Show your work: The EU's new Guidelines on AI transparency

A tale of two sittings

Anyone educated in the '80s or '90s will remember that there were two fundamentally different kinds of math test.

The first was the multiple-choice paper, the one where you were handed a pre-printed answer sheet and a soft pencil, and your sole task was to shade in a small oval beside the letter A, B, C, D, or E. The sheet was fed through a machine. The machine did not care about your reasoning, your elegant proof, or the inspired shortcut you had spotted on a prior response. It cared about the position of the graphite. You either shaded the right oval, or you did not. There was something faintly unsettling about the whole arrangement: You were being graded by a machine and you knew that no human would so much as glance at your paper.

The second kind of exam was the written paper. Here, the rules were different. A human being would read your answers, and the instruction at the top of every page was the same: show your work. You could arrive at the correct answer, but if the margin beside it was empty, if the steps that led you there were invisible, you would find a red line through it and a short comment reminding you that guesswork, however inspired, does not count. The injustice seemed obvious. You had the answer. What did it matter how you got there?

But the teacher was making a deeper point. Showing one’s work is not a formality. It demonstrates whether the answer was achieved through an act of understanding or random chance. An answer without proof is merely an assertion. An answer that shows its work is an argument (one that could still receive positive marks if you had solid lines of reasoning).

The European Union (EU), in its new Guidelines on artificial intelligence (AI) transparency, has drawn on both lessons at once. On July 20, 2026, the European Commission (Commission) published its Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act, which went into effect on August 2, 2026. The message to the AI industry is twofold.

First, as with the multiple-choice paper, it must be disclosed whether a person is dealing with a machine on the other side of an interaction so that users can make informed decisions.

Second, as with the written paper, the machine must show its work. If an AI system produces an output (a generated image, a synthetic voice, a block of text), the origin of that output must be marked, detectable, and verifiable.

The Guidelines are non-binding; if necessary, an authoritative interpretation may only be given by the Court of Justice of the EU. In addition, the Guidelines complement the Code of Practice on Transparency of AI-Generated Content (Code of Practice), published on June 10, 2026, which we discussed in a previous client alert using the analogy of hallmarking. If the Code of Practice shows how to walk the path, the Guidelines say where its edges lie.

Four obligations, one principle

Article 50 of the AI Act contains four distinct transparency obligations, two applying to providers and two applying to deployers, each aimed at a different type of AI system or output:

• First, providers of AI systems that interact directly with natural persons (e.g., chatbots, AI agents, voice assistants, avatars) must design and develop those systems so that individuals are informed they are interacting with an AI system (Article 50(1)).
  
  
• Second, providers of AI systems that generate or manipulate synthetic content (e.g., audio, image, video, or text) must ensure that the outputs are marked in a machine-readable format and are detectable as artificially generated or manipulated (Article 50(2)).
  
  
• Third, deployers of emotion recognition and biometric categorization systems must inform exposed individuals that those systems are in operation (Article 50(3)).
  
  
• Fourth, deployers who use AI to generate or manipulate deepfakes, or to produce text published on matters of public interest, must disclose the artificial origin of the content (Article 50(4)).

These four obligations can apply cumulatively to a single AI system, potentially engaging the responsibility of different actors along the value chain. For example, a financial services firm could deploy an AI avatar adviser on its website, where:

• Users speak to a realistic avatar (Article 50(1))
  
  
• The avatar's speech and appearance are AI-generated (Article 50(2))
  
  
• The system analyzes facial expressions and tone of voice to assess customer stress (Article 50(3)), and
  
  
• The system also generates market commentary published on investment issues of public interest without substantive human editorial review (Article 50(4) public-interest text disclosure).

Such a single use case could potentially engage every transparency category, but responsibility may be split between the AI vendor and the organization using the system, depending on the functionality in question.

Knowing you are talking to a machine

The first obligation – ensuring people are told they are interacting with AI – appears to be simple, but there are some nuances.

Four cumulative elements must be present:

1. The system must qualify as an AI system (which excludes simple rules-based auto-replies)
   
   
2. It must be intended to interact (i.e., designed for bidirectional exchange with natural persons, enabling them to provide input (e.g., text, voice, or physical actions) and receive contextual output in return)
   
   
3. The interaction must be direct (i.e., indirect or mediated interaction is excluded, such as where a customer service representative separately uses an AI assistance tool to improve communications with a person), and
   
   
4. It must interact with natural persons.

The purpose, as Recital 132 of the AI Act explains, is to enable individuals to take informed decisions, avoid overreliance, and calibrate their trust.

The treatment of AI agents is noteworthy. AI agents are covered by Article 50(1) if they are intended to interact with the persons instructing them or with other natural persons in the execution of tasks, such as making bookings, managing correspondence, negotiating or concluding contracts, or executing purchases. They must disclose both their artificial nature and the person on whose behalf they are acting.

Where a provider cannot reliably determine in advance whether its AI agent will directly interact with a natural person, the AI agent should be designed at the architecture level to disclose itself as AI in every situation where such interaction is possible. Like the multiple-choice test, the user must know that a machine, and not a human, is grading its responses.

The Guidelines contemplate a range of disclosure techniques, such as textual labels, auditory statements, visual cues, and multimodal combinations. They also highlight what does not suffice for disclosure, such as disclosures buried in terms and conditions, and machine-readable markings invisible to the user. The Guidelines state that vague references to an "assistant" will not suffice, nor will a general banner reading "Services on this website use AI." The notification must be specific, contextual, and delivered at the point of interaction.

There is an exception for interactions where the AI origin is deemed “obvious” to a reasonably well-informed, observant, and circumspect natural person, but it is interpreted restrictively. The general awareness that chatbots and AI agents exist does not mean people will recognize them in practice. The bar is high: There must be “almost no doubt left” about the artificial nature of the interaction.

An AI-powered code assistant available only to professional developers may qualify as obvious, as may an internal employee-facing assistant used by properly trained, AI-literate staff. But a chatbot embedded in a consumer helpdesk does not, because users may perceive replies as human-generated. Nor does a robotic companion pet that looks highly similar to a real animal, nor AI avatars in immersive environments where children, the elderly, or persons with disabilities may struggle to tell human from machine. Assumptions regarding what is obvious may differ materially depending on the audience and their level of digital literacy. In riskier contexts, a single notification at the start will not be enough. Where users may form emotional attachments or dependencies (AI companions, for instance) or where the system provides financial advice, legal assistance, or health guidance, periodic reminders and context-aware disclosures could be necessary.

Providers must also ensure disclosure whenever the system is asked questions about its own nature, or where the exchanges suggest the person is likely to be misled about the AI-origin of the interaction. If someone asks the machine if it is a machine, it must answer honestly.

In companion and high-stakes advisory settings, the guidelines aim for calibrated trust, not just a box ticked to confirm that a machine announced itself. Just because the oval was shaded does not mean that the reader understood the question.

Marking the work

The second obligation concerns synthetic content, such as AI-generated or manipulated audio, images, video, or text. Notably, content manipulation can include the alteration of already existing material by an AI system, such as modifying an image or voice recording. Article 50(2) requires providers to ensure such content is marked in a machine-readable format and detectable.

Marking and detection are two distinct but interlinked requirements. Marking without detection is writing your working in a cipher nobody can read; detection without marking is looking for footprints on a surface that has since been raked smooth.

Providers may rely on a single technique or a combination, such as watermarks, metadata, cryptographic provenance methods, logging, fingerprinting – so long as the overall technical solution meets four requirements. It must be effective, interoperable, robust, and reliable.

The Guidelines carefully define these four requirements. Effectiveness means the solution enables people to distinguish AI-generated content. Reliability means it works accurately across the variety of content that the system produces. Robustness means it survives both common alterations and adversarial attacks. Interoperability means the marks can be detected across systems, regardless of which provider struck them.

Not everything is caught. Source code, short sequences of numbers or symbols, machine-to-machine communications not perceived by humans, and outputs used in closed-loop industrial environments all sit outside scope. A proportionality-based exemption also covers narrowly defined business-to-business and industrial applications, provided the output is strictly technical, intended only for a limited professional audience, and not shared externally. Real-time ephemeral content (e.g., video games or virtual reality consumed immediately without being stored) may also be exempted where marking is not technically feasible and in-experience disclosure is provided instead.

A mark, however, is not a certificate. It can show that content was generated or manipulated by AI, but it says nothing about whether that content is accurate, lawful, authorized, original, or untouched. Nor is the absence of a mark proof of a human hand. Marks may be technically infeasible, exempted, stripped out, worn away by ordinary processing, or simply missing from older content. Detection is one line of work, not the final answer.

In addition, the "assistive function for standard editing" draws a boundary that may be a concern for organizations. Standard editing (e.g., fixing grammar, adjusting layout, formatting for accessibility, removing dust spots from a dirty lens, correcting red eye from flash photography, rescaling a video clip) sits safely outside the marking obligation.

But the moment the AI begins to change the substance (e.g., summarizing text, replacing faces, synthesizing speech in a specific person's voice, generating realistic video depicting events that never occurred, or altering body shape or skin color) it crosses the line into manipulation and the marking obligation takes effect.

The Guidelines also contain clarification about what does not count as synthetic content at all. Music playlists, recommender systems, and search results merely arrange existing material rather than generating anything new; therefore, these functions are out of scope. Furthermore, observations and recordings from the physical world, such as GPS data from vehicles, readings from smart meters, and manufacturing data from robot sensors are also out of scope.

AI agents present a more nuanced case: Their actions are caught if they produce content perceptible by natural persons, but intermediate processing steps (e.g., reasoning, chain-of-thought) that no human is intended to see do not need to be marked. The agent’s private working-out, as it were, is not the same as the work shown in the margin.

The deepfake question

The fourth obligation shifts the focus from providers to deployers, addressing two specific categories: deepfakes and AI-generated text on matters of public interest.

First, the Guidelines help clarify what kind of activity qualifies someone as a deployer. Per the AI Act, individual using a generative system in a private, non-professional capacity is not considered a deployer. A system must be used in a commercial, occupational, or freelance activity for its operator to qualify as a deployer. But note the limit of the carve-out. It lifts the individual’s deployer obligations only. The provider’s separate duty to mark in-scope outputs remains firmly in place.

The term “deepfake” is defined in the AI Act and given further clarification under the Guidelines, which provide that the following criteria must all be met:

1. The content must resemble existing persons, objects, places, entities, or events
   
   
2. The subject must be something that exists, could plausibly exist, or could plausibly have existed, and
   
   
3. The content must falsely appear to a person to be authentic or truthful.

The “false appearance” criterion is assessed objectively, weighing the level of resemblance, the substantive message, the deployment context, and the audience’s expectations. The practical examples are well worth reading for any organization in the creative, media, or advertising sectors, because the boundary lines contain some nuance. An AI-generated image of a sphinx flying over the Eiffel Tower is not a deepfake, because it defies the laws of physics. An AI-generated video of mice arguing in human language about the best type of cheese is not a deepfake, because biology does not recognize the scenario.

But an AI-generated video of a synthetic avatar of a company CEO congratulating employees is a deepfake, as is an AI-generated audio that clones the voices of a newspaper podcast’s regular presenters. AI-generated special effects in a film, where no one in the audience expects the backdrop to be real, may not trigger the obligation, but fully AI-generated actors, digital replicas of real or deceased performers, and simulated performances are treated differently. The dividing line is not whether AI was used, but whether the output could reasonably mislead someone about what they are looking at.

The audience-expectation element adds a further layer. Deployers may wish to consider not only their intended audience, but also incidental audiences – if children, the elderly, or people with lower digital literacy might encounter the content, their greater susceptibility to being misled may be sufficient to bring it within scope.

There is, however, a limit: Deployers are not expected to plan for viral redistribution by unknown third parties. For example, content shown solely on a subscriber-only website or in a corporate newsletter does not require the deployer to assume broad public accessibility by default.

According to the Guidelines, transparency requirements are lighter for evidently artistic, creative, satirical, fictional, or analogous works: Deployers must still disclose the AI origin but may do so in a manner that does not hamper the display or enjoyment of the work. For example, disclosures may be allowed in end credits, exhibition notes, or equivalent notices.

The word “evidently” is key: The artistic or fictional character must be apparent to the viewer. Ambiguous content does not qualify, and where a piece combines informative and creative elements, the informative character prevails and full labeling applies.

For AI-generated text on matters of public interest, the Guidelines make an important exception: Text that has undergone genuine human review or editorial control, where a natural or legal person holds editorial responsibility for the publication, does not require a disclosure. However, the review must include elements such as fact-checking, editorial judgment, the authority to approve, alter, or reject content on substantive grounds. Spell-checking and grammar review may not meet this threshold.

The Commission’s Q&A contains information on the scope of “matters of public interest,” which includes politics and democratic processes, public administration, the administration of justice, fundamental rights, public security, public health, environmental protection, consumer safety, and any economic, financial, scientific, or cultural development that may be a relevant subject of public debate.

For example, an AI-generated summary of a newspaper article discussing a town council decision or an AI-manipulated corporate report on a listed company’s website containing investor information could be in scope. Whereas AI-generated fantasy novels, product advertisements, or a news summary generated by a chatbot visible only to the user who prompted it would fall out of scope.

For organizations that rely on the human-review exception, the Guidelines make clear that where AI systems are used to modify, supplement, or reformulate content after editorial sign-off, the resulting content must be treated as AI-generated or manipulated, and the exception becomes void. The order of operations matters: A human must have the last substantive word. If the machine rewrites after the editor has signed off, the working has changed and the examiner needs to know.

Nor can that last word be a rubber stamp. According to the Guidelines, review counts only where the reviewer has the competence to judge the content, the information and time to do so, and genuine authority to change or reject it. A sign-off from someone who cannot realistically depart from the output is considered insufficient.

The extraterritorial pencil

As with other EU Digital Decade laws, the Guidelines clarify the extraterritorial reach of these obligations. If the output of their AI system is used in the EU, providers must comply with the transparency obligations whether they are established within the EU or in a third country.

Deployers must comply if they are established in the EU, or if they are in a third country but anticipate that their content will reach EU audiences, including by posting deepfakes on the globally accessible internet.

However, the Guidelines note that incidental, unforeseeable, or unauthorized downstream use should not alone trigger obligations for third-country providers who have not placed their systems on the EU market. Similarly, third-country deployers are not bound where content reaches EU audiences through channels that are unforeseeable and outside their control.

For organizations outside the EU, the question is not where the technology was built. For providers, it is whether the system’s output is used in the EU. For deployers, it is whether the content is foreseeably directed at an EU audience. Geography of development is not necessarily determinative. Reach is what counts.

Open-source AI systems are not exempt from these obligations. The Guidelines confirm that AI systems released under free and open-source licenses remain within the scope of Article 50 and must comply with the relevant transparency obligations. However, providers of open-source AI components (e.g., software, data, models, tools, or services that do not in themselves constitute an AI system) are not directly covered, though they are encouraged to design those components in a way that facilitates downstream compliance.

The watched and the watchers

The third obligation, emotion recognition and biometric categorization, may require special attention.

Under the AI Act, a biometric categorization system is an AI system that uses biometric data, such as facial images or voice characteristics, to assign individuals to particular categories, rather than identifying or authenticating who they are, such as a facial analysis tool estimating age bracket (e.g., 18–25, 26–35), or a voice analytics tool classifying callers by accent or gender. An emotion recognition system attempts to infer emotions or intentions (e.g., happy, angry, stressed, frustrated). A single system could do both.

Article 50(3) requires deployers of these systems to inform natural persons who are exposed to them that they are in operation. The obligation applies regardless of whether the exposure is in real time or operates ex post. Crucially, the Guidelines note that compliance with this information obligation does not, in itself, render the use of such a system lawful, nor does it legitimize intrusive or discriminatory uses that might be prohibited under Article 5 of the AI Act or other Union law (particularly the General Data Protection Regulation). Transparency is necessary, but it is not sufficient.

The Q&A document confirms that the obligation does not require deployers to disclose the purpose of the system’s operation, merely that it is operating (which disclosure must be clear, distinguishable, and accessible to the relevant audience). Practical examples include a pop-up message before a computer game indicating that the player’s face is being recorded to capture emotions, or a visible notice at each entrance to an exhibition room informing visitors that their facial images are captured for age categorization in addition to data protection transparency requirements where the deployer is a controller.

The value chain and the vanishing mark

Modern AI supply chains rarely involve a single actor. Foundation model providers, platform hosts, system integrators, and downstream deployers may all handle the same content before a human being sees it. The allocation of transparency obligations depends on the role each actor performs.

According to the Guidelines, providers must embed transparency into the system’s design from when it is placed on the market or put into service. A downstream provider or deployer who builds on someone else's model retains its own responsibility for compliance. The obligation cannot be outsourced by outsourcing the technology.

Actors along the value chain who are not themselves providers or deployers – platforms, content management systems, digital asset managers – are encouraged to preserve markings as content passes through their systems. The mark must survive the journey from model to end user.

For procurement teams, this translates into a practical due diligence question at the buying stage: Can the vendor explain which outputs are marked, which technique is used, whether detection can be accessed by an application programming interface (API) or specification, and whether markings survive the ordinary transformations expected in the customer’s environment? The answer should be testable.

Not an island: Interplay with the wider regulatory landscape

Transparency is not the endpoint of compliance. Following the Digital Omnibus on AI, certain AI-generated content may be prohibited outright under Article 5 of the AI Act irrespective of compliance with Article 50 transparency requirements. In such cases, marking or disclosure cannot legitimize content whose creation or dissemination is itself prohibited.

Moreover, Article 50 does not operate in a vacuum. The transparency obligations sit alongside data protection, consumer protection, intellectual property, media law, and the Digital Services Act (DSA), complementing rather than displacing those frameworks.

The interplay with the DSA is considerable. Machine-readable marks embedded under Article 50(2) can help providers of very large online platforms and search engines meet their own obligations under Articles 34 and 35 of the DSA to identify and mitigate systemic risks arising from AI-generated content, including risks to democratic processes, civic discourse, and electoral processes. Consumer protection law adds another layer: The Unfair Commercial Practices Directive already prohibits misleading actions or omissions about the main characteristics of a product or service, and the Consumer Rights Directive requires clear pre-contractual information, obligations that apply regardless of whether the AI interaction might count as “obvious” under Article 50(1).

For deployers creating deepfakes, data protection, intellectual property, and personality rights all layer on top of the transparency obligation. A deepfake incorporating personal data relating to an identifiable living person triggers the deployer's duties as a data controller, including the requirement for an appropriate legal basis for the processing. Incorporating protected copyright or trademark material may trigger those obligations too.

The transparency label does not make an unlawful deepfake lawful; it merely tells the viewer what they are looking at, which is another example of a step that is necessary but not sufficient for full compliance with the wider legal requirements. The marking obligations themselves must comply with data protection principles: Providers must not process information about the creator of the content for marking purposes, and any personal data used must be deleted once detection has served its purpose.

The liability landscape is evolving in parallel. Directive (EU) 2024/2853, which extends product liability rules to software and AI, must be transposed by Member States by December 2026.

Two instruments, one framework

Deployers and providers are encouraged to consider how Guidelines and the Code of Practice fit together. The latter, published on June 10, 2026, is a voluntary tool for demonstrating compliance with the marking and labeling obligations under Article 50(2), (4), and (5) of the AI Act. It has been assessed as adequate by both the Commission and the AI Board. Signing up is the straightforward way of showing your work to the regulator, which may provide legal certainty and predictability in a single step. For more information on the Code of Practice, please see our previous client alert.

Organizations that choose not to adhere must demonstrate compliance through alternative adequate means and may face a heavier supervisory touch. For the remaining obligations, interactive AI under Article 50(1) and emotion recognition under Article 50(3), there is no code of practice at all, and organizations must chart their own course with the Guidelines as their map.

Enforcement and the cost of empty margins

The penalty framework gives these transparency obligations teeth. Non-compliance could attract administrative fines of up to EUR15 million or three percent of total worldwide annual turnover, whichever is higher under the AI Act. However, it is important to remember there could be multiple infringements, such as additional infringements of data protection requirements.

EU institutions, bodies, and agencies face fines of up to EUR750,000. Proportionality is taken into account for small and medium enterprises and start-ups, but the maximum figures could still be substantial.

Enforcement sits principally with national market surveillance authorities, with the AI Office responsible for AI systems built on general-purpose AI models where the same entity provides both the model and the system, and the European Data Protection Supervisor responsible for EU institutions.

A limited transitional grace period exists for Article 50(2) whereby generative AI systems placed on the market before August 2, 2026 have until December 2, 2026 to comply with the marking and detection obligation. Content generated before that date does not need to be marked or labeled retroactively, although the Commission encourages it.

Next steps

With the August 2, 2026 application date now past, organizations may wish to consider the following actions as a matter of priority:

• Map your obligations: Determine whether you are a provider, a deployer, or both, and identify which of the four Article 50 obligations apply to your AI systems and their outputs.
  
  
• Assess your disclosure mechanisms: For interactive AI systems, ensure that notifications are specific, contextual, and provided at the point of interaction – not buried in terms and conditions.
  
  
• Implement or audit marking solutions: For generative AI systems, ensure that outputs are marked with machine-readable marks and that corresponding detection tools are available, effective, interoperable, robust, and reliable.
  
  
• Review your value chain: Audit the contracts and operational processes along your AI supply chain to ensure that responsibility for marking, preservation, and detection is allocated explicitly rather than assumed.
  
  
• Consider the Code of Practice: Adherence to the Code offers a path to demonstrating compliance; non-adherence may invite greater scrutiny from regulators.
  
  
• Prepare for deepfake and text obligations: Deployers publishing AI-generated content should pilot the EU icon, review editorial control processes, and ensure disclosure is clear, distinguishable, and accessible.

For more information, please visit DLA Piper's focus page on AI, the Algorithm to Advantage portal, and DLA Piper's AI Scorebox. If you would like to discuss any of the issues raised in this article, please contact the authors or your usual DLA Piper attorney.