may label only that part of the text which is AI-generated or manipulated"), but
not mandatory — the line per content item is the simpler and more robust way. As text it is
readable for screen readers; an icon needs alternative text wherever accessibility requirements
apply (Article 50(5), second sentence; para 144; chapter 04, section 7).
SEO. One line of boilerplate is not a duplicate-content problem: navigation, footer and cookie
notices are identical on every page anyway, and one more short line changes nothing measurable in
the ratio of main content to surrounding frame. Marked up as an element of its own (for instance
an <aside> with its own class), it also stays cleanly separated from the body text.
Timing. Subject to the labelling duty are texts published on or after 2 August 2026 — para 154: "if texts that have been generated or manipulated before 2 August 2026 are published on or after that date, they need to be labelled"; texts published earlier need no retroactive labelling. Voluntary retroactive labelling does no harm — para 154 even expressly encourages it (worded there for pre-existing deepfakes, "without being expected to engage in disproportionate efforts"). Anyone labelling retroactively dates honestly: the as-of date is the state of the content, not the day of the retroactive labelling (section 6).
The line says what; the provenance page says how. It is the counterpart to the label line: one page that explains how content comes about, which checks run — and which do not — and how errors are reported. The model is this repository's Provenance, to which every chapter's line links. Three reasons to create one:
- This is where quality is shown rather than claimed. The process facts the line names are demonstrated here: which sources, which comparison, which result. Facts about the process may be set out at length — value judgements ("professional", "high quality") still may not (section 3.2).
- It supplements without replacing. Para 38: disclosures in documentation "may complement, though not replace, in-context disclosure" — the line stays on the content, the page hangs off it.
- It is the answer to enquiries. Para 148 expects non-signatories "to explain how the measures they implement ensure compliance", and expressly announces requests for information to deployers as well: "Deployers may also be subject to such requests with regard to their labelling practices under Article 50(4) AI Act." The provenance page is that explanation, written in advance.
What it is not: not a review record and not a claim of editorial review. The sentence "no editorial review by a human with relevant subject-matter expertise" stands on the page just as it does in the line — "editorially reviewed" belongs only where it is true (section 1).
Minimum content — checklist:
- Scope: which content the label applies to — areas, start date, what is excluded.
- Origin: which systems or agents produce the content, in which passes; what role the human plays (commissioning, selection, approval) — and that this approval is not a substantive review.
- Verification steps: what is machine-checked and how — named sources, comparison method, links, tests — with the result.
- Non-verification: what expressly does not take place (no substantive editorial review, no legal review) and why the exception is not claimed — the competence test in one sentence.
- Currency and ageing: where the as-of date comes from, which reference state was checked and which events cause the content to go out of date.
- Responsible entity: who publishes the content as deployer, with contact details — a reference to the Impressum suffices (§ 5 DDG, the German Digital Services Act; chapter 03, section 6).
- Correction loop: how errors are reported, what happens then (check, correction, new as-of date) and how corrections are made recognisable.
Why it is the most effective single measure. A dated statement is a statement about a point in time: "As of 24 August 2026, the position is …" does not become untrue because the situation changes later — the objective recipient's perspective qualifies it in time from the outset. On this playbook's reading, it is the most effective single measure against the charge of misleading practice (§ 5 UWG) and of untrue assertion of fact (§ 824 BGB) at once — more effective than any exclusion of liability, because the date works on the expectation rather than on the legal consequence; a decision specifically on as-of dates attached to online texts is not documented. On the same reading, there is no statutory duty to keep freely accessible content up to date — updating duties such as § 327f BGB are tied to a contract (§ 327(1) and (3) BGB). Such a duty arises only through a promise of currency — or, on the same reading well arguable but not settled by the highest courts, through positive knowledge that published content has since become wrong and dangerous; it must then be corrected or removed (section 5, correction loop). A promise of currency is exactly what a wrong date gives.
| Where the date comes from | Fit for purpose? |
|---|---|
| The last substantive generation or verification run — the run that produced statements or compared them against sources | ✅ |
| The reference state that was checked: "Legal position: AI Act as amended by Regulation (EU) 2026/1744; guidelines C(2026) 5054 final of 20 July 2026" | ✅ better than a bare date |
| Build or deploy timestamp | ❌ dates the rendering, not the content |
| Commit date of a layout, spelling or link change; "last modified" in the CMS | ❌ dates the file, not the statement |
| Manually entering "today" on publishing with no verification run behind it | ❌ a promise with nothing to back it |
A wrong as-of date is worse than none. If "As of: 08/2026" sits above a text whose substance has not been touched since 2024, the date itself is an untrue statement — on this playbook's reading, about an essential characteristic of the service and open to challenge as such (§ 5(2) no. 1 UWG; whether that reference point holds for free content has not been decided). If there is no verification run, there is no date — or it honestly reads "First published: …".
For existing content what counts is the last substantive change in the version history, not the last commit: changes only to layout, spelling, links or metadata do not count. In practice: go through the history per content file, look for diffs beyond whitespace and formatting, take the last state with a substantive change — or, where that cannot be reconstructed, run a new verification run and date that. Keep them technically separate: the date is a field in the content that only the verification or generation run writes; the build reads and renders it, but never sets it itself.
- Neighbouring duties remain. Para 124 with footnote 34: copyright, data protection and personality rights as well as the prohibition of misleading practices — chapter 08, sections 4 and 6. The label gives relief on origin, not on truth.
- Legal texts are the bigger risk — even though they are not subject to the labelling duty. On the convincing reading of the guidelines, a privacy policy and general terms and conditions do not fall under Article 50(4), second subparagraph: self-disclosure towards one's own users and contractual terms are not "matters of public interest" within the meaning of para 131 iii ("relevant to society at large", "meriting public debate or scrutiny"); the question has not been expressly decided anywhere. In substance they are nevertheless the bigger warning-letter risk: since the CJEU, judgement of 4 October 2024, C-21/23 ("Lindenapotheke"), and the BGH, judgement of 27 March 2025, I ZR 186/17 ("App-Zentrum III"), inadequate privacy notices can be challenged under unfair-competition law; invalid clauses in general terms and conditions fall away without replacement under § 306 BGB — with no reduction to the permissible extent (geltungserhaltende Reduktion), that is, precisely when they are needed. If legal review is to be invested anywhere, it is here.
- Statements about third parties and about your own services remain to be checked (section 2) — the label changes nothing about that, and the OLG Hamm's attribution applies regardless.
- AI literacy under Article 4 AI Act remains a standing organisational duty, independent of any label — since 27 July 2026 a duty to take measures that support AI literacy, without a prescribed level (chapter 08, section 6).
- Images, audio and video know no other branch anyway. The playbook's first rule of thumb: the editorial exception applies to text only (Article 50(4), second subparagraph; para 133). For deepfakes, labelling is the only path in any case — there the deepfake test alone decides (para 113 f.), and the art exception merely relaxes the form (paras 119–123). Featured images and videos therefore go separately through stage 2 of the tree (chapter 01).
This path, too, is a defensible position, not a safe harbour: it fulfils the transparency duty in a traceable way and keeps readers' expectations honest — the responsibility for what is said stays with the deployer.
- Answer the competence test honestly for the specific text and record the result (chapter 01, stage 5; chapter 03, section 3). One "no" is enough: this path.
- Create no review record that claims an editorial review (para 135). Instead, one sentence: "labelled because …".
- Build the line according to the grammar: origin statement + only true process facts + as-of date. No value judgements, no legal citations (section 3).
- One component, one place: the line at the head of every affected content item, below the title, before the first paragraph; visual check without click, hover or scrolling. Not in the Impressum, not in the banner, not in the footer (section 4; para 142; Code of Practice Sub-measure 1.2.2 lit. f).
- Create the provenance page, link to it from the line, work through the minimum content (section 5; model: Provenance).
- Set the as-of date from the actual generation or verification run, name the reference state, keep the data source separate from the build; date existing content by the last substantive change (section 6).
- Work through what the line does not cover: your own figures, your own service promises, statements about third parties, legal texts — and images, audio and video separately through the deepfake test (sections 2 and 7).