EU & cross-borderUpdated 22 November 20269 min read

AI Act Article 50 Checklist for Indian GenAI Startups

A task-level checklist for Indian GenAI startups to meet EU AI Act Article 50 transparency duties: chatbot disclosure, machine-readable marking, deepfake and public-interest text disclosure.

By Suhas Bhairav · Reviewed against the DPDP Rules, 2025 (G.S.R. 846(E), 13 Nov 2025)

Provision
Article 50, Regulation (EU) 2024/1689
Applies from
2 August 2026
Provider duties
AI interaction disclosure; machine-readable output marking
Deployer duties
Deepfake, emotion recognition, public-interest text disclosure
Maximum fine
€15m or 3% of worldwide turnover
Supporting guidance
Commission-facilitated code of practice on marking and labelling

Quick answer

Article 50 of the EU AI Act, applying from 2 August 2026, requires providers to tell people they are interacting with AI unless obvious, and to mark synthetic audio, image, video and text outputs in a machine-readable, detectable format. Deployers must disclose deepfakes, emotion recognition and AI-written public-interest text. Indian GenAI startups with EU users are in scope via Article 2.

  • Article 50 applies from 2 August 2026 and reaches Indian providers whose systems are used in the EU or whose outputs are used there.
  • Providers must design interactive AI so users know it is AI, unless obvious from context.
  • Providers of generative systems must mark synthetic audio, image, video and text in a machine-readable, detectable way.
  • Deployers must disclose deepfakes, inform people exposed to emotion recognition or biometric categorisation, and label AI-written public-interest text unless human-edited.
  • Breaching Article 50 can attract fines up to €15m or 3% of worldwide turnover.

Who this is for: Indian GenAI startup founders, Product managers for chat, image, video and voice tools, ML and platform engineers, API providers serving EU developers, Trust-and-safety leads

What does Article 50 require?

Article 50 is the AI Act's transparency article. It applies to many AI systems that are not high-risk, which is why it matters to most GenAI startups. It has applied since 2 August 2026 under the Regulation as adopted.

#DutyWhoTrigger
1Inform people they are interacting with an AI system, unless obviousProviderAI system intended to interact directly with natural persons
2Mark outputs in a machine-readable format, detectable as artificially generated or manipulatedProviderSystem generates synthetic audio, image, video or text
3Inform people exposed to emotion recognition or biometric categorisationDeployerDeploys such a system
4Disclose that content is artificially generated or manipulatedDeployerDeploys a system producing deepfakes
5Disclose AI-generated text on matters of public interestDeployerPublishes such text to inform the public, unless human-reviewed with editorial responsibility

Indian startups are usually providers (duties 1 and 2) for their own products, and sometimes deployers (duties 3–5) when they use AI to produce content they publish themselves, such as marketing videos.

The Commission is facilitating a code of practice on marking and labelling AI-generated content. Check its status on the Commission's AI Act page; following it is a practical route to showing compliance.

Are you in scope?

Article 2 extends the AI Act to providers outside the EU who place AI systems on the EU market, and to providers and deployers outside the EU where the system's output is used in the EU.

Quick scope test for Indian GenAI startups

  • Do EU residents sign up for your app, or do EU businesses call your API? → In scope as provider.
  • Do EU customers embed your model in their products? → In scope as provider; your customers are deployers (or providers of their own systems).
  • Do you generate content for EU clients who publish it? → Your output is used in the EU; check both roles.
  • Is your product India-only, geo-restricted, with no EU customers? → Likely out of scope, but verify sign-up and payment flows.

Exceptions to be aware of

Article 50 contains exceptions — for instance, where AI interaction is obvious to a reasonably well-informed person, for certain law enforcement uses, and for artistic, creative, satirical or fictional works where the deepfake disclosure can be limited so it does not hamper enjoyment. Read the article's text before relying on an exception.

For the wider AI Act picture, see EU AI Act for Indian companies.

Duty 1: AI interaction disclosure — how to implement

Text chat

  • Show a clear statement at the start of the conversation: "You're chatting with an AI assistant."
  • Keep a persistent indicator (label or avatar badge) in the chat UI.
  • If a human agent can take over, label the switch clearly in both directions.

Voice

  • Open calls with an audible disclosure before collecting any information.
  • Do not use voice personas that imply a human without disclosure.

Email and messaging agents

  • Sign AI-drafted messages sent autonomously with a disclosure line.
  • Where a human reviews and sends, decide whether the "obvious" test is met; if unsure, disclose.

API products

If you provide a model or system that EU developers embed, document how they can surface the disclosure, and consider making it the default in SDKs and widgets. You are the provider of the system as designed; design it so disclosure is the easy path.

Do not confuse with privacy notices

An AI disclosure banner does not satisfy GDPR transparency or India's DPDP Rule 3 notice. Keep both. See DPDP privacy notice requirements.

Duty 2: machine-readable marking — how to implement

Providers of generative systems must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The Act expects solutions to be effective, interoperable, robust and reliable as far as technically feasible.

By output type

OutputPractical techniques
ImagesEmbedded provenance metadata (C2PA-style content credentials) plus invisible watermark
VideoContainer-level metadata plus frame-level watermark
AudioAudio watermark plus file metadata
TextMetadata where the format allows; statistical watermarking where feasible; API response fields flagging AI generation

Engineering rules

  1. Mark server-side at generation. Do not rely on client code that API users can bypass.
  2. Preserve through your pipeline. Resizing, compression and format conversion must not strip marks; add CI tests.
  3. Provide detection. Publish how marks can be detected, or offer a verification endpoint.
  4. Version your marking. Log which marking method was applied to which output.
  5. Configure visible labels separately. Visible labels help deployers meet deepfake disclosure and help with India's IT Rules, but they are not a substitute for machine-readable marking.

If you also serve Indian users, India's IT Rules amendments on synthetic content (final text notified February 2026) add labelling duties; compare in AI labelling in India vs EU Article 50.

Duties 3–5: when you are the deployer

Startups become deployers when they use AI systems themselves in a professional context — for example, generating ad creatives, running an AI news digest, or using emotion analytics in a product.

Deepfakes

If you publish image, audio or video content that resembles real people, places or events and would falsely appear authentic, disclose that it is AI-generated or manipulated. For evidently creative or satirical works, the disclosure can be limited to not hamper the work, but it is still required.

AI-written public-interest text

If you publish AI-generated text to inform the public on matters of public interest, disclose it — unless a human has reviewed or edited it and a person or organisation holds editorial responsibility. Build an editorial workflow that records the reviewer.

Emotion recognition and biometric categorisation

If your product infers emotions or categorises people by biometric data, inform the people exposed. Note that some emotion recognition uses — for instance in workplaces and education — fall within the AI Act's prohibited practices, which have applied since 2 February 2025. Check the prohibition list before building.

Supporting your customers

Your enterprise customers are deployers too. Give them documentation, configurable labels and disclosure templates so they can meet duties 3–5 with your product.

Transparency under Article 50 sits beside, not instead of, data protection law.

  • GDPR applies to personal data of people in the EU in prompts, uploads and outputs. Your EU customers will expect an Article 28 DPA and SCCs for any access from India. See DPA and SCCs for Indian vendors.
  • DPDP applies to Indian users' personal data from 13 May 2027: Rule 3 notice, Rule 6 security, Rule 7 breach intimation and Rule 8(3) one-year log retention.
  • Marking logs. Records of which outputs were marked, and for which user, are personal data. Retain and protect them accordingly.

Common mistakes

  • Burying "AI-generated" in terms of service.
  • Marking images but not audio or text.
  • Stripping metadata in a CDN or image optimisation step.
  • Assuming geo-blocking the EU in marketing means EU users never reach the product.
  • Treating the Digital Omnibus proposal as having changed Article 50 timing without checking whether it was adopted and what it says.

Further reading on suhasbhairav.com

Before 13 May 2027

Compliance checklist

  1. Confirm EU exposure

    Check sign-ups, API keys and customer contracts for EU users or EU use of outputs.

    AI Act Art. 2

  2. Add AI interaction disclosure

    Show disclosure at the start of chat and voice sessions and keep a persistent indicator.

    AI Act Art. 50(1)

  3. Mark images and video

    Embed provenance metadata and robust watermarks server-side.

    AI Act Art. 50(2)

  4. Mark audio

    Apply audio watermarking and metadata to generated speech and sound.

    AI Act Art. 50(2)

  5. Flag generated text

    Add metadata or response fields marking AI text; evaluate text watermarking.

    AI Act Art. 50(2)

  6. Test mark persistence

    Add CI tests that marks survive resizing, compression and transcoding.

    AI Act Art. 50(2)

  7. Publish detection guidance

    Document how deployers and the public can detect your marks.

    AI Act Art. 50(2)

  8. Label your own deepfakes

    Disclose AI generation in marketing or media content you publish.

    AI Act Art. 50(4)

  9. Set an editorial workflow

    Record human review for AI-assisted public-interest text.

    AI Act Art. 50(4)

  10. Screen for prohibited emotion recognition

    Check emotion recognition features against the prohibited practices list.

    AI Act (prohibited practices)

  11. Follow the code of practice

    Track the Commission's marking and labelling code and align your methods.

    AI Act Art. 50

Key dates

  1. EU AI Act enters into force

  2. Prohibited practices, including certain emotion recognition uses, apply

  3. Article 50 transparency obligations apply

  4. DPDP core obligations apply to Indian users' data

Answers

Frequently asked questions

Does Article 50 apply to Indian AI startups?

Yes, if they place AI systems on the EU market or their system's output is used in the EU. Article 2 applies the AI Act regardless of where the provider is established. An Indian chatbot or image generator with EU users must meet provider duties under Article 50.

Do I need to watermark AI text under the EU AI Act?

Providers of systems generating synthetic text must mark outputs in a machine-readable format and make them detectable, as far as technically feasible. For text, that may mean metadata, API flags or statistical watermarking. The Commission's code of practice on marking and labelling gives practical guidance.

Is a disclaimer in my terms enough for Article 50?

No. People must be informed they are interacting with an AI system, normally at the latest at the first interaction, in a clear and distinguishable way. A clause in terms of service that users rarely read does not meet that standard. Use in-product disclosure.

What if my chatbot is obviously AI?

Article 50 does not require disclosure where it is obvious to a reasonably well-informed, observant and circumspect person that they are interacting with AI, taking context into account. Relying on this is risky for humanlike voice or text agents, so most providers disclose anyway.

What are the penalties for breaking Article 50?

Breaches of most AI Act obligations, including Article 50 transparency, can attract fines up to €15 million or 3% of worldwide annual turnover, whichever is higher, with lower ceilings for SMEs and startups in some cases. Enforcement is by national market surveillance authorities.

Has the Digital Omnibus changed Article 50?

The Commission proposed the Digital Omnibus in November 2025 to amend parts of EU digital law, including some AI Act timelines. It is not law until adopted and published. Check its status and final text on EUR-Lex before assuming any change to Article 50 or its dates.

Legal references and sources

This guide explains the law for engineering and business teams. It is not legal advice; confirm how it applies to your organisation with qualified counsel, and check the Gazette for later notifications or amendments.

About the author

Suhas Bhairav

Suhas Bhairav

Systems architect and applied AI practitioner who designs production AI systems for Indian and global companies, with a focus on privacy-by-design under the DPDP Act and the EU AI Act.