Article 50: what every SaaS must ship by 2 August 2026

Article 50 of Regulation (EU) 2024/1689 starts to apply on 2 August 2026. For most SaaS founders this is the first part of the AI Act that bites, because Article 50 does not care whether your product is high-risk or low-risk under Annex III. It cares about one question: inside your product, does a natural person interact with an AI system, or does your product produce AI-generated content? If the answer is yes for any feature, you are bound. The penalty for an Article 50 violation runs up to €15 million or 3% of worldwide annual turnover under Article 99(4), whichever is higher. This piece is the single-source readout the practice publishes for founders preparing for the 2 August deadline. It covers what Article 50 actually requires, the edge cases our auditors see founders miss, and the minimum you need to ship.

The four rules in Article 50

Article 50 is short. Most founders skim it once and miss two of the four rules. The full text is in EUR-Lex; the four operative paragraphs are these.

Article 50(1): chatbot disclosure. Providers of AI systems intended to interact directly with natural persons must design them so that the person is informed they are interacting with an AI system. The obligation falls on the provider (the entity that places the system on the market) and, in practice, on the deployer (the SaaS that integrates a GPAI through an API and exposes it to end users). Disclosure must be "clear and distinguishable" at the latest at the first interaction. The exemption for systems that are "obvious from the circumstances" is narrower than founders assume; our auditors do not rely on it.

Article 50(2): synthetic-content marking. Providers of generative AI systems must mark synthetic audio, image, video, and text output so that the content is detectable as artificially generated. The marking has to be machine-readable. This means provenance metadata: C2PA, IPTC, or an equivalent watermarking standard. A visible "made with AI" label is not by itself sufficient under Article 50(2); the obligation is technical, not cosmetic.

Article 50(3): emotion and biometric categorisation. Deployers of emotion-recognition systems or biometric-categorisation systems must inform the natural persons exposed to them. This rule attaches at deployment, not at training. A SaaS that runs sentiment analysis on a sales call is a deployer of an emotion-recognition system and owes disclosure to the call's participants.

Article 50(4): deepfake and AI-generated public-interest text. Deployers of AI that produces or manipulates image, audio, or video constituting a deepfake must disclose that the content has been artificially generated or manipulated. AI-generated text published to inform the public on matters of public interest carries the same obligation unless the content has undergone human editorial review and a person or organisation holds editorial responsibility.

Four rules. Two attach to providers, two to deployers. Most SaaS occupy both roles depending on the feature, which is why our intake form classifies feature-by-feature, not product-by-product.

Edge cases that decide whether you are bound

The plain reading of Article 50 catches obvious cases. Edge cases decide most of the engagements we run.

Function calling and tool use. A user types into a search box. Your backend routes the query through a GPAI to determine intent, then returns structured results. Is the user "interacting with an AI system"? Recital 132 reads narrowly (interaction means conversation), but the practice treats any user-facing surface that routes through an AI as in-scope when the AI's output materially shapes what the user sees. The safer position is to disclose.

Voice agents. Article 50(1) explicitly survives voice. A voice agent that takes a call and books a meeting is an interactive AI system. Disclosure must be conveyed in the modality of the interaction; a written disclosure in the user's account settings does not discharge the obligation.

Embedded widgets. SaaS A embeds SaaS B's chatbot as a widget. The user interacts with B's surface inside A's product. Both A (the deployer of B's system in A's context) and B (the provider) carry Article 50(1) obligations. The disclosure can be shared, but liability is joint.

B2B and internal tools. Article 50 does not carve out B2B or internal deployment. A natural person interacting with an AI system is a natural person whether they are a customer, an employee, or a partner. Internal HR chatbots, customer-support agents, and sales-enablement copilots are all in scope.

"Obvious from the circumstances." The Article 50(1) exemption is invoked by founders far more often than it works. Our auditors apply a strict test: the average reasonable person, with no special expertise, must immediately and unambiguously know the counterpart is an AI. A product called "AI Assistant" with an avatar that looks like a person does not pass. Tone and branding are insufficient.

Article 50(5) accessibility. Disclosure must be accessible to persons with disabilities. Visual-only disclosure on a voice agent fails. Audio-only disclosure on a deaf-accessible chat fails. The obligation is dual-modality where the interaction is dual-modality.

How the practice implements each rule in production

Disclosure is not a paragraph in your privacy policy. It is a UI surface, a metadata field, and a published statement. The practice uses the same pattern across engagements.

Chatbot disclosure (Article 50(1)). First-message preamble: "You are chatting with an AI assistant. It can make mistakes; verify anything important." Persistent indicator: "AI" badge near the input. Accessible label: aria-label on the conversation container. The exact copy is in our open template at /blog/article-50-disclosure-code.

Synthetic-content marking (Article 50(2)). For images: embed C2PA manifests at generation time via the upstream API (OpenAI Images, Stability, Adobe Firefly all support it). For text: store the generation metadata server-side and expose it via a Content-Provenance HTTP response header when the content is served as a standalone artefact. For audio and video: SynthID or an equivalent. The practice does not accept "the upstream provider handles this" as a discharge of Article 50(2) without verifying which marking standard the upstream actually emits.

Emotion and biometric notice (Article 50(3)). A short-form notice at the point of capture: "This call may be analysed for sentiment to improve service quality." Stored consent record where the legal basis requires it (GDPR overlap, not AI Act). The AI Act's Article 50(3) duty is informational; the GDPR's Article 9 duty for biometric data is consent-based. Implement both.

Deepfake and public-interest text (Article 50(4)). Visible label on the artefact ("AI-generated" or "AI-edited") plus the C2PA metadata from rule 2. For public-interest text, document the editorial review chain in a /provenance endpoint per piece of content. The editorial-responsibility exemption is real but only if a named person or organisation actually carries that responsibility on the record.

Every implementation choice above maps to a specific paragraph of Article 50 and a specific recital. The mapping is what an auditor or DPA will ask for first.

Penalties, enforcement, and what happens after 2 August

Article 99(4) sets the Article 50 penalty band at up to €15 million or 3% of worldwide annual turnover, whichever is higher. This is the same band as most high-risk obligations. It is lower than Article 5 (prohibited practices, €35M or 7%) and higher than misleading information (€7.5M or 1%). For SMEs and startups, Article 99(6) requires authorities to consider the proportionality of the fine; this is a moderating factor, not a safe harbour.

Enforcement runs through national market-surveillance authorities. Member States had until 2 August 2025 to designate them. Where a national authority has not been formally designated by 2 August 2026, the Commission has signalled it will not impede national consumer-protection or data-protection authorities from acting on Article 50 grounds in the interim. Pragmatically, the first wave of Article 50 enforcement will run through DPAs and consumer authorities that already have AI on their workplan.

What we expect in the first twelve months. Industry-led complaints will dominate volume; consumer-association complaints will dominate visibility. The first published decisions will almost certainly target chatbots without first-message disclosure and generative-image features without C2PA. Both are the easiest violations to demonstrate from a single screen recording. Our auditors prioritise these two surfaces above all others in the run-up to 2 August.

Three dates founders should track. 2 August 2026: Article 50 applies. 2 December 2027: Annex III high-risk obligations apply (delayed from 2 August 2026 by the AI Omnibus). 2 August 2028: Annex I embedded-product obligations apply. These are independent deadlines, and we run separate workstreams for each.

The minimum you ship by 2 August 2026

If you read nothing else in this piece, ship the following before 2 August 2026.

  1. First-message disclosure on every chatbot, voice agent, and conversational surface. Copy and aria-label per the template at /blog/article-50-disclosure-code.
  2. A persistent "AI" badge or label adjacent to the input field of any conversational surface.
  3. C2PA or equivalent provenance metadata on every image, audio, and video your product generates. If the upstream provider emits it, surface it; if not, embed it server-side before storage.
  4. A short-form notice at point of capture for any emotion-recognition, sentiment-analysis, or biometric-categorisation feature, regardless of how the analysis is later used.
  5. A visible "AI-generated" or "AI-edited" label on any output that could constitute a deepfake under Article 3(60), defined broadly enough that the practice treats all AI-edited human imagery as in scope.
  6. A published Article 50 statement at /article-50 or /transparency listing the AI systems your product uses, which Article 50 paragraphs apply to each, and how the obligation is discharged. This becomes your defence document if a complaint lands.
  7. A logging trail that captures, per session, which AI disclosure was shown, in which modality, and at what timestamp. This is not required by Article 50 but is required by every audit we run.

What you do not need to ship by 2 August: anything from Annex III (those obligations apply 2 December 2027), conformity assessment under Article 43, or any provider-side obligation if you are only a deployer. The pressure to over-comply by August is real, and it consumes engineering time you will need for Annex III.

The practice runs a five-day audit that covers the seven items above and produces a signed assurance letter. Details at /audit and /pricing.

Frequently asked questions

Does Article 50 apply if my chatbot's first message says "Hi, I'm an AI assistant"?

Partially. Article 50(1) requires disclosure that is clear, distinguishable, and made at the latest at the first interaction. "Hi, I'm an AI assistant" inside the first message satisfies the timing rule. It does not by itself discharge the accessibility duty under Article 50(5), the persistent-indicator practice our auditors expect, or the logging trail that lets you prove disclosure happened. Treat the first-message text as one of four requirements, not the complete fix.

I use OpenAI's API. Am I a provider or a deployer under Article 50?

You are almost certainly a deployer. The provider is the entity that develops the AI system and places it on the EU market; for general-purpose chat models that is OpenAI, Anthropic, Google, or Mistral. As a deployer, Article 50(1) still binds you for chatbot disclosure, and Article 50(3) and 50(4) bind you for emotion-recognition and deepfake use cases. Article 50(2) synthetic-content marking is a provider obligation, but you are responsible for surfacing the provenance metadata to your users; "the upstream emits it" is not by itself a defence if your product strips it.

What is the relationship between Article 50 and Article 52?

Article 52 covers procedural matters for the AI Office and harmonised standards; it does not impose disclosure on chatbot deployers. Founders sometimes confuse Article 52 with the original draft text or with another regulation. The operative chatbot rule is Article 50(1). If you read advice that points to Article 52 for chatbot disclosure, verify the source against the final Regulation (EU) 2024/1689 text on EUR-Lex.

Does this apply to internal tools used only by employees?

Yes. Article 50 applies wherever a natural person interacts with an AI system or sees AI-generated content. There is no carve-out for internal or B2B use. The practical implication: an internal HR chatbot, a sales copilot, or an engineering assistant that takes natural-language input from employees is in scope. Employer-employee context does not change the rule, though it may change which legal basis you rely on for any overlapping GDPR duties.

My SaaS has no chatbot, but it generates AI summaries of user data. Am I still bound?

Probably. Article 50(2) marking applies to synthetic content; an AI-generated summary is synthetic text. If the summary is consumed only by the user who triggered it (private context), the duty under 50(4) for public-interest text does not attach, but the provenance-marking duty does where you are the provider of the generative system. If the summary is shared, published, or shown to a third party, treat it as content that needs a visible "AI-generated" label plus machine-readable provenance.

Sources

Last updated: 2026-06-09