Article 50 kept its date, and it lands on the chat window

The week

For a long time, August 2, 2026 had been circled on compliance calendars across Europe and beyond. It was the date on which most of the EU AI Act, the European Union’s law governing artificial intelligence, was due to apply. Then, weeks before, the EU changed the timetable.

The EU finished the AI part of its digital omnibus, a package of amendments meant to simplify and adjust several digital laws at once. Regulation (EU) 2026/1744 was published in the Official Journal, the EU’s official record of law, on July 24 and entered into force on July 27. It moved the obligations for standalone high-risk AI systems, the category covering uses such as credit scoring and certain decisions about people, to December 2, 2027. It moved obligations for AI built into regulated products, such as medical devices, to August 2, 2028.

Many readers took that as a general reprieve, but the omnibus left the transparency duties in Article 50 on their original date. From August 2, providers of AI systems that interact with people had to make clear that users are dealing with AI. Deployers, the organizations that put AI to use, had to disclose deepfakes and certain AI-generated text. So if a customer is chatting with a bot, the customer has to be told; if a video shows a realistic person who was generated by software, viewers have to be told.

There was only one piece of relief. Machine-readable marking of synthetic content, a hidden digital tag that lets software recognize AI-made images, audio, video or text, has until December 2, 2026 for systems already on the market. Everything else in Article 50 applied on schedule. On the same date, the European Commission’s power to fine providers of general-purpose AI models, the large models that power many chatbots and tools, also took effect.

So in the first week of August, organizations serving people in the EU woke up to a split reality. The heaviest obligations had moved more than a year away. The most visible ones, the ones a customer can see by opening a chat window, were already in force.

The omnibus confirmed that the Act’s high-risk regime would wait until late 2027. It did not touch the part of the Act that customers and citizens actually see.

Article 50 is short, and it is easy to dismiss as a labeling rule. In practice it reaches every chat assistant, voice bot, generated image and drafted notice that a regulated organization puts in front of people in the EU. Breaches sit in a penalty tier of up to €15 million or 3 percent of worldwide turnover. For a large bank or insurer, the turnover-based figure is the one that sets the ceiling.

How the AI Act sorts risk, and where Article 50 fits

The EU AI Act sorts AI by risk. A few uses are banned outright. A larger group of high-risk uses, such as AI that helps decide who gets credit or a job, carry heavy obligations: risk management, documentation, human oversight and more. Those are the obligations the omnibus pushed back. Then there is a lighter, broad category of transparency duties, which apply to many ordinary systems simply because people might not realize AI is involved. That is Article 50.

Two words do a lot of work in the Act. A provider is the organization that develops an AI system, or has it developed, and puts it on the market or into service under its own name. A deployer is the organization that uses an AI system in its work. A useful comparison is a car: the manufacturer is the provider, the delivery company that runs a fleet is the deployer. Each has different duties, and the same company can be both for different systems.

A few other terms recur. A general-purpose AI model is a large model trained on broad data that can be adapted to many tasks; many chat assistants are built on one. Synthetic content is anything generated or substantially altered by AI: an image, a voice clip, a video, a paragraph. A deepfake, in the Act’s sense, is generated or manipulated image, audio or video content that resembles real people, places or events and could falsely appear authentic.

Article 50 sets three distinct duties, and each lands on a different team. Providers must design interactive systems so people know they are dealing with AI. Providers of generative systems must mark synthetic output in a machine-readable way. Deployers must disclose certain uses and outputs to the people affected. Each needs its own owner and its own evidence. Organizations should confirm with counsel which apply to them, especially where they serve EU customers from outside the EU.

Duty one: telling people they are dealing with AI

The first duty falls on providers of systems intended to interact directly with people. Unless it is obvious from context, the person must be told no later than the first interaction. Not after a few exchanges, not in the terms and conditions: at the start.

For a bank or insurer, the first question is who the provider is. A firm that buys a vendor’s assistant and deploys it unchanged is usually a deployer. A firm that builds its own assistant on a general-purpose model and puts it into service under its own name is likely the provider of that system. Many digital teams have done exactly that, often quickly and with enthusiasm, without anyone stopping to ask what role that made the firm play under the Act.

The duty then becomes a design requirement with test cases. The disclosure has to appear at the first turn in every channel: web chat, app, messaging platforms and voice. Voice is where gaps tend to hide. An intelligent call routing system that answers in natural speech is an AI interaction, and the greeting script has to say so. A caller who hears a friendly, fluent voice asking how it can help may reasonably assume it is a person.

A hand-off rule matters too. When a human agent takes over, or an AI drafts replies a human sends, the customer should not be misled about which is which. Product owners should add these checks to release testing, the set of checks run before any new version goes live, so a redesign cannot drop the notice. A disclosure that existed in March and vanished in a September redesign is exactly the kind of gap a regulator, or a journalist, will find.

Suppose an insurer launched a claims assistant on its website last year, built in-house on a general-purpose model. The web version opens with a clear line explaining that the customer is talking to an AI assistant. But the same assistant was later connected to a popular messaging app, and that version opens straight into “Hi, how can I help with your claim?” The insurer is the provider, the messaging channel is in scope, and the gap is visible to any customer who tries it.

Duty two: marking what the system generates

The second duty applies to providers of systems that generate synthetic audio, images, video or text. Their outputs must be marked in a machine-readable format and be detectable as AI-generated, where technically feasible. Machine-readable marking is a watermark or a label in the file’s metadata, the hidden information that travels with a digital file, describing where it came from. People may not see it, but software can check it.

Systems already on the market before August 2 have until December 2, 2026 to comply. In July the Commission and the AI Board, which brings together representatives of EU member states, judged a voluntary code of practice on marking and labeling adequate, which gives providers a recognized route.

Most regulated organizations will meet this duty through their vendors. The procurement task is to confirm that image, voice and text generation tools used in marketing, customer communications and claims handling apply marking, and that internal pipelines do not strip it. Re-encoding a video, flattening an image or pasting generated text into a document template can remove metadata.

A simple control is to sample published assets each quarter and check that marking survived. Where an organization fine-tunes or wraps a model and offers the result to customers, it may be a provider in its own right and should plan accordingly. A bank that offers business clients a branded tool for drafting marketing images, built on someone else’s model, needs to ask that question carefully.

Duty three: disclosing what deployers use and publish

The third set of duties falls on deployers, which covers almost every bank, insurer, hospital and agency using AI. Three uses matter.

Emotion recognition and biometric categorization. People exposed to these must be informed. Emotion recognition means software that infers how someone feels from their voice, face or words; biometric categorization means sorting people into groups based on physical or behavioral traits. Contact-center analytics that infer caller sentiment or emotion are the usual case. Emotion recognition in the workplace has been prohibited since February 2025, so the same tools pointed at staff need separate review. A vendor’s dashboard that scores both the customer’s mood and the agent’s mood may be fine on one side and prohibited on the other.

Deepfakes. Generated or manipulated images, audio or video resembling real people, places or events must be disclosed. A synthetic spokesperson in a bank video or a cloned executive voice on a hold message both qualify. Marketing teams may not think of these as deepfakes, a word associated with deception, but the duty applies all the same.

AI-generated public-interest text. AI-generated text published to inform the public on matters of public interest must also be disclosed. The exception is text that has undergone human review or editorial control, with someone holding editorial responsibility. That exception suits agencies and health systems well, provided the review is real and recorded. A public notice on benefit changes, or a hospital’s public health update, drafted with AI and approved by a named communications lead, can rely on it. The control is a sign-off record that names the reviewer, not a general policy statement.

How it works in practice: one inventory, three owners

Abstract duties become manageable once they are attached to real systems. Take a European health system with a dozen hospitals, a patient app and a busy communications office. Its compliance lead starts with a single spreadsheet listing every place AI touches a patient, a member of the public or a member of staff.

The list is longer than anyone expected. There is an appointment-booking assistant in the patient app, built by the digital team on a general-purpose model. There is a vendor’s voice system on the main switchboard that answers calls in natural speech and routes them. There is a contact-center analytics tool that scores calls for sentiment. The communications office uses a text generator to draft public health updates and an image tool for campaign posters. A recruitment team has been piloting a video interview product that claims to assess candidates’ engagement.

Each line gets a role and an owner. The booking assistant: the health system is the provider, so the digital product lead owns the first-interaction disclosure and adds it to release testing in every channel. The switchboard: the vendor is the provider of the system, but the health system checks that the greeting says the caller is speaking with an automated assistant. The analytics tool: the health system is a deployer, so callers are informed, and the staff-scoring feature is reviewed against the workplace prohibition. The text generator and image tool: procurement confirms with the vendors how marking is applied, and the communications office records a named reviewer for each AI-drafted public update. The recruitment pilot is paused for legal review.

The exercise requires no new technology, only a list, a decision for each line, a named person and a record. The work is therefore cheap to do early and awkward to reconstruct after a complaint.

A note for organizations outside the EU

Many US banks, insurers, health systems and public bodies assume European AI law is someone else’s concern. That assumption deserves a second look. The AI Act can reach organizations based outside the EU when their AI systems are placed on the EU market or their outputs are used in the EU. A US bank with European retail customers, an insurer with an EU branch, or a hospital network that serves international patients through a shared app may find that Article 50 applies to some of its channels.

The practical test is not where the organization is headquartered but where its users are. Counsel should map which products and channels reach people in the EU, and the inventory above should flag them. Where a single global chat assistant serves customers everywhere, the simplest answer is often to apply the disclosure everywhere, which also avoids the awkwardness of telling European customers something American customers are not told.

The part of the Act anyone can test

Article 50 is the part of the AI Act most visible from outside. Anyone can open a chat window, call a contact center or watch a promotional video. Customers, consumer groups and journalists do not need access to internal documents to test whether an organization discloses its use of AI. That makes reputational risk arrive faster here than in the high-risk regime, where failures are often discovered only through audits.

It is also a matter of trust. Customers who discover they were talking to software without being told may reasonably feel misled. A bank or hospital that is open about where it uses AI, and clear about when a human is involved, builds credibility. One that is caught out loses it.

And there is a cost argument. Each of these duties is cheap to meet early and costly to retrofit after a complaint. Interface notices, marking checks and editorial sign-off logs are ordinary product and communications controls. The difference after August 2 is that a regulator can ask to see them.

The shape of a working response

A well-run response has a recognizable shape:

• An inventory of every customer-facing AI system, with a decision recorded on whether the organization is its provider, its deployer or both.
• A named owner for each of the three duties: typically digital product for interaction disclosure, procurement and marketing operations for marking, and communications for deployer disclosures.
• Disclosure checks built into release testing for every channel, including voice.
• Quarterly sampling of published assets to confirm marking survives internal tools.
• A sign-off log for AI-drafted public notices that names the reviewer.
• A review of contact-center analytics, separating customer-facing uses from any use on staff.

Most failures follow from misreading the omnibus. The first is treating the high-risk delay as a general pause and parking all AI Act work until 2027. The second is assuming that buying a vendor’s tool makes the organization a deployer in every case, when building on a general-purpose model can make it a provider. The third is covering web chat and forgetting voice and messaging. The fourth is relying on a policy that says “we review AI content” without any record showing who reviewed what. The fifth is assuming that because the organization is based outside the EU, none of this applies, when it may serve EU customers directly.

Where the disclosure gaps are likely to be

1. Which customer-facing systems use AI, and for each, are we the provider, the deployer or both?
2. Does every channel, web, app, messaging and voice, tell people at the first interaction that they are dealing with AI?
3. Do our vendors’ image, voice and text tools apply machine-readable marking, and have we checked that our own pipelines keep it intact?
4. Do our contact-center tools infer emotion or sentiment, from customers or from staff, and have we reviewed both uses?
5. When we publish AI-drafted notices on matters of public interest, who signs them off, and where is that recorded?

Behind all of these sits a scoping exercise for counsel: mapping which of the organization’s activities reach people in the EU, and confirming which Article 50 duties apply to each.

With hindsight

The deferral became settled law, and the next dates are now close. December 2, 2026 brings the marking deadline for generative systems already on the market, along with two new prohibitions added by the omnibus, on non-consensual intimate imagery and child sexual abuse material.

The Commission adopted its final Article 50 guidelines on July 20, after consulting on a draft in May and June, so providers and deployers already have the reference text regulators will read. That removes one common excuse for delay. Organizations can no longer say they are waiting to see how the duties will be interpreted; the interpretation has been published.

We have not seen public enforcement under Article 50, and national authorities’ priorities are not yet clear. That silence is better read as uncertainty than as safety, and the uncertainty argues for evidence rather than waiting. When an authority does act, or a complaint arrives, the organizations in the strongest position will be those that can show a record: the inventory of systems, the provider-or-deployer decision for each, the release tests, the marking samples, the editorial sign-offs.

The duties that arrived in August are visible to any customer or journalist who opens a chat window, so gaps are easy to find from outside. That is the particular character of Article 50. Unlike many compliance obligations, its failures do not wait for an auditor; they can be screenshotted.

Organizations that split the three duties by owner can show what they do. Those that read the high-risk delay as a general pause have catching up to do before the December dates. The December deadline for marking also falls on vendors, so procurement teams should already have written confirmation of how each generation tool will comply. The practical step now is to ask to see the inventory and the owners, and to ask when each channel was last tested.