Korean companies are increasingly using generative AI to create advertising images and promotional videos, or publishing AI-generated text on their websites.
For content targeting the EU market, simply adding a short disclosure at the bottom may not always be sufficient to satisfy the applicable obligations.
The transparency obligations under Article 50 of the EU Artificial Intelligence Act (AI Act) will apply from 2 August 2026. On 20 July 2026, the European Commission adopted final guidelines clarifying the scope of application and relevant exceptions.
1. Distinguishing the Obligations of Providers and Deployers
The EU AI Act imposes different obligations depending on the role a company plays in the AI value chain.
An AI system provider is a business that develops an AI system, or has one developed, and places it on the EU market or puts it into service under its own name or trademark.
Companies that develop and provide generative AI services or image/video generation solutions are typical examples.
If a company integrates another provider’s AI model into its own service and offers it under its own name, it may also be assessed as a provider rather than merely a user.
An AI system deployer is a company or organization that uses an AI system under its authority for business purposes. This may include using generative AI to create advertising images or promotional videos, or publishing AI-generated text on a corporate website.
If an employee uses AI in accordance with company instructions, the company—not the individual employee—may be regarded as the deployer. By contrast, where a natural person uses AI purely for personal and non-professional purposes, deployer obligations generally do not apply.
| Category | Provider Obligations | Deployer Obligations |
| Main disclosure target | AI-generated output itself | Deepfakes and AI-generated text on matters of public interest |
| Disclosure method | Machine-readable marking or metadata | Human-readable text, icons, or audio notice |
| Primary verification | System or platform | Human reviewer |
2. Providers Must Apply Machine-Readable Marking
Providers of AI systems that generate synthetic audio, images, video, or text must ensure that it is technically detectable that the output was generated or manipulated by AI.
The required disclosure is different from text that is directly visible to users. Providers are expected to apply machine-readable methods to outputs, such as metadata or content provenance information. Such methods should be effective, interoperable, robust, and reliable, taking into account the characteristics of the content and the state of technology.
Certain types of output may be excluded from, or subject only to limited, marking obligations, including:
- Short numerical, symbolic, or textual outputs with little substantive significance;
- Source code or other output written in programming languages;
- Outputs processed automatically within a system and not exposed to people;
- Intermediate outputs in a closed production environment rather than final content; and
- Assistive editing functions that do not materially alter the meaning of the original content, such as spelling correction.
However, B2B AI outputs are not automatically exempt. Whether an exception applies must be assessed based on the nature of each output.
3. Deployer Disclosure Obligations: Deepfakes and Public-Interest Text
The fact that a provider has implemented technical marking does not mean that every company using AI must attach a visible disclosure to all AI-generated outputs.
Typical situations in which deployer disclosure obligations arise are deepfakes and AI-generated text published on matters of public interest.
A deepfake generally refers to AI-generated or manipulated image, audio, or video content that resembles real people, objects, places, entities, or events and could falsely appear authentic or truthful to a person.
The mere fact that content was created with AI does not automatically make it a deepfake. The analysis should consider factors such as:
- Similarity to a real person or subject;
- The substantive message conveyed by the content;
- The context in which the content is published; and
- The likelihood that users will believe the content is genuine.
For example, a promotional video using a synthesized version of a real CEO’s face and voice so that it appears the CEO personally made the statement may be highly likely to require deepfake disclosure.
By contrast, where users can clearly recognize from the outset that the content is fictional or staged, the analysis may differ. Disclosure obligations do not necessarily disappear for artistic, satirical, fictional, or similar works, but disclosure may be provided in an appropriate manner that does not interfere with the display or enjoyment of the work.
The key question is therefore not simply whether AI was used, but whether the content could realistically mislead users into believing it is authentic.
4. Scope of Disclosure for AI-Generated Text on Matters of Public Interest
Disclosure obligations may also arise where AI-generated or manipulated text is published for the purpose of informing the public on matters of public interest.
Representative areas may include:
- Politics and democratic processes;
- Public administration and public services;
- Judicial systems and law enforcement;
- Fundamental rights and public safety;
- Public health and the environment;
- Consumer safety; and
- Major economic, financial, scientific, or cultural issues.
For example, automatically publishing AI-generated news articles, policy explanations, financial market analysis, or legal and administrative information without human review may trigger transparency obligations.
However, product descriptions, ordinary marketing copy, or personal posts do not automatically become “public-interest text” merely because they were generated by AI. The purpose, substance, and expected audience of the publication must be assessed.
5. Meaningful Human Review and the Exception Requirements
Even for AI-generated text concerning matters of public interest, a separate disclosure obligation may not apply if certain requirements are met.
First, the content must undergo meaningful human review or editorial control.
Second, a natural person or legal entity must assume editorial responsibility for publication.
A merely formal check—for example, running a spelling or grammar review—is generally insufficient. The final guidance explains that simple proofreading or procedural confirmation may not qualify as meaningful human review and editorial control.
For review to be considered substantive, the responsible person should actually exercise authority and responsibility such as:
- Editing, approving, or rejecting content;
- Verifying facts and sources;
- Reviewing the appropriateness of legal or policy-related statements;
- Deciding whether the content should ultimately be published; and
- Assuming editorial and legal responsibility for published content.
Accordingly, companies should not merely state in internal guidelines that “a responsible person reviews the content.” They should also maintain records showing who reviewed the content, under what criteria, and when approval was given.
6. Placement, Method, and EU Icons
Where disclosure obligations apply to deepfakes or public-interest AI-generated text, users should be able to recognize that the content was AI-generated or manipulated at the time they are first exposed to it.
The disclosure should be distinguishable from other information, clearly recognizable, and accessible to persons with disabilities.
In practice, companies may use methods such as:
- Placing an “AI-generated” or “AI-edited” notice near an image;
- Inserting a notice at the beginning of a video or on-screen;
- Providing an audio notice before synthesized audio is played; or
- Placing a disclosure above or near the title of AI-generated text.
The EU has also introduced iconography through its transparency-related code of practice to distinguish between fully AI-generated content, AI-edited content, and content involving AI in the production process.
Use of such EU icons is not mandatory. A company may satisfy the disclosure obligation without using the icons, provided the disclosure itself meets the legal requirements. Conversely, merely using an icon does not automatically mean all transparency obligations have been fulfilled.
Provider-side machine-readable markings and deployer-side user-facing disclosures are separate obligations and must be assessed independently.
7. Whether Pre-Application Content Must Be Relabeled
Article 50 will apply from 2 August 2026, but whether content created before that date must be relabeled depends on the type of content.
For image, audio, and video deepfakes, outputs generated or manipulated before 2 August 2026 generally do not need to be retroactively labeled merely because they are published after that date. The relevant reference point is the date of generation or manipulation.
By contrast, for AI-generated text relating to matters of public interest, the relevant point may be the date of publication. Even if the content was written before the application date, disclosure obligations may arise if it is first published after 2 August 2026, unless the requirements for meaningful human review and editorial responsibility are satisfied.
Where content published before the application date is later reused in a new advertisement or new publication after the application date, the company should assess whether that reuse constitutes a new distribution rather than simple archival storage.
8. Status of the Digital Omnibus Transitional Period
As a general rule, AI systems falling within Article 50 should be prepared to comply with transparency requirements from 2 August 2026, regardless of when they were placed on the market.
However, under the EU’s proposed AI regulatory simplification measures under the Digital Omnibus, a transitional period until 2 December 2026 is expected to apply to certain machine-readable marking obligations under Article 50(2) for generative AI systems placed on the market or put into service before 2 August 2026.
The amendment has completed adoption procedures by the Council and Parliament, but formal publication in the Official Journal of the European Union has not yet been completed.
Companies should therefore not assume that compliance will automatically be deferred until December 2026. They should separately monitor final publication and the effective date and, as a practical matter, prepare based on the original application date of 2 August 2026.
9. Extraterritorial Application to Korean Companies
The EU AI Act does not apply only to companies established in the EU.
A non-EU business may fall within the scope of the Act where, for example:
- It supplies an AI system or general-purpose AI model to the EU market;
- It makes an AI system available for use in the EU; or
- The output of an AI system provided or deployed by a non-EU operator is used in the EU.
Accordingly, Korean companies should assess applicability where they provide AI services to EU users or where AI-generated content created in Korea is distributed to EU customers, subsidiaries, or platforms.
However, the mere fact that a company has an EU business partner or that its website can be accessed from the EU does not automatically mean Article 50 applies.
The analysis should focus on how the AI system and its outputs are actually supplied, deployed, and used in the EU market.
10. Penalties for Non-Compliance
A breach of general obligations under the AI Act, including Article 50, may result in administrative fines of up to €15 million or 3% of the undertaking’s total worldwide annual turnover for the preceding financial year, whichever is higher.
For SMEs, the lower of the fixed amount and the turnover-based amount may apply.
The actual amount of the fine may depend on factors such as:
- The nature, severity, and duration of the infringement;
- The scale of harm;
- Whether the infringement was intentional or negligent;
- The size of the undertaking; and
- The degree of cooperation with market surveillance and supervisory authorities.
Decent Law Firm International Practice
The EU AI Act’s transparency obligations do not create a blanket rule requiring every piece of AI-generated content to carry an “AI-generated” label.
AI system providers must assess whether technical marking obligations apply, while companies publishing AI-assisted content must determine whether the content falls within the categories of deepfakes or public-interest AI-generated text.
Where a company seeks to rely on a human-review exception, it must establish not only meaningful review and editorial control but also clearly defined editorial responsibility.
In practical terms, compliance is less about adding a single disclosure sentence and more about building an internal operating framework for content classification, review, approval, and recordkeeping.
Decent Law Firm’s International Practice advises Korean companies providing AI services or content to the EU market on matters including:
- Assessment of AI content usage and provider/deployer status;
- Classification standards for deepfakes and public-interest AI-generated text;
- Design of meaningful human review, editorial responsibility, and internal documentation procedures;
- Review of AI solution and outsourcing agreements and allocation of transparency obligations; and
- Assessment of extraterritorial application of the EU AI Act and related compliance strategies.
Companies preparing for the 2 August 2026 application date should review their current AI-content publication practices in advance.
This content is provided for general informational purposes only and does not constitute legal advice for any specific matter or business.