EU AI Act Article 50: What Went Live on 2 August 2026
On 2 August 2026, Article 50 of the EU Artificial Intelligence Act became enforceable across all 27 EU member states. The European Commission published implementation guidelines on 20 July 2026 to clarify what the obligations require in practice. Article 50 is the transparency chapter of the EU AI Act, covering the disclosure of AI involvement in consumer interactions, machine-readable marking of synthetic content, and labelling of deepfake media. Its scope is broad: the obligations apply to any provider or deployer whose AI systems interact with EU users, regardless of where the company is incorporated or headquartered. A company based in Bengaluru, Tel Aviv, or San Francisco that deploys a chatbot serving EU residents is within scope.
The Three Core Obligations
The first obligation covers conversational AI. Any provider deploying an AI system designed to interact directly with users — chatbots, virtual assistants, customer service bots, voice agents — must ensure users are informed they are interacting with an AI system before or at the start of the interaction. The disclosure must be clear and upfront. The obligation applies unless the artificial nature of the system is entirely self-evident from context, which the Commission's guidance indicates is a narrow exception. A customer service chatbot with a human name and conversational tone does not qualify for the self-evident exception simply because AI chatbots are widely known to exist.
The second obligation covers synthetic content at the generation layer. Providers of generative AI systems producing text, images, audio, or video must mark outputs in a machine-readable format — a digital watermark, cryptographic metadata, or C2PA-style provenance signal — so that third parties and detection tools can identify the content as artificially generated or manipulated. The Commission's guidance specifies that the marking must survive common distribution steps including compression, re-encoding, and social media upload.
The third obligation covers deepfakes specifically. Any entity using AI to generate or manipulate synthetic media depicting realistic representations of real people, places, or events must display a clear, human-perceivable label on the content — not only a machine-readable marker, but a label that a viewer can see and understand as indicating AI involvement. Media publishers, platforms, and applications that host such content without visible labelling are within scope.
The Grace Period: Machine-Readable Marking Until December 2026
The AI Omnibus provisional agreement of May 2026 included a four-month transition period for the machine-readable content marking requirement. Providers of generative AI systems already on the market before 2 August 2026 have until 2 December 2026 to implement watermarking infrastructure. The grace period applies only to the machine-readable technical marking. Chatbot disclosure and deepfake labelling carry no transition period — both apply from 2 August 2026 with immediate effect.
Penalties: Up to 15 Million Euros or 3% of Global Turnover
Non-compliance with Article 50 obligations can result in fines of up to 15 million euros or 3 per cent of worldwide annual turnover, whichever is higher. Enforcement sits with national market surveillance authorities in each EU member state, coordinated through the AI Office. The European Commission's July 2026 guidance explicitly states that enforcement will be outcome-focused: regulators will assess whether users were actually informed, not whether a disclosure clause appeared in a terms of service document.
What Software Teams Need to Build
For engineering and product teams building AI products that serve any EU users, Article 50 compliance is concrete implementation work. Chatbot and conversational AI deployments need a disclosure mechanism that fires visibly before the first AI-generated response in any session — a system prompt instruction or back-end flag is not sufficient if users cannot see the disclosure. Generative AI outputs need watermarking or C2PA provenance metadata embedded at generation time and durable across compression and re-encoding. Any application that generates synthetic media depicting real people needs a human-readable label on the output before it reaches users.
For Indian software companies and product teams that build AI applications for European clients or serve EU consumers directly, the compliance window is active now. The machine-readable watermarking deadline is 2 December 2026 — four months away. Teams that have not scoped watermarking implementation should begin now: robust content provenance infrastructure requires changes at the model output layer, not just at the user interface.
The Bottom Line
Article 50 of the EU AI Act took effect on 2 August 2026, with European Commission enforcement guidelines published 20 July 2026. The three obligations are: chatbot disclosure — users must be told they are interacting with AI before the interaction starts; synthetic content machine-readable marking — watermarks or provenance metadata on AI-generated text, images, audio, and video; and deepfake labelling — human-perceivable labels on AI-generated or AI-manipulated realistic media. Machine-readable marking carries a grace period to 2 December 2026 for systems already on the market; chatbot disclosure and deepfake labelling apply from 2 August 2026 with no transition time. Fines run up to 15 million euros or 3 per cent of worldwide annual turnover. For Indian software teams building AI products that serve EU users, concrete implementation work — disclosure mechanisms, watermarking infrastructure, deepfake labels — is now a live compliance requirement.
Frequently Asked Questions
What is EU AI Act Article 50 and when did it take effect?+
Article 50 of the EU Artificial Intelligence Act is the transparency chapter, covering three categories of obligation: disclosure of AI involvement in consumer interactions, machine-readable marking of synthetic content, and labelling of deepfake media. It took effect on 2 August 2026 across all 27 EU member states, with European Commission enforcement guidelines published on 20 July 2026. The obligations apply to any provider or deployer whose AI systems interact with EU users, regardless of where the company is incorporated or based. Non-compliance can result in fines of up to 15 million euros or 3 per cent of worldwide annual turnover, whichever is higher.
What are the three main transparency obligations under EU AI Act Article 50?+
The three obligations are: first, chatbot and conversational AI disclosure — providers must ensure users are informed they are interacting with an AI system before or at the start of the interaction, with a narrow exception only where the artificial nature is entirely self-evident. Second, synthetic content machine-readable marking — providers of generative AI systems producing text, images, audio, or video must embed machine-readable markers such as digital watermarks or C2PA-style provenance signals in their outputs, durable enough to survive common compression and re-encoding steps. Third, deepfake labelling — any entity generating or manipulating synthetic media depicting realistic real people, places, or events must display a human-perceivable label indicating AI involvement, not only a machine-readable marker.
Is there a grace period for the EU AI Act Article 50 watermarking requirement?+
Yes, but only for the machine-readable content marking obligation. The AI Omnibus provisional agreement of May 2026 granted providers of generative AI systems already on the market before 2 August 2026 a four-month transition period, running until 2 December 2026, to implement technical watermarking infrastructure. The other two obligations — chatbot disclosure and deepfake human-perceivable labelling — carry no grace period and apply from 2 August 2026 with immediate effect. Teams building new generative AI products launched after 2 August 2026 must implement machine-readable marking from the start; only pre-existing products qualify for the December transition window.
What do Indian software teams need to implement to comply with EU AI Act Article 50?+
Indian software companies building AI products that serve any EU users need to implement three categories of compliance work. For conversational AI and chatbots, they need a disclosure mechanism that fires visibly at the start of every user session, surfacing clearly in the product interface — not buried in terms of service or handled solely at the system prompt level. For generative AI producing text, images, audio, or video, they need watermarking or C2PA provenance metadata embedded at generation time and durable across common distribution and compression steps. For any application generating synthetic media depicting real people, they need a human-readable label displayed on the output. Teams without watermarking infrastructure should begin scoping it now, as the December 2 deadline is approximately four months away and robust provenance systems require implementation at the model output layer.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.