EU Enforces Mandatory AI Labeling Rules Starting August 2

Starting Sunday, August 2, companies operating in the European Union face strict new requirements for how they deploy and label artificial intelligence. The next major phase of the EU AI Act takes effect, mandating that businesses disclose when users are interacting with AI and requiring visible labels and machine-readable watermarks on synthetic media.

The European Commission’s AI Office, operating alongside national authorities across the bloc, will begin enforcing Article 50 of the AI Act. This specific provision establishes transparency obligations that apply universally to certain AI systems, regardless of whether the software is classified as high-risk, minimal-risk, or prohibited.

The primary goal of the legislation is to ensure individuals know when they are dealing with artificial intelligence rather than a human, and when the images, video, or audio they consume have been artificially generated or manipulated.

The End of Undisclosed AI Chatbots

For consumer-facing technology, the most immediate change involves interactive AI systems. Providers of AI platforms that interact directly with individuals must now design them so users are explicitly informed they are engaging with a machine.

This disclosure must be clear and presented upfront. Companies can no longer hide AI disclosures deep within their terms of service. The European Commission noted that the rule applies directly to customer service chatbots, virtual assistants, and automated phone systems.

There is a narrow exception to this rule: if it is entirely obvious from the circumstances and context that the user is interacting with an AI system, a formal disclosure may not be strictly required. However, legal experts advise companies to default to active disclosure to avoid regulatory scrutiny.

Mandatory Watermarking and Deepfake Labels

The most complex technical obligations fall on developers and deployers of generative AI systems. Systems capable of generating synthetic audio, image, video, or text must now ensure their outputs are marked in a machine-readable format.

These digital watermarks must allow the content to be reliably identified as artificially generated or manipulated by subsequent detection tools. According to the European Commission’s guidelines, companies can meet this requirement using a variety of techniques, including embedded metadata, cryptographic provenance indicators, or digital fingerprinting.

The rules differentiate between the companies building the AI models (providers) and the businesses using them (deployers). Deployers of AI systems that produce deepfakes—defined as image, audio, or video content that resembles real people, places, or events—must visibly or audibly disclose that the content has been artificially generated or manipulated.

This labelling must be accessible and provided at the latest when a human is first exposed to the content. A general disclaimer on a social media profile or within a company’s general terms and conditions is insufficient. Each deepfake must be individually labeled. To facilitate this, the EU has released a set of standardized icons that companies can use to mark their generative outputs.

E-commerce and Retail Impact

The transparency rules carry significant implications for the online retail and advertising sectors. E-commerce sellers frequently use generative AI to create lifestyle scenes, swap product backgrounds, or generate synthetic models to display clothing.

Under the new framework, product imagery or marketing visuals created or significantly altered with generative AI tools and presented to EU consumers must be transparently disclosed.

Not all edits trigger the labeling requirement. The Commission’s guidelines exclude AI systems that merely assist with basic editing processes, such as minor color corrections, noise reduction, or background formatting, provided they do not generate entirely new synthetic elements. However, replacing a model’s face, materially altering a product’s appearance, or fabricating a realistic background context crosses the threshold into synthetic content that requires disclosure.

The financial risks for non-compliance are severe. Companies that fail to properly label AI-generated content face administrative fines of up to €15 million or 3% of their global annual turnover, whichever figure is higher.

Regulating General-Purpose AI Models

Sunday’s enforcement milestone also expands the European Commission’s oversight of the foundation models that power generative AI. The AI Office can now actively enforce the AI Act’s rules for providers of general-purpose AI (GPAI) models.

These models, characterized by their ability to perform a wide range of tasks and serve as the base for downstream applications and AI agents, are now subject to strict governance.

The rules place additional, heavier obligations on developers of the most advanced GPAI models that the EU determines pose “systemic risks”. Providers of these top-tier models must implement specific measures to address risks of large-scale harm, including potential vulnerabilities related to cybersecurity, biological and chemical threats, harmful manipulation, and risks to fundamental rights.

The Code of Practice

To help the industry navigate the technical complexities of watermarking and detection, the European Commission facilitated a multi-stakeholder process to draft a Code of Practice on Transparency of AI-generated Content.

On Friday, the Commission published an initial list of over 180 organizations that have signed the Code. The document outlines practical rules for providers regarding the marking and detection of synthetic content, and rules for deployers concerning the labeling of deepfakes and AI-generated text published on matters of public interest.

While adherence to the Code of Practice is voluntary, the transparency requirements under Article 50 are legally binding. Companies that sign the Code and implement its measures are granted a presumption of conformity, providing them with legal certainty across all EU member states.

Companies that choose not to adopt the Code must independently prove to market surveillance authorities that their alternative compliance methods are equally adequate.

A Phased Implementation

The August 2 enforcement date represents the third major phase of the EU AI Act’s gradual rollout. The legislation officially entered into force two years ago, in August 2024, triggering a 24-month transitional period designed to give businesses time to adjust.

The first set of rules took effect in February 2025, outlawing a specific list of “prohibited practices”. These bans targeted AI applications deemed to carry unacceptable risks, including systems that deploy subliminal techniques to manipulate human behavior, social scoring systems operated by public authorities, and untargeted scraping of facial images from the internet or CCTV footage to build facial recognition databases.

In August 2025, the initial governance requirements and foundational rules for general-purpose AI models became applicable.

While this weekend’s activation of the transparency rules covers the majority of the AI Act’s general provisions, the regulatory timeline extends further into the future. The final major phase is scheduled for August 2, 2027, when the strict conformity assessment requirements for certain high-risk AI systems—particularly those serving as safety components in regulated products—become mandatory.