Skip to content

Since August 2, 2026, chatbots used in the EU must say they are AI and generated content must carry a machine-readable mark. For products launched earlier, the grace period on marking ends December 2. Who is covered, what to change and how your own image pipeline can strip the marks.

Published
Reading
13 min
Author
Yaroslav Tsarenko
–views
Glowing shapes on a dark field, each wrapped in a thin marking ring, with a timeline and a bright point ahead

On August 2, 2026, Article 50 of the EU AI Act started to apply: the transparency rules for AI. A chatbot has to say it isn't human. Generated images, video, audio and text have to carry a machine-readable mark. And on December 2, four weeks from now, the grace period ends for products that were already on the market before August: from that date, marking generated content is mandatory for them too (European Commission FAQ).

This article is for founders and product teams with AI features and users in the EU: a support assistant, a product description generator, an AI avatar, a tool that produces images or voice. We'll cover what Article 50 actually requires, who is responsible for what, why a product built on the OpenAI API can still be out of line, and what to do in four weeks. This is an engineer's reading, not legal advice: bring in counsel for final calls. My part is showing which obligations turn into tickets.

What applies since August 2, and what starts December 2

The AI Act (Regulation (EU) 2024/1689) is being phased in since 2025. In summer 2026 the EU adopted the Digital Omnibus on AI, Regulation (EU) 2026/1744. It pushed back the heaviest high-risk requirements but left Article 50 on schedule. The Omnibus was published in the Official Journal on July 24 and entered into force on July 27 (Cooley summary).

Feb 2, 2025ProhibitedpracticesAug 2, 2025General-purposeAI modelsAug 2, 2026Article 50:chatbots, deepfakesDec 2, 2026Marking forexisting systemsDec 2, 2027High-risk,Annex IIIAug 2, 2028High-risk,Annex INovember 4: you are here

What matters to a product team:

  • August 2, 2026. Article 50 applies to every new AI system. Chatbot disclosure and deepfake labelling are already mandatory, with no grace period at all.
  • December 2, 2026. Generative systems placed on the market before August 2 must start marking their output (Article 50(2)). The same day two new prohibitions kick in: generating intimate imagery of identifiable people without consent, and generating child sexual abuse material.
  • December 2, 2027. Requirements for Annex III high-risk systems (hiring, education, credit scoring and so on). If your product isn't in that list, Article 50 is the part of the AI Act that affects you next.

Content generated and published before August 2, 2026 doesn't need to be marked retroactively, though the Commission encourages it.

Does this apply to your product

The two most common objections are "we're not an EU company" and "we just call the OpenAI API". Neither works the way people hope.

The AI Act reaches providers outside the EU when the output of their AI system is used in the EU. A US startup with customers in Berlin or Paris is in scope just like a European one. And "provider" in the AI Act doesn't only mean OpenAI or Google. A provider is whoever develops an AI system and places it on the market under their own name. The model can belong to someone else, but the chatbot on your site is your system.

1Is the AI outputused in the EU?Article 50does not applynoyes2Does the AI talk toa person directly?"I am an AI" disclosurefrom the first messageyesnext3Does it generate text,images, audio or video?Machine-readablemark on the outputyesnext4Do you publish a deepfakeor public-interest text?A visible labelfor peopleyes

Walk the tree above for each AI feature separately. Typical answers:

Feature in your product What Article 50 requires Who is responsible
Support assistant in a chat widget Say it's AI from the first message You, as provider of the system
AI agent that emails customers or books appointments Disclosure, because the agent talks to a person directly You
Image or video generator for your users A machine-readable mark on every file You, even if the model is third-party
Generated product descriptions in your catalog Marking on the text; for public-interest publications, disclosure too You, as provider and deployer
Internal tool: call summary for a sales rep Usually nothing, if there's no dialogue with outside people and nothing is published —
Background moderation or classification Outside Article 50: no dialogue and no generated content for people —

The last two rows answer the "do we have to rebuild everything" question. No. Article 50 is about the moments when AI talks to a person or creates content someone will see.

Provider or deployer: two roles, different duties

Article 50 splits duties between two roles. The provider builds the system and is responsible for how it behaves out of the box: disclosure in conversation and marking of output. The deployer uses a system in its own business and is responsible for what it publishes: labels on deepfakes and public-interest text, notices for emotion recognition.

Provider of the system

  • Chatbot says it is AIfrom the first message, clear and visible
  • Content is markedmetadata and a watermark on every output
  • Detection is possiblea way to check that content is generated

Deployer

  • Deepfakes are labelledvisible to people, not only to machines
  • Public-interest texta label, or human editorial review
  • Emotion recognitionpeople are told the system is running
Provider vs deployer: who owns what

A typical SaaS with AI features plays both roles at once. You're the provider of the banner generator you give your customers. You're also a deployer when your marketing team publishes a generated video ad. One detail people miss: a provider's machine-readable mark doesn't satisfy the deployer's duty to show a visible label to people. They're separate obligations and are checked separately.

Chatbots and AI agents: "I'm an AI" from the first message

Article 50(1) requires systems that interact with people to be designed so the person knows it's AI. In its guidelines of July 20, 2026 and the FAQ, the Commission makes clear this covers not just chatbots but also AI agents and avatars. The test has four conditions that must all hold:

  1. it's an AI system;
  2. there's a genuine two-way exchange, not a one-off answer in the background;
  3. the AI communicates directly, not through a human intermediary;
  4. the counterpart is a natural person.

An agent that emails or messages a customer on your behalf and keeps the conversation going meets all four. An agent exchanging data with another service with no person involved doesn't.

There is an exception: no disclosure is needed when it's obvious to a reasonably well-informed, observant person that they're dealing with AI. But the Commission stresses that this exception is read narrowly. An "AI" badge on a button may not be enough if the bot introduces itself in the conversation as "Emma from support".

Article 50(5) adds requirements on form: the information must be clear and distinguishable, appear no later than the first interaction, and meet accessibility requirements. In practice:

  • the bot's first message or a line above the input: "You're chatting with an AI assistant";
  • for voice, a sentence at the start of the call, not a line in the privacy policy;
  • for emails sent by an agent, a plain statement in the body, not only in technical headers;
  • text a screen reader can announce, not just an icon.

When I built the AI assistant for the AllShip logistics marketplace, the core rule was that it never acts on its own: it proposes an action and a person confirms. Disclosure follows the same logic: the person should know who they're talking to and who makes the decision. How that assistant is put together is in my post on the OpenAI API in production.

Marking content: metadata plus a watermark

Article 50(2) requires generated audio, images, video and text to be marked in a machine-readable format and detectable as AI-generated. Solutions must be effective, interoperable, robust and reliable as far as technically feasible. Features that only assist with standard editing and don't substantially alter the input are exempt: a spellchecker or auto color correction isn't generation.

The law doesn't say how to mark. That's what the voluntary Code of Practice is for. The Commission published it on June 10, 2026; in July the Commission and the AI Board assessed it as adequate, and by July 31 about 190 organizations had signed it, including OpenAI, Google, Anthropic, Microsoft, Meta and Mistral (Commission announcement). You don't have to sign it, but it's the clearest available answer to "how much is enough".

Images, video, audioTextC2PA metadatasignature and timerequirednot possibleWatermarkinside the contentrequiredone layer is enoughFingerprint, logorigin registryoptionaloptionalVisible labeldeployer dutyfor deepfakespublic interest

The Code's core idea is at least two layers:

  • Metadata: a digital signature and timestamp embedded in the file in a tamper-evident way. The de facto standard is C2PA Content Credentials.
  • A watermark: an invisible signal inside the content itself that survives some transformations, such as a screenshot.
  • Fingerprinting or logging: an optional third layer, a registry that lets you look up where a file came from.

Text gets an exception: you can't attach metadata to plain text, so a single layer, a watermark, is accepted. The Code also asks providers to offer a detection tool and to prohibit intentional removal of metadata in their terms of use.

The good news for most products: if you generate images through a major provider, part of the work is done. OpenAI's content provenance docs say images from its API carry C2PA metadata and a SynthID watermark. The bad news: your own code can easily lose them.

The trap: your pipeline strips the marking

The same OpenAI docs warn that editing, converting or sharing a file can remove its metadata. And that is exactly what a typical product does to an image: resize it, convert it to WebP, store it, serve it through a CDN with image optimization.

The usual wayGenerationC2PA +SynthIDResize, WebPmetadatastrippedStoragefile withouta markCDNone morepassUserno markThe right wayGenerationC2PA +SynthIDkeepMetadataor re-signTest in CIfails withouta markCDNverifiedUsermark present

You can check your pipeline in five minutes. The script below creates a JPEG carrying the IPTC trainedAlgorithmicMedia XMP tag (the standard "made by AI" label), then runs it through sharp, the image library under the hood of a lot of Node.js apps, twice: the usual way and with keepMetadata().

ts
import sharp from "sharp";

const xmp =
  '<x:xmpmeta xmlns:x="adobe:ns:meta/"><rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#">' +
  '<rdf:Description xmlns:Iptc4xmpExt="http://iptc.org/std/Iptc4xmpExt/2008-02-29/" ' +
  'Iptc4xmpExt:DigitalSourceType="http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia"/>' +
  "</rdf:RDF></x:xmpmeta>";

const original = await sharp({ create: { width: 64, height: 64, channels: 3, background: "#74ecae" } })
  .withXmp(xmp)
  .jpeg()
  .toBuffer();

const plain = await sharp(original).resize(32).webp().toBuffer();
const kept = await sharp(original).resize(32).webp().keepMetadata().toBuffer();

for (const [name, buf] of [["original", original], ["plain", plain], ["kept", kept]] as const) {
  const meta = await sharp(buf).metadata();
  console.log(name, meta.format, "xmp:", Boolean(meta.xmp), "ai-label:", buf.includes("trainedAlgorithmicMedia"));
}

Output on sharp 0.34.5:

text
original jpeg xmp: true ai-label: true
plain webp xmp: false ai-label: false
kept webp xmp: true ai-label: true

A plain resize().webp() silently dropped the label: by default sharp removes all metadata. Image optimization services in your CDN can behave the same way, so test the file your user actually receives. C2PA is stricter still: the manifest is bound to a hash of the file, so after any re-encode the old signature no longer matches. The right approach is to write a new manifest after processing, with the original recorded as its ingredient, rather than hoping the metadata survives.

What to do about it:

  • list every place a generated file is changed: resize, format conversion, cropping, adding a logo;
  • where possible, keep metadata (keepMetadata() or the equivalent in your library) and guard it with an automated test;
  • for C2PA, re-sign after processing, or serve the original with its manifest next to the optimized copy;
  • confirm the provider's watermark survives your processing: the same OpenAI docs point to a web checker at openai.com/verify, so run the file your user actually gets through it.

Deepfakes and public-interest text: a visible label

Article 50(4) is the deployer's duty, meaning whoever publishes. If you show people a generated or manipulated image, video or audio that resembles real people, places or events (a deepfake), you must disclose that it's AI-generated. For evidently artistic, satirical or fictional work, it's enough to disclose that the content exists in a way that doesn't spoil the work.

The text rule is narrower: a label is needed when generated text is published to inform the public on matters of public interest and no person with editorial responsibility reviewed it. The Commission clarifies that a spellcheck doesn't count as editorial review. A sneaker description in your catalog usually isn't a public-interest matter. Automated fuel-price news on your media site is.

For visible labels, the Code offers three user-tested EU icons: one for fully AI-generated content, one for AI-modified content, and a basic one for an interactive layer. Using the ready-made icon is easier than inventing your own and having to prove people understand it.

Fines, enforcement and what not to panic about

Breaking Article 50 can cost up to 15 million euros or 3% of total worldwide annual turnover, whichever is higher. The Commission notes that proportionality can be taken into account for SMEs. National market surveillance authorities enforce it; the AI Office has a limited role.

A realistic read for a small product, and this is my estimate, not an official position: early enforcement will most likely start with complaints and high-profile cases rather than mass audits of startups. Still, two reasons not to wait:

  • enterprise customers will ask. A European company that is itself on the hook as a deployer will check that your product doesn't create problems for it, and "we haven't looked yet" is a weak answer in a procurement review;
  • it's cheap now and expensive later. Chat disclosure is a few hours of work. Marking in the pipeline is a few days. Rework after a complaint, with lawyers and a deadline, is a different story.

What to ask your model provider and your contractor

If a contractor built your AI features, or you assembled them in Lovable, Bolt or Cursor, here's a short list of questions:

  • Which models and providers are in use? Did they sign Section 1 of the Code, and what exactly do they mark: metadata, a watermark, both?
  • What happens to a file after generation? Where are the resize, conversion and CDN steps, and does the marking survive to the user?
  • Where does the product talk to people? Chat, voice, email, agents, and where is the disclosure in each.
  • Can we switch models without a rewrite? Providers retire models, and every swap means checking the marking again. How to avoid the retirement trap, I covered using the GPT-4o shutdown as the example.
  • Where do the API keys live? If your generation key leaks, someone can generate content in your name. What to do about keys is in my breakdown of the LiteLLM and Railway leaks.

A four-week plan to December 2

Four weeks is enough if you start with an inventory rather than legal memos.

  1. week 1
    Inventorywhere the product talks to people and what it generates
  2. week 1
    Disclosure in the UIchatbots, agents, voice assistants, emails sent by AI
  3. weeks 2–3
    Markingyour model provider's metadata and watermarks survive the pipeline
  4. week 3
    Labels for peoplean icon on deepfakes and generated publications
  5. week 4
    Evidencetests, a decision log, one transparency page
Four weeks to December 2
  • Week 1: inventory and disclosure. List every AI feature: where AI talks to people and what it generates. For each, the role (provider or deployer) and the duties from the table above. Fix chat and agent disclosure this week: it has applied since August 2.
  • Weeks 2–3: marking. Check that your model provider's metadata and watermarks reach the user. Add an automated test that fails when the pipeline strips the mark. If the model is your own or open-weight, pick a marking solution using the Code as the benchmark.
  • Week 3: visible labels. EU icons on deepfakes and generated publications, and an editorial review rule for public-interest text.
  • Week 4: evidence. A short AI transparency page for the product, a decision log (why feature X is outside Article 50), tests in CI. That's what you show a customer, an investor or a regulator.

Checklist: audit your product in 20 minutes

Check How Red flag
Chat disclosure Open the chat as a new user The first message doesn't say it's AI
Bot name See how the bot introduces itself A human name with no explanation
Voice and email Listen to the start of a call, read an agent's email Disclosure only in the privacy policy
Image metadata Download a generated file the way a user gets it and check it with openai.com/verify or a C2PA tool No mark
Optimization Search the code for resize, toFormat and your CDN image settings Metadata stripped by default
Deepfakes Review ads and marketing assets Realistic people with no label
Publications Find automated text on the site News or analysis with no human review
Documentation Look for a customer-facing description of AI features Nothing beyond "we use AI"

Two or more red flags mean it's worth booking a week of work before December 2.

What to do today

Article 50 isn't a ban on AI. It's about honesty: people should know when they're talking to a machine, and content should carry a trace of where it came from. For most products this isn't a rebuild but a handful of targeted changes: a line in the chat, a test on metadata, an icon on generated publications, and a page that explains it all.

Start with the checklist: 20 minutes and you'll know the size of the job. If you have many AI features, or someone else built them, the fastest route is to run the inventory together and turn it into a four-week plan. How I work on AI features in products is on the AI integration page.

Was this useful?

– likes · – views

Need this for your product?AI features grounded in your data, not a chat widget glued onLearn more →

Questions people ask

What does Article 50 of the EU AI Act require?

Four things. Chatbots and AI agents must tell people they are interacting with AI. Systems that generate text, images, audio or video must mark their output in a machine-readable way. Deployers of emotion recognition must inform the people exposed to it. And anyone publishing deepfakes or AI-generated text on matters of public interest must disclose it.

What happens on December 2, 2026?

The Article 50(2) grace period ends for generative AI systems placed on the market before August 2, 2026: from then on their output must carry machine-readable marking. The grace period was added by the Digital Omnibus, Regulation (EU) 2026/1744. Every other Article 50 duty, including chatbot disclosure, has applied since August 2 with no grace period.

Does the EU AI Act apply to US or other non-EU companies?

Yes, if the output of your AI system is used in the EU. The regulation covers providers established outside the Union, so a US SaaS with customers in Germany or France falls under Article 50 just like a European one.

What are the fines for breaking Article 50?

Up to 15 million euros or 3% of total worldwide annual turnover for the preceding year, whichever is higher. The Commission notes that proportionality can be taken into account for SMEs. National market surveillance authorities enforce it.

Do I have to sign the Code of Practice?

No. The Code is voluntary and not signing it is not a breach. But the Commission and the AI Board assessed it as adequate, so following its measures is the simplest way to show your marking and labels meet the law.