Watch the video with chapters & transcript →
AI TOOLS
In short: Article 50 of the EU AI Act took effect on 2 August 2026 — before the high-risk rules, which were pushed back to 2 December 2027.
In this video, Ani Björkström, the Stockholm-based AI consultant behind the YouTube channel "Ani Björkström | AI for Finance," walks through Article 50 of the EU AI Act — the transparency provision that quietly went live on 2 August 2026 while most attention was fixed on the high-risk AI rules now delayed to December 2027. The video is aimed at anyone running a chatbot, publishing AI-generated content, or using AI on customer faces and voices, and it distils the law into four practical rules plus a five-question self-audit.
Article 50 is not a safety or paperwork rule — it is a honesty rule. It covers four distinct situations: AI chatbots talking to people, AI-generated media (images, video, audio, text), emotion-recognition and biometric-categorisation systems, and deepfakes of real people, places or events. Each situation carries its own disclosure requirement, and the video stresses that the disclosure must be visible where people actually see it — not buried in a caption, a description, or hidden metadata inside a file.
The rule most companies are already breaking, according to the video, is the chatbot rule: if a customer could reasonably believe they are talking to a human, the AI must say otherwise at the start of the conversation, in clear and easy-to-notice language — a single sentence such as "You are chatting with an AI assistant" is enough.
The video draws a sharp line between the system and the content it produces. If an AI system was already running before 2 August 2026, the operator has until 2 December 2026 to build machine-readable marking into that tool. That deadline applies to the tool, not to past posts — content made and published before 2 December 2026 does not need to be labelled retroactively. Once the tool passes that deadline, everything it publishes afterward must carry the mark. A small logo in the corner of an image is described as not sufficient on its own; the label must be readable by a machine, not just visible to a human eye.
The video highlights a liability split that catches people off guard. For the chatbot disclosure rule, the company that built the AI system carries the responsibility. But for emotion-recognition, biometric categorisation, and deepfakes, it is the user of the AI — the company deploying it in a call centre, a shop, or a marketing campaign — that carries the legal responsibility. "We just bought the tool" is explicitly called out as not a valid defence.
| Rule | Who it applies to | What to do |
|---|---|---|
| Rule 1 — Chatbots | Website chat, phone support, booking assistants | State clearly at the start of the conversation that the user is talking to AI |
| Rule 2 — AI-generated content | AI images, video, audio, text | Machine-readable mark; systems running before 2 Aug 2026 have until 2 Dec 2026 to comply |
| Rule 3 — Emotion/biometric AI | Call centres, shops, interview-scoring tools | Disclose at the moment the AI is used, not in a policy document |
| Rule 4 — Deepfakes | Marketing, AI avatars of real people, AI-written news text | Disclose the content is synthetic; softer disclosure allowed for art/satire; human editor can substitute for disclosure on news text |
The video closes with five self-check questions companies are told to answer this week: does every AI chat clearly say it is AI at the start; does every AI-generated output carry a machine-readable mark; does the company use anything that reads emotion or sorts people by appearance; does it publish AI images, video or voice of real people; and does it publish AI-written text about news or public affairs without a human editor taking responsibility. A "yes" to the fourth or fifth question without an existing disclosure is flagged as something to fix immediately, while the 2 December 2026 date is the one deadline every operator of a pre-existing AI system should already have in their calendar.
On 2 August 2026. The high-risk AI rules that most coverage expected on that date were separately delayed to 2 December 2027.
Up to €15 million or 3% of global annual turnover, whichever is larger. For a company earning €1 billion a year, that means a €30 million exposure, not €15 million.
No. If your AI system was already running before 2 August 2026, its labelling deadline is 2 December 2026, and that applies only to the tool going forward — content published before that date does not need to be retroactively marked.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?