Transparency
AI Act Article 50 for bank chatbots: what applies since 2 August 2026
Since 2 August 2026, Article 50 of the EU AI Act applies to every assistant a bank puts in front of customers: the website chatbot, the app assistant, the voice line that answers before an adviser. The question fits in one line: does the person writing to your assistant know they are talking to a machine, and do they still know after several exchanges?
This guide applies the article's five paragraphs to a bank, says who carries each one, and gives six tests to run on the live system. The Article 50 guide covers the article itself in full.
What Article 50 asks of a bank chatbot
Article 50 does not make your assistant high-risk. Its transparency duties apply whatever the risk class, and they apply now. The first paragraph speaks directly to a chatbot:
Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use. (Article 50(1))
The information must be built into the product, not left in a legal notice. The "obvious" exception is judged in context. A customer who opens a chat window in their online banking, where human advisers also answer, has no reason to guess they are talking to an AI, so do not build your compliance on that exception.
An assistant that assesses the creditworthiness of a person also falls under Annex III, point 5(b), with Chapter III from 2 December 2027; the Annex III guide draws that line. Article 50 already applies.
Who carries each duty in a bank
In the usual case, the bank buys the assistant from a vendor. The vendor is the provider and carries paragraphs 1 and 2; the bank is the deployer and carries paragraphs 3 and 4. Paragraph 5 sets the form for both, and the bank controls the screen: the assistant's name, its avatar, the opening banner and the system prompt, the standing instructions the assistant follows.
A bank that has the assistant built and puts it into service under its own brand fits the definition of provider (Article 3). Treat the duties you can influence as yours, and check the contract.
The five paragraphs, applied to a bank assistant
Paragraph 1, say it is an AI. The assistant must present itself as an AI, and keep doing so when asked. A human first name and a photographic avatar point the wrong way. The decisive test is not the opening banner but the answer to "are you a robot?" in the middle of a conversation.
Paragraph 2, mark what it generates.
Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated. (Article 50(2))
Your assistant's replies are generated text. Marking is the provider's duty, but you will have to show the supervisor how it works on your channel, so ask the vendor for its technique. The exception for standard editing does not by itself cover an assistant that writes its own replies; have counsel confirm.
Paragraph 3, emotions and biometrics.
Deployers of an emotion recognition system or a biometric categorisation system shall inform the natural persons exposed thereto of the operation of the system ... (Article 50(3))
A text chatbot is not concerned: the definition in Article 3, point 39, covers emotions inferred from biometric data. A voice line that infers stress in a customer's voice may be. If your vendor turns that feature on to route calls, the duty to inform is yours.
Paragraph 4, deep fakes and public-interest text.
Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated. (Article 50(4))
The assistant itself rarely triggers it. The paragraph matters when the bank reuses the same engine for a video with a synthetic adviser, or for a market note published without human review.
Paragraph 5, the form of the information.
The information referred to in paragraphs 1 to 4 shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure. The information shall conform to the applicable accessibility requirements. (Article 50(5))
"Clear and distinguishable" rules out a greyed-out line in the terms and conditions. "At the latest at the time of the first interaction" means before the assistant's first reply. The information must also reach a screen reader and read well on a phone.
What changes on 2 December 2026
Systems placed on the market before 2 August 2026 have until 2 December 2026 for the machine-readable marking of paragraph 2. If your vendor placed the chatbot on the market before 2 August 2026, that is the date to put in your plan; the 2 December 2026 guide covers the deadline. Article 111(4) speaks of systems placed on the market: if the bank built the assistant and put it into service itself, ask counsel whether the grace applies.
Two texts frame how to comply: the Code of Practice on Transparency of AI-Generated Content (10 June 2026) and the Commission Guidelines on Article 50 (20 July 2026). The Code of Practice is voluntary. A provider that does not follow it must show another way the marking duty is met. Ask your vendor whether it follows the Code. Two new Article 5 prohibitions apply on the same day; they do not target a bank assistant.
Six tests to run on the live assistant
A banner on the first message proves nothing about how the model behaves after several exchanges. Test transparency the way you test a safeguard: by trying to break it, with a pass criterion written in advance.
What to send, and what counts as a pass
Send each test ten times, because a model can answer the same question differently each time: one reply proves nothing, and the finding is a pass rate. Check the form too, before the first message, on a computer, on a phone and with a screen reader, in every language you offer. How we test sets out the method.
What to keep as evidence
Keep six things: the role allocation; the vendor's documentation; the real configuration, with its versions; dated screenshots per channel and language; the test transcripts, with the decision of the reviewer (the person who decides); and a record of every change, with the test run that followed it. Store it all in an audit trail that can be neither changed nor deleted; the Article 50 guide lists each item.
Our banks page shows how Vidimus covers a bank's agents, and our Article 50 page how it tests disclosure on a live agent.
To learn which paragraphs apply to your assistant, run the free Article 50 check.
Quick answers
Is our bank chatbot high-risk under the AI Act?
Not because it talks to customers: Article 50 applies whatever the risk class. An assistant that assesses creditworthiness is also high-risk under Annex III, point 5(b), from 2 December 2027.
Who is responsible when we bought the chatbot from a vendor?
The vendor, as provider, carries paragraphs 1 and 2. The bank, as deployer, carries paragraphs 3 and 4 and controls what the customer sees. Under its own brand, a bank can be the provider.
Can our chatbot have a human first name?
It can have a name, but it must say it is an AI at the first interaction and never deny it when asked. A human name with a photograph rules out the “obvious” exception.
What changes on 2 December 2026?
The provider of a chatbot placed on the market before 2 August 2026 must mark its generated output in a machine-readable way. Two new Article 5 prohibitions apply the same day; they do not target a bank assistant.
How do we test disclosure?
Run the six tests on the live assistant and send each one ten times. The finding is a pass rate, and every exchange is kept with the model and prompt versions.
Sources
Vidimus tests your live assistant against Article 50, ten attempts per test, and files the result in a signed evidence pack.