What do the final Article 50 guidelines change?

The final Article 50 guidelines in one paragraph

If you built your Article 50 compliance on the consultation draft, at least one of your calls is now backwards. The Commission's final Article 50 guidelines under the AI Act, adopted on 20 July 2026 as C(2026) 5054 final, reverse the draft on AI translation, tighten it on AI agents, deep fakes and editorial review, and split the retroactivity rule in two. They are issued under Article 96(1)(d) and, as paragraph 5 says, they are non-binding: authoritative interpretation belongs to the Court of Justice. Their stated aim is to assist competent authorities as well as providers and deployers in applying Article 50 in a consistent and uniform manner, and paragraph 4 says they were informed by stakeholder input on the draft published for consultation on 8 May 2026.

The final text is on the Commission's AI Act Service Desk, and the 8 May 2026 draft remains published for comparison. Paragraph numbers below are the final text's unless marked as the draft's, because the numbering shifted between versions: the draft's retroactivity paragraph 142, for example, is paragraph 154 in the final text. For the duty map itself, who owes what under each limb, start with the Article 50 transparency obligations guide. This page is about the guidelines document.

What changed from the 8 May 2026 draft

Each row below compares the wording of both documents directly. Where the final text simply carries a draft passage over, it is not listed.

Two rows deserve emphasis. The translation reversal changes the answer for any product that runs machine translation of text, and the marking-technique row removes a draft statement that a single technique could not currently satisfy Article 50(2). Neither change relaxes the four quality criteria themselves: paragraph 80 still requires the combined solution to be effective, interoperable, robust and reliable "to the legally required degree".

Chatbots and AI agents under Article 50(1)

Paragraph 30 sets four cumulative elements for Article 50(1): an AI system, intended to interact, directly, with natural persons. The final text adds two anti-avoidance sentences the draft did not have. The "mere possibility for humans to intervene or review the AI system's outputs should not be used to circumvent" disclosure, and products "blending AI-generated responses with human-curated content" need disclosure for the AI outputs unless those outputs are "properly reviewed and sent by humans as the main interlocutors".

On scope, paragraph 30 excludes closed physical environments such as industrial machinery, "backend machine-to-machine calls between AI systems (including AI agents) whose outputs are not intended to reach natural person", and virtual environments without direct contact with people. The examples in paragraph 31 put recommender systems, spam filters, automated translation or transcription tools and search and retrieval tools outside the provision. That is not a business-to-business carve-out: a support bot answering a client's staff still talks to natural persons. The chatbot disclosure page covers what the notice should say.

Agents carry the heaviest new text. Paragraph 31 requires agents capable of interacting with the people instructing them or with others to "disclose both their artificial nature and the person on whose behalf they are acting". Where the provider cannot tell in advance whether an agent will meet a person, it "should be designed at the architecture level, and instructed, to disclose itself" wherever a person is reasonably likely to be on the other side, "including where a natural person may represent or act for a legal entity". Agents should also disclose themselves to the people instructing them at key steps such as authorisation, reporting and validation, and at every new interaction. Paragraph 36 suggests verifiable AI identifiers, and footnote 21 points to attestations held in EU Digital Identity Wallets.

The obviousness exception is narrow. Paragraph 45 says it "should be interpreted restrictively" and limited to cases "where there is almost no doubt left". The examples that pass are narrow: code assistants for developers, trained internal staff, diagnostic tools for health professionals, appliance assistants and single-player game characters. Chatbots on platforms and helpdesks are listed as failing it.

Marking under Article 50(2): what is in and what is out

The final guidelines spell out in more detail what sits outside Article 50(2), adding examples to two exclusions the draft already had and one new exclusion for agent actions, and the marking duty page covers the techniques in more depth.

• Short and technical outputs. Paragraph 68 now gives examples for "a short sequence of numbers, symbols or letters": single words, image captions, alt-text, UI labels and icon-scale graphics. Source code is defined broadly and can include natural-language comments forming an integral part of it, and SQL, infrastructure-as-code, YAML and JSON configuration are expressly excluded as well.
• Closed-loop production. Intermediate outputs used only in closed-loop workflows, "for example for film, animation, games or advertising production", are outside, but the final output must be marked (para 68). The draft's version of this exclusion named only film production.
• Agent actions. Reasoning steps, chain of thought, web requests and browser actions are not synthetic content; an agent's text, image, audio or video output perceptible to people is (para 63).

Three narrow proportionality cases follow. Generative systems embedded in physical products with a closed, mainly instructive output, the vehicle navigation example, may use "less robust metadata markings" (para 86). Strictly technical, internal outputs of industrial or business-to-business applications, "excluding public and consumer-facing AI systems", need no marking if three cumulative conditions are met (para 87). Ephemeral real-time content in games or VR may be exempt where marking "is not technically feasible" and people are told the content is AI-generated (para 88).

On the duty itself, paragraph 70 is blunt: marking without available detection "will not suffice". Paragraph 76 requires reliance on "publicly-available industry standard detection solutions", ideally locally executable, with a provider's own detection allowed only until harmonised standards and a provider-agnostic solution emerge. Paragraph 74 lets providers rely on an upstream model provider's marking, without shifting responsibility for demonstrating compliance.

Deepfakes and the artistic, satirical and fictional regime

Paragraph 113 reads the Article 3(60) definition as four cumulative criteria: resemblance, to existing, persons, objects, places, entities or events, that would falsely appear authentic or truthful. The final text adds three clarifications that move borderline cases.

• Plausibility, not existence. A subject that "can plausibly exist or could have plausibly existed" is enough; dragons and elephants driving cars are outside because they "have no potential to mislead" (para 113).
• Photorealism is not the test. It makes a deep fake more likely "but photorealism alone is not determinative"; the question is whether the content is capable of deceiving in its deployment context (para 114). Background scenes and special effects in standard film production are unlikely to qualify, while fully AI-generated actors, digital replicas of deceased actors and de-ageing are likely to.
• Foreseeable audience, not every audience. Paragraph 115 says deployers need not consider dissemination "beyond the reasonably foreseeable audience", giving subscriber-only pages and corporate newsletters as examples. Where children or the elderly are foreseeably in that audience, deception of that part of it may suffice.

The final example list adds a synthetic CEO avatar congratulating staff, and a product image in advertising or packaging that makes the product look "more appealing or with improved quality than in real life". A real car against an AI-generated background is not a deep fake unless it misleads about the product.

The attenuated regime is narrow. Article 50(4) limits disclosure for "evidently artistic, creative, satirical, fictional or analogous" works to a form that does not hamper display or enjoyment. Paragraph 122 says the categories "should be interpreted strictly", excludes content whose nature is "exclusively informative or commercial and is recognisable as such", and adds the tie-break the draft lacked: where a deep fake combines informative and creative characters, "the informative character should always prevail". Paragraph 123 confirms these works "are not excluded from the transparency obligation". The examples in paragraph 124 treat a satirical image of a politician as eligible, and a teleshopping-style deep fake or a synthetic influencer demonstrating a sponsored product as not. The deepfake disclosure page covers where the label goes.

Paragraph 16 confirms that hosting services, online platforms and broadcasters that only disseminate third-party content are not deployers, and paragraph 14 adds that a company which merely commissions an advertising agency, without deciding whether and how it uses AI, is not a deployer either.

Public-interest text and editorial responsibility

The second limb of Article 50(4) requires deployers to disclose AI-generated or manipulated text "published with the purpose of informing the public on matters of public interest", unless the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility. Paragraph 131 reads "published" as accessible to an indeterminate, fairly large number of unrelated readers, so closed messaging groups and internal corporate networks are outside. Its examples put advertising copy and product descriptions outside, but not where they carry claims related to matters such as health, consumer safety or sustainability.

The final text raises the bar for the exception in three places and slightly adjusts a fourth:

1. Fact-checking is the floor. Human review means deliberate examination of substance by people with relevant knowledge, and "Fact-checking the accuracy of the content is a minimum requirement" (para 134).
2. Automated review does not count. Paragraph 135 adds "automated review processes" to spell-checks, the mere existence of an editorial policy and cursory approval as things that cannot satisfy the condition.
3. Post-sign-off AI voids it. "Any substantive AI intervention occurring after the human review or editorial control process has taken place will therefore cause the exception to become void" (para 136).
4. Name the responsible person or function. The draft already asked for the identity and contact details of the person with editorial responsibility to be publicly available, for example in website legal information or a colophon. The final text allows the legal person, the natural person "or the function" to be named (para 138).

Paragraph 140 ties the concept to Article 2(8) of the European Media Freedom Act, so media service providers can rely on existing editorial processes. Non-media deployers can still qualify: a reviewed AI-supported translation and a sustainability report reviewed by compliance staff are among the examples in paragraph 138.

Dates, retroactivity and who is reached

Paragraph 153 states that Article 50 applies from 2 August 2026 to all in-scope systems "regardless of their date of placement on the market". The only grandfathering is the Omnibus transition for Article 50(2) marking on generative systems placed on the market before 2 August 2026, with compliance due by 2 December 2026, and paragraph 153 adds that a system that is partly interactive and partly generative gets that transition for marking only, not for its Article 50(1) disclosure. The grace period page works through who qualifies. Footnote 50 confirms that the Article 111(2) high-risk grandfathering does not reach Article 50 duties.

Paragraph 154 governs content already in existence. Article 50(2) outputs and Article 50(4) deep fakes generated or manipulated before 2 August 2026 "do not need to be marked or labelled retroactively". For public-interest text the publication date decides: if text generated earlier is published on or after 2 August 2026, "they need to be labelled". Holders of older unlabelled deep fakes are encouraged to label them without "disproportionate efforts", such as auditing content databases or changing printed packaging.

On territorial reach, paragraph 13 catches third-country deployers where "the deployer itself foresees dissemination and use of the AI outputs in the Union", including "by posting deep fakes on the globally accessible internet", but not where content arrives through "channels that are unforeseeable and outside their control". Paragraph 10 says incidental, unforeseeable or unauthorised downstream use "should not alone trigger" the obligations for third-country providers not on the Union market. The guide for US companies sets that against Article 2.

What to check against the final text

If your Article 50 assessment was written against the draft, these are the points to re-run.

1. Translation features. If you treated machine translation of text as requiring 50(2) marking on the strength of the draft, paragraph 92 now places it among minor alterations. Summaries and rewrites that change style or meaning still need marking.
2. Agents. Check that agents disclose who they act for, not only that they are AI, and that they disclose at authorisation and reporting steps (para 31).
3. Satire and campaigns. Any creative work with an informative strand gets the standard label, not the lighter one (para 122).
4. Editorial workflows. Confirm fact-checking happens, that no AI rewrite follows sign-off, and that the responsible person or function is published (paras 134 to 138).
5. Back catalogue. Separate content by generation date for marking and deep fakes, and by publication date for public-interest text (para 154).
6. Code of Practice position. Signatories of a code assessed as adequate face supervision focused on adherence; non-signatories should expect more requests for information and are expected to run a gap analysis against the code (paras 147 and 148).

Paragraph 25 repeats that complying with Article 50 does not make a system lawful, and that the same system may be prohibited or high-risk. Biometric categorisation and emotion recognition notices are covered on the emotion recognition page. To see which Article 50 limbs and which other parts of the Act reach your own system, run it through the free triage classifier.