EU releases transparency guidelines for AI labeling and deepfakes
The European Commission issued new guidelines to clarify how companies must disclose AI interactions and label synthetic content under the EU AI Act.
The short version
- Companies must clearly inform users when they are interacting with an AI system.
- Synthetic content must be marked in a way that is both human-readable and machine-readable.
- Deepfakes are defined by their similarity to real people or places and their ability to feign authenticity.
- Fines for non-compliance can reach up to 15 million Euro or 3% of global annual turnover.
- Companies outside the EU must comply if they target an EU audience with deepfakes on the open internet.
Why it matters
These guidelines provide a concrete roadmap for businesses to avoid heavy fines while integrating AI into their products. For users, it means a higher standard of transparency regarding whether the content they see or the service they use is human-led or machine-generated.
New rules for AI transparency
The European Commission has published new guidelines to clarify the transparency requirements found in Article 50 of the EU AI Act. These guidelines, released in the EU Official Journal, serve as a primary orientation for companies to ensure they are following the law correctly. While the final legal interpretation rests with the European Court of Justice, these documents provide the specific practical steps for implementation.
The guidelines focus on four core areas of transparency. Organizations are required to design interfaces so that users can clearly recognize when they are interacting with an AI system. An exception is made only if the use of AI is completely obvious to an average consumer without any extra information.
For synthetic content, the rules require that providers make the content recognizable and mark it in a machine-readable format. This means the information must be accessible to both human eyes and automated software systems. The Commission emphasizes that a simple label is not enough; technical solutions must be effective, robust, and reliable based on what is technically feasible.
Defining deepfakes and labels
A significant portion of the guidelines clarifies what constitutes a deepfake under the new regulations. A deepfake is defined by four cumulative criteria: it must have a recognizable similarity to real people, objects, places, or events, and it must be objectively suitable to feign authenticity. These rules apply even if there is no intent to deceive, as the standard is based on the impact on the real audience.
The guidelines specifically highlight how these rules apply to advertising and product images. A manipulated image on a package or in an ad is considered a deepfake if it can mislead the audience about the actual appearance, properties, or benefits of a product. However, minor edits like color corrections, background extensions, or swapping a background do not require labeling as long as the product itself is not misrepresented.
Furthermore, the Commission clarified that invisible watermarks cannot replace visible labels. A label must be perceivable without any additional tools or help. These requirements are designed to protect vulnerable groups, such as minors or people with lower digital literacy, from being misled by synthetic media.
Exceptions and human oversight
The regulations also address specific types of content, such as AI-generated text of public interest. In these cases, an exception exists if the text undergoes substantial human review by experts who take on the editorial responsibility. This means that simple automated checks, general editorial guidelines, or basic spell-checking do not qualify as sufficient human oversight for this exemption.
If a text is rewritten using AI after it has been released, the privilege of the human oversight exception is completely lost. This ensures that the final version of public-interest content remains under human control. The rules are intended to balance the utility of AI tools with the need for accountability in public discourse.
Additionally, operators of systems for emotion recognition and biometric categorization must inform all exposed individuals beforehand. This includes a specific requirement to inform minors before such systems are used. These layers of disclosure are meant to ensure that users are aware of the specific types of data processing occurring in real-time.
Deadlines and global reach
The transparency obligations under Article 50 generally begin on August 2, 2026. However, there is a specific transition period for providers of generative systems that were already on the market before that date. These providers have until December 2, 2026, to implement the machine-readable marking requirements for their systems.
Further obligations for high-risk systems will take effect on December 2, 2027, and August 2, 2028. Companies that fail to comply with these rules face significant penalties. Fines can reach up to 15 million Euro or 3% of a company's global annual turnover, with the lower of the two values applying to small and medium-sized enterprises.
The reach of these rules extends beyond the borders of the European Union. Companies located outside the EU are still subject to these regulations if they publish deepfakes on the open internet and can reasonably be expected to reach an EU audience. To help with compliance, the Commission notes that following an approved code of conduct can make it easier to prove compliance and may lead to milder sanctions in the event of a violation.
What to watch
- How companies will balance 'machine-readable' requirements with user privacy.
- The development of standardized 'codes of conduct' to mitigate potential fines.
Try it
Check the product packaging or advertisements for your favorite brands to see if they use visible labels for AI-generated images.
Sources and links
- ReportingBornCity: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by BornCity · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
