As of November 2, 2026, any company publicly distributing AI-generated images, videos, audio, or text must apply a specific label, machine-readable and/or visible as appropriate, under penalty of fines up to 7.5 million euros or 1% of worldwide turnover.
In brief
- November 2, 2026, marks the entry into force of the obligation for machine labeling of AI-generated content in Europe.
- Article 50 of the AI Act already mandates informing users during interactions with AI since August 2026; machine labeling is added in November.
- All companies publishing AI-generated images, videos, audio, or text are affected, particularly for marketing or editorial purposes.
- Two distinct obligations: visible labeling for deepfakes, and machine-readable labeling for all large-scale generative content.
- Non-compliance with Article 50 exposes companies to fines of up to 7.5 million euros or 1% of worldwide turnover.
- Major tools (OpenAI, Google, Mistral) integrate watermarking, but it is the responsibility of each company to ensure the correct transmission of the label.
As of **November 2, 2026**, all companies that produce and publicly distribute content generated by artificial intelligence (AI) – whether images, videos, audio, or text – must imperatively apply a specific label. This requirement, introduced by the **AI Act (Regulation EU 2024/1689)**, is part of the Digital Omnibus calendar and aims to strengthen transparency regarding the origin of digital content.
Machine labeling will thus become mandatory on this precise date, without conditions for postponement. Unlike other obligations of the regulation (notably those in Annex III on high-risk systems, whose application could be deferred to December 2, 2027), the date of November 2, 2026, is maintained and must be anticipated by all concerned organizations.
Article 50 AI Act: Scope and Detailed Timeline
**Article 50** of the AI Act regulation structures transparency obligations into four components, each with its own scope and schedule:
1. Disclosure during AI interactions (Art. 50§1)
Since **August 2, 2026**, any person interacting with an AI system (chatbot, virtual assistant, conversational agent) must be explicitly informed that they are communicating with a machine. This obligation does not apply if the context makes the artificial nature obvious.
2. Information on emotion recognition and biometric categorization (Art. 50§3)
AI systems capable of detecting emotions or classifying individuals based on biometric characteristics must, since **August 2, 2026**, clearly inform the affected persons of this automated analysis.
3. Visible labeling of deepfakes and manipulated content (Art. 50§4)
Any audio or visual content created or modified by AI to represent real people, places, or events must include a **visible label** indicating that it is artificial or manipulated. This label must be understandable to any recipient. This includes:
- AI avatar videos resembling existing people in marketing campaigns
- AI-generated photos used as customer testimonials or fictional portraits
- Synthetic voices imitating presenters or public figures
- News or product presentation videos created with AI speakers
- Images of real places modified or created by AI and presented as authentic
4. Machine-readable labeling (Art. 50§2) – the major innovation of November 2026
The main technical evolution concerns the obligation to integrate a **machine-readable label** (watermarking) into all generative content distributed on a large scale. This label, inserted directly into the file (image, audio, video, text), allows platforms, regulators, and detection tools to automatically identify the AI origin of the content.
> Key takeaway: Visible labeling (Art. 50§4) is the responsibility of the content publisher, while machine labeling (Art. 50§2) is the responsibility of AI solution providers. However, if the label is missing or removed during distribution, the company remains responsible for compliance.
Digital Omnibus: Ongoing Calendar Adjustments
The Digital Omnibus, whose political trilogue is expected to conclude around **April 28, 2026**, adjusts certain deadlines of the AI Act. Regarding Article 50:
- **Chatbots (Art. 50§1)**: Date confirmed for August 2, 2026
- **Deepfakes (Art. 50§4)**: Application also as of August 2, 2026
- **Machine labeling (Art. 50§2)**: The European Parliament proposes an entry into force on **November 2, 2026**, for systems already on the market, while the Council suggests February 2, 2027. The final date will be decided during the trilogue, but preparation must begin without delay.
For companies, the difference between November 2026 and February 2027 matters little: the essential point is to anticipate the integration of machine labeling into your publication processes and to verify that your AI tool providers are already compatible.
Progress status among major AI providers
Major industry players are gradually adapting their solutions:
- **OpenAI**: Deploys C2PA labeling on images from DALL-E, with integration of metadata attesting to the AI origin and file authenticity.
- **Google**: Uses SynthID technology in its Imagen (image) and Lyria (audio) models, offering an invisible watermark resistant to common modifications.
- **Adobe**: A pioneer of the C2PA standard, which it integrates into Firefly and Photoshop.
- **Mistral**: Implements labeling on its generative image models.
Warning: Labeling, often stored in metadata, can be lost during format conversions, compression, or uploads to platforms that strip this information.
Steps to prepare before November 2026
1. **Map your AI generation tools**: List all software and platforms used to produce images, videos, audio, or text (Midjourney, DALL-E, ElevenLabs, etc.) and verify the presence of C2PA labeling or equivalent in their exports.
2. **Audit your publication chain**: Identify points where labeling metadata might be deleted (conversion, compression, social media upload). Some platforms like LinkedIn or YouTube are starting to display C2PA labels, others are not.
3. **Apply visible labeling to deepfakes**: Any image or video of a real person generated by AI must include an explicit mention (e.g., "Generated by AI", "Synthetic voice"), in accordance with Art. 50§4 as of August 2026.
4. **Update your legal notices and T&Cs**: Clearly indicate the use of AI content and the nature of the labels applied on your site or application.
5. **Train your marketing and content teams**: Raise awareness among your staff regarding the identification of content subject to obligations and the verification of labeling before any publication.
For a personalized diagnosis of your obligations, the compaia diagnostic guides you in evaluating your usage and preparing your compliance documents.
AI content labeling addresses real issues: fighting deepfakes, preventing fake advertising, and election manipulation. For companies that use AI transparently, compliance remains accessible and low-cost. Conversely, ignoring these obligations exposes you to significant financial risks. Consult the AI Act timeline to plan your compliance before November 2026.
FAQ on AI content identification
On what date does AI-generated content labeling become mandatory in Europe?
The obligation for visible labeling of deepfakes (Article 50§4) takes effect on August 2, 2026. Machine labeling (Article 50§2) is expected for November 2, 2026 (European Parliament) or February 2, 2027 (Council), with the final date to be confirmed during the April 2026 trilogue.
Which AI content must be identified according to the AI Act?
Article 50 of the AI Act makes it mandatory to identify deepfakes and audio/visual content representing real people, places, or events (Art. 50§4), content generated on a large scale (machine labeling, Art. 50§2), and interactions with chatbots (Art. 50§1). Purely fictional works are generally exempt.
My company uses AI images in its communication: am I affected?
Yes, as soon as your images represent real people or are presented as authentic. Article 50§4 requires an explicit mention for any AI content featuring identifiable individuals. The compaia diagnostic helps clarify your obligations.
Is the C2PA standard sufficient to comply with the AI Act's machine labeling?
C2PA is a technical standard recognized for the integration of authenticity metadata. It meets the requirements of Article 50§2 if the metadata remains intact during distribution. The regulation does not prescribe a single standard, but C2PA is the reference adopted by the majority of providers.
What are the risks in case of missing labels on my AI content?
Infringements of Article 50 can lead to fines of up to 7.5 million euros or 1% of the total worldwide annual turnover, whichever is higher. Sanctions apply to both providers and publishers of non-compliant content. In France, ARCOM coordinates regulation with the CNIL and the DGCCRF.
Does AI labeling also apply to text generated by tools like ChatGPT?
The machine labeling obligation (Article 50§2) targets text generated on a large scale by AI. It is primarily aimed at model providers. For common uses (emails, articles proofread by a human), visible labeling is not systematic. However, for texts published without human supervision or simulating real testimonials, the obligation applies.