Digital Marketing

December 2 Is Your AI Tool Vendor's Deadline. The One That Binds You Passed on August 2.

Leo Wang November 12, 2026
December 2 Is Your AI Tool Vendor's Deadline. The One That Binds You Passed on August 2.

December 2 Is Your AI Tool Vendor's Deadline. The One That Binds You Passed on August 2.

If you publish AI-assisted content into the EU, you have probably been told that 2 December 2026 is the day your AI-generated content must carry a label. That date is real and it is in the law. It is also not addressed to you.

The deadline comes from Article 111(4), inserted into the AI Act by the Digital Omnibus on AI in July 2026. Read who it names: providers of AI systems that generate synthetic audio, image, video or text content, where those systems were placed on the market before 2 August 2026, must take the necessary steps to comply with Article 50(2) by 2 December 2026 [2]. Article 50(2) is the machine-readable marking obligation, and it sits on providers. The Omnibus recital describes the whole thing as a four-month transition for those providers [2].

A brand that uses ChatGPT or Claude to write its German product pages is not a provider. It is a deployer. Its obligations live in Articles 50(3), 50(4) and 50(5), and those have applied since 2 August 2026 with no transition period at all [1][2].

So the question worth your time is not what to do before December. It is which paragraphs reached you in August, and whether your content already sits inside or outside them.

---

Quick answer

  1. You are in scope even with no EU entity and no European domain. Article 2(1)(c) names providers and deployers established or located in a third country where the output produced by the AI system is used in the Union [1]. Server location, company registration and domain suffix are not the test.
  2. Your role is almost certainly deployer, which means Article 50(2) is not yours. Machine-readable marking belongs to whoever puts the generative system on the market under its own name [1].
  3. The Commission has already spoken on advertising copy, and the answer is narrower than either rumor. Its Article 50 guidelines list AI-manipulated text inside a company's advertisement or product descriptions as falling outside Article 50(4) — with a bracket excluding claims related to matters such as health, consumer safety or sustainability [3].
  4. The human-review exemption is real and it is conditional. Article 50(4) does not apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication [1]. The guidelines then exclude spell-checking, holding an editorial policy, automated review and perfunctory sign-off from counting as that review [3].
  5. Most of the tools your content team actually uses have not signed the Code of Practice. The signatory list carries the foundation model providers; it does not carry the marketing and writing SaaS layer [5].

The boundary on all of it: the AI Act, the Digital Omnibus, the Article 50 guidelines and the Code of Practice say nothing whatsoever about how labeled content is treated in search rankings or in AI-generated answers. Anyone selling you either version of that claim is not citing these documents.

---

Article 50 reaches you without an EU entity

Article 2 lists who the regulation applies to, and the list has seven limbs. Limb (a) covers providers placing systems on the Union market, explicitly regardless of whether they are established in the Union or in a third country. Limb (b) covers deployers having their place of establishment or located within the Union. If the list stopped there, a Chinese exporter with no European subsidiary would be outside it.

It does not stop there. Limb (c) covers providers and deployers of AI systems established or located in a third country, where the output produced by the AI system is used in the Union [1]. Deployers are named in that limb alongside providers. The trigger is where the output is used, which makes your incorporation documents, your hosting region and your top-level domain all irrelevant to the question.

The Commission's guidelines narrow limb (c) in one way that matters. For a third-country deployer, the guidelines treat the obligation as engaged where the deployer itself foresaw the dissemination and use of the output in the Union — their illustrations include directing or authorizing distribution in the Union, and posting a deepfake to the globally accessible internet. Where content reaches a Union audience through channels that were unforeseeable and outside the deployer's control, the guidelines do not treat it as in scope [3].

Run that against what a cross-border team actually does. Translating your catalog into German, buying ads against German queries and shipping to German addresses is the foreseeable case, not the accidental one. The narrowing protects a company whose content drifted into Europe by accident. It does not protect a company that chose Europe as a market.

One more door that people expect to be open is closed. Article 2(12) exempts AI systems released under free and open-source licences from the regulation in principle, except where they fall under Article 5 or Article 50 [1]. Running an open-weights model on your own hardware does not move you outside this article.

Provider or deployer: the split that decides which paragraphs bind you

Article 50 is titled as transparency obligations for providers and deployers, and each paragraph picks one. Getting your own side wrong is how teams end up preparing for the wrong deadline.

ParagraphWhat it requiresWho owes it
50(1)Design the system so a person is told they are interacting with an AI system, unless that is obvious to a reasonably observant person in the circumstancesProvider
50(2)Mark outputs in a machine-readable format and make them detectable as artificially generated or manipulatedProvider
50(3)Inform people exposed to emotion recognition or biometric categorization that the system is operatingDeployer
50(4) first subparagraphDisclose that deepfake image, audio or video content is artificially generated or manipulatedDeployer
50(4) second subparagraphDisclose that text published to inform the public on matters of public interest is artificially generated or manipulatedDeployer
50(5)Give that information clearly and distinguishably, no later than first interaction or exposure, meeting applicable accessibility requirementsBoth

Source for the whole table: Article 50 of Regulation (EU) 2024/1689 [1].

Two definitions decide your side. A provider develops an AI system, or has one developed, and places it on the market or puts it into service under its own name or trademark. A deployer uses an AI system under its own authority [1]. Using a third party's generative tool to draft copy is the deployer case.

There is a route across the line, and it is widely misdescribed. People point at Article 25, which converts distributors, importers, deployers and other third parties into providers. Read its triggers: all three are gated on the system being high-risk, and the consequence is the Article 16 provider obligations. Article 50 is not in Article 25's list of consequences, and the Digital Omnibus replaced only Article 25(2) and the first subparagraph of Article 25(4), leaving the trigger list untouched [1][2].

The actual route runs through the definitions. Article 3(11) defines putting into service as supplying a system for first use directly to a deployer or for own use in the Union [1]. Combine that with the provider definition and the consequence is concrete: if you build or commission a content generation system and put it into service under your own name, you become a provider even when the only users are your own team. The guidelines add the adjacent case, where a company takes someone else's generative system already on the market, modifies it — their example is new training data — and puts it into service under its own name, becoming the provider of that new system. They also confirm one entity can hold both roles at once, owing machine-readable marking and human-facing disclosure together [3].

Where the line falls for prompt orchestration and retrieval layers, the guidelines do not say. If your pipeline is an internal product with a name and a roadmap, that silence is worth legal advice rather than an assumption.

What December 2 actually moves

Article 111(4) does one thing: it gives providers whose generative systems were already on the market before 2 August 2026 until 2 December 2026 to comply with Article 50(2) [2]. Systems placed on the market on or after 2 August 2026 had to comply from day one.

For a deployer, the date changes nothing about your own duty and something about your supply chain. What arrives on 2 December is the point by which the marking and detection capability of the pre-August tools you already use is supposed to exist. That capability is the thing you can start asking your vendors about.

The guidelines close one escape here as well. A system with both an interactive feature and a generative feature gets the transition only for the Article 50(2) marking obligation; the Article 50(1) interaction disclosure had to be in place from 2 August 2026 [3].

There is also a question about your archive. The guidelines state that content generated or manipulated before 2 August 2026 does not need retrospective marking — but text generated before that date and published on or after it does fall in. The Commission encourages voluntary marking of existing unlabeled content while stating it does not require disproportionate effort such as auditing an entire back catalog or reworking already printed packaging [3]. If your team drafted a quarter of EU landing pages in July and shipped them in September, that is the scenario the guidelines describe.

Where marketing copy sits in Article 50(4)

The deployer text obligation covers AI-generated or manipulated text published with the purpose of informing the public on matters of public interest [1]. The guidelines break that into three cumulative conditions: the text must be published, it must be intended to inform the public, and it must concern a matter of public interest.

Published, in their reading, means accessible to an indeterminate and reasonably large number of unconnected potential readers, paywalled or not. It excludes closed private groups, professional private correspondence and texts internal to an organization such as intranet posts. Matters of public interest they describe as relevant to society as a whole and worthy of public debate or scrutiny, listing political and democratic processes, public administration, justice, fundamental rights, public safety, public health, environmental protection and consumer safety — and then any economic, financial, political, scientific or cultural development capable of becoming a subject of public debate. They note the scope evolves over time and context [3].

That last clause is why "it's only marketing" is not a safe answer. The guidelines then give worked examples on both sides, and the pair below is the useful one for brand teams.

Guideline exampleWhich side of Article 50(4)
AI-manipulated text that is part of a company's advertisement or product descriptions, excluding claims related to matters such as health, consumer safety or sustainabilityOutside
An AI-manipulated passage in a lifestyle website article comparing the effect of different diets on a disease in middle-aged womenInside
AI-generated fantasy fictionOutside
An AI-generated summary, on a newspaper site, of a human-written article about a town council decisionInside
An AI-manipulated company report containing investor information, published on a listed company's websiteInside
A chatbot news summary visible only to the single user who asked for itOutside

All six examples come from the Commission's Article 50 guidelines [3].

Read the first two rows together and the operating rule appears. The test runs on what the text does, not on which content type your CMS calls it. A product description stays outside until it makes a health, consumer-safety or sustainability claim. A lifestyle article that is really about a disease is inside, whatever the brand calendar labels it. The guidelines take no position on "brand blog" as a category, which means the category is not the unit of analysis.

For a cross-border catalog this reframes the audit. The exposure concentrates in supplement and personal-care claims, child and electrical safety language, recyclability and carbon statements, and the health-adjacent editorial that consumer brands publish to earn traffic. The plain spec sheet is the easy part.

The human-review exemption, and four ways to void it

Where Article 50(4) does reach your text, one exemption discharges it: the content underwent a process of human review or editorial control, and a natural or legal person holds editorial responsibility for its publication [1]. Both conditions, together.

The guidelines then say what does not qualify. Superficial, purely formal or procedural checks such as spell-checking and grammar correction do not qualify. The mere existence of an editorial policy does not qualify. Automated review processes do not qualify. Perfunctory sign-off where the reviewer or editorial entity was not substantially involved does not qualify [3].

Two further points carry real operational weight.

AI touching the file after sign-off voids the exemption. Where content is modified, supplemented or restated by AI after human review or editorial control is complete, the result must be treated as AI-generated or manipulated for Article 50(4) purposes [3]. A pipeline that runs a localization or SEO-rewrite pass after editorial approval loses the exemption at that step. The fix is sequencing, and it is free if you do it before the pipeline is built.

Editorial responsibility has to be published. The guidelines ask that the identity and contact details of the legal person, natural person or function holding editorial responsibility be made available in an easily accessible place — online through website terms or other user-facing legal information, offline in a colophon [3]. The Code of Practice spells out the internal side for signatories: a pre-publication human review or editorial control policy naming the person who holds editorial responsibility with their title and contact details, plus an overview of the organizational measures and human resources allocated. It states explicitly that per-item records of individual reviews are not required [4].

That combination tells you how much headcount this actually costs. You need a named owner, a written policy and a review step before publication. Per-item sign-off records are explicitly outside the requirement.

What a label has to look like if you need one

Article 50(5) sets the form: the information goes to the people concerned clearly and distinguishably, no later than the time of first interaction or exposure, and conforms to applicable accessibility requirements [1]. The guidelines confirm Article 50 itself imposes no new accessibility standard, pointing instead to the existing web accessibility and accessibility act directives [3].

They are more useful on what fails. Information buried in an instruction manual, hidden behind several layers of an online interface's menus, or placed in terms of use that typically go unread does not meet the clear and distinguishable test [3].

The Code of Practice gives signatories a concrete form. The icon or equivalent marking uses the capitalized abbreviation AI as its main visual element. For published text, the placement is above or at the top of the text, close to the heading, or in a colophon at the start — and signatories may mark only the portion of the text that is AI-generated or manipulated. For very short text where marking would harm readability, the obligation still stands but can be met with an in-interface contextual cue [4].

Note what the officials have not endorsed. No document read for this article approves a single line in the page footer, and none rejects it either. Treat the published placement guidance as the available instruction and the footer line as untested.

Your content tools are mostly not signatories

Signing the Code of Practice lets an organization rely on its measures to demonstrate compliance, regardless of where it is established or which market surveillance authority supervises it [4]. The list of who has signed is therefore a map of where responsibility has been publicly accepted. As read on 11 November 2026, the list shows 95 Section 1 signatories and 192 Section 2 signatories, against an official count of around 190 organizations by the end of July 2026, roughly half of them small, new companies [5].

LayerOn the list
Foundation model and generative system providersOpenAI, Google, Microsoft, Anthropic, Mistral, Meta, Cohere, Aleph Alpha, Black Forest Labs, Synthesia
Marketing, agency and retail deployersDentsu, Bulgari, E. Breuninger, Iberdrola, Lufthansa, Lenovo, Getty Images, Fastweb
Content and marketing SaaS commonly used by export teamsNot present: Jasper, Copy.ai, Writesonic, Surfer, Semrush, HubSpot, Canva, Adobe
Chinese AI vendorsNot present: Alibaba, ByteDance, Tencent, Zhipu, Baidu, Huawei

Names checked against the published signatory list on 11 November 2026 [5]. Absence from the list is not non-compliance — the Code is voluntary, and the Commission's own position is that an organization choosing another route must demonstrate compliance by alternative and equally adequate means, assessed separately by each market surveillance authority [4][6].

Two conclusions follow for a deployer. First, the middle layer of your stack has made no public commitment about marking, so whether your Jasper or HubSpot output carries a machine-readable mark is a question to put to your account manager in writing. Second, and more important, a provider's mark does not discharge your duty anyway: the Commission's FAQ states that a deployer cannot rely on the provider's machine-readable marking to satisfy its own disclosure obligation [6]. Machine-readable marking and human-facing disclosure are two different jobs.

What the law says about AI visibility

The most common claim attached to this topic is that labeling content as AI-generated makes AI engines less likely to cite it. The second most common is the reassuring inverse.

Neither is in these documents. The AI Act, the Digital Omnibus, the Article 50 guidelines and the Code of Practice were searched in full for terms covering search engines, search results, ranking, recommender systems, citation, crawling, visibility, indexing, snippets, AI answers, retrieval, downranking, demotion and exclusion. Every hit resolved into something else: ranking appears once, in the guidelines, restating the Digital Services Act's transparency requirement for online platform recommender systems. recommender appears only where the guidelines place recommendation and retrieval systems outside Articles 50(1) and 50(2). The search engine hits all concern very large online search engines under the DSA, where Article 35(1)(k) obliges platforms and search engines to make realistic-looking content distinguishable on their own interfaces — a duty on the platform, not a consequence for marked content [1][2][3][4].

The honest statement of this result is narrow, and it is worth keeping narrow. The law does not address how labeled content is treated in search or in AI answers. That silence supports no prediction in either direction. Establishing what actually happens would require a published ranking or citation policy from the engines, or reproducible measurement — and neither was available for this article. When a vendor quotes Article 50 while promising an effect on your citation rate, the gap between those two things is the whole problem.

What non-compliance costs

Article 99(4) sets the band, and point (g) of that paragraph names the transparency obligations for providers and deployers under Article 50 explicitly. The exposure is administrative fines up to EUR 15 000 000 or, for an undertaking, up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher [1].

Two adjustments apply. Article 99(6) makes the fine the lower of the percentage and the absolute amount for SMEs including start-ups, and the Digital Omnibus added Article 99(6a) extending the same treatment to small mid-cap enterprises [1][2]. Member States set their own penalty rules within the regulation's conditions and ceilings, and the Omnibus added a requirement that they take the interests of SMEs and small mid-caps and their economic viability into account [2].

Enforcement sits mainly with national market surveillance authorities. The AI Office's role is limited to AI systems built on a general-purpose AI model where the model and the system come from the same provider, or systems integrated into platforms and search engines designated under the DSA; the European Data Protection Supervisor covers Union institutions, with a EUR 750 000 ceiling for breaches outside Article 5 [1][3][6].

Which authority moves first, and how hard, is not something this article can tell you. No national enforcement legislation, authority designation or published case was verified here.

China's marking rules are a parallel system

Teams working from China already carry a domestic obligation, and it is not the same shape as the European one. The Measures for Labeling AI-Generated and Synthetic Content were issued jointly by the Cyberspace Administration of China, the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration, published on 14 March 2025 and in force from 1 September 2025 across 14 articles [7].

Three provisions show the difference. The duty holder is the service provider — the tool or platform side — rather than being split between a provider and a deployer. Explicit labels for text may sit at the start, the end or an appropriate position in the middle, with implicit labels added to file metadata carrying content attribute information, the service provider's name or code, and a content number. And where a user requests content without an explicit label, the service provider may supply it after setting out the user's own labeling duty in the service agreement, retaining logs for no less than six months [7].

The accompanying mandatory national standard, GB 45438-2025, was published on 28 February 2025 and took effect on 1 September 2025. Its registration record was verified; its technical clauses were not read for this article, so nothing here describes what formats it requires.

What cannot be concluded is that satisfying one regime satisfies the other. No official equivalence finding, mutual recognition arrangement or mapping between the Chinese measures and Article 50 was located. Treat them as two systems with overlapping vocabulary and different allocations of duty.

FAQ

Do I have to label our EU product pages as AI-generated?

Probably not on the strength of Article 50(4) alone, and the reason matters. The Commission's guidelines list AI-manipulated text in a company's advertisement or product descriptions as outside that paragraph, but the same example excludes claims related to matters such as health, consumer safety or sustainability [3]. So the audit unit is the claim, not the page. A spec table is outside; the paragraph on that same page about skin sensitivity or recyclability needs a separate look. Note also that Article 50(4) is one of several laws touching product text — consumer protection and product compliance rules apply on their own terms.

If our AI tool marks its output, are we covered?

No. The Commission's FAQ states that a deployer cannot rely on the provider's machine-readable marking to meet its own disclosure obligation [6]. Article 50(2) marking is designed to be read by software; Article 50(4) disclosure is designed to be noticed by a person. Where both apply, both have to happen.

Our editor reviews everything before publication. Does that settle it?

It settles it if the review is substantial and someone holds editorial responsibility, and if nothing touches the file afterwards. The guidelines exclude spell-checking, the mere existence of an editorial policy, automated review and perfunctory sign-off from counting as human review, and they treat content modified, supplemented or restated by AI after review as AI-generated again [3]. A post-approval AI localization or rewrite step is the common way teams lose an exemption they had.

Will labeling hurt how often AI engines cite us?

The documents that create this obligation do not address it. A full-text search of the AI Act, the Digital Omnibus, the Article 50 guidelines and the Code of Practice found no provision on the treatment of labeled content in search rankings or AI-generated answers. That means the law is silent, which is different from the question being answered in either direction. Anyone asserting a direction needs an engine's published policy or reproducible measurement, and should be asked for it.

Should we sign the Code of Practice?

That is a legal and commercial decision rather than a technical one, and the official framing is narrow on both sides. Signing lets you rely on the Code's measures to demonstrate compliance irrespective of where you are established, but the Code itself states that adherence does not constitute conclusive evidence of compliance with the AI Act. Declining means demonstrating compliance by alternative and equally adequate means, assessed separately by each market surveillance authority, with the Commission noting that lower transparency about the method may attract more requests for information [4][6]. Section 2 is the section that applies to deployers.

Bottom line

December 2 is a supply-chain date. It tells you when the marking capability of AI tools that were already on the market before August is supposed to exist, which makes it a question to put to your vendors rather than a task on your own compliance calendar [2].

Your own date was 2 August 2026, and the work it implies is smaller and more specific than the panic suggests. Find the claims in your EU content that touch health, consumer safety or sustainability, and the health-adjacent editorial that sits under a lifestyle heading. Decide whether each one carries disclosure or qualifies for the human-review exemption. Name the person who holds editorial responsibility, publish their contact details, write the short pre-publication policy, and make sure no AI step runs after sign-off [1][3][4].

Then stop reading claims about what labeling does to your AI citation rate. These four documents do not contain that answer, and the people quoting them do not have it either.

References

[1] - Regulation (EU) 2024/1689 (Artificial Intelligence Act) — Articles 2, 3, 25, 50, 99, 100, 111, 113 and recitals 132–136
[2] - Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI) — amendments to Articles 25, 50, 99, 111 and 113, OJ L, 2026/1744, 24.7.2026
[3] - Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689, C(2026) 7702 final
[4] - Code of Practice on Transparency of AI-generated Content — Section 1 for providers, Section 2 for deployers, published 10 June 2026
[5] - Strong backing for the Code of Practice on Transparency of AI-generated Content — full signatory list, 95 Section 1 and 192 Section 2 signatories
[6] - Transparency obligations under Article 50 of the AI Act — European Commission FAQ
[7] - 人工智能生成合成内容标识办法 (Measures for Labeling AI-Generated and Synthetic Content), Cyberspace Administration of China and three co-issuing authorities, in force 1 September 2025

Related Articles