This post tries to sum up the essentials of the AI Act, and on top of that you get the downloadable icons the EU just released for labelling AI-generated content.
1. Risk classification
Regulation (EU) 2024/1689 sets out four tiers. Which one applies doesn't depend on the tool you use, but on what you do with it and who it acts on.
Prohibited
In force since Feb 2025. Social scoring, harmful manipulation, emotion recognition at work or in education, certain biometric uses. No exceptions.
High risk
Hiring, educational assessment, credit, migration, justice. Requires controls, documentation, traceability and real human oversight — not a rubber stamp.
Transparency
Chatbots, deepfakes, artificially generated content. People need to know there's a machine behind it. This is where the icons come in.
Minimal risk
The bulk of everyday uses. No specific obligations under the Regulation.
2. What sets the risk tier
This is the point most often misread. Summarising internal minutes is very different from automatically filtering job applications. In the second case, you're making decisions that affect someone's access to employment.
And it applies to any organisation using AI — even AI contracted out to a third party. Knowing which platform you use isn't enough. You have to look at what it's for, who it acts on, and what the consequences are.
Before classifying a system, ask yourself: what decision does it produce? About whom? Could it affect their rights or opportunities? Is anyone actually reviewing the output?
3. The labelling icons
The transparency obligation takes full effect in August 2026, but the icons are already available. The European Commission has published them as a free resource to make compliance easier. Three situations, three labels:
Basic
When AI played a part in creating deepfake content or published text, or when a custom text label is added instead. E.g. a video with "voices generated with AI" plus the icon.
100% AI — AI Generated
The entire piece is AI-generated with no human input beyond the prompt. E.g. deepfakes of fictional politicians, AI-composed music, news summaries with no editorial review.
Modified — AI Modified
Real content partly modified with AI, turning it into a deepfake or into text of public interest. E.g. swapping a politician's face in a real photo, AI-furnishing photos of a flat.
All three come in four variants: black and white, each at 100% and 50% opacity, to suit any background. Here are the official files, ready to download:
The icon has to meet four conditions:
- Be visible from the first moment
- Be built into the piece — not overlaid on top
- Stay visible when the content is shared or downloaded
- Come with plain-language text alongside it
When you don't need to label: when the content is evidently art, satire or fiction; when it's legally authorised for criminal investigation; or when the text has gone through human review with clear editorial responsibility. If a person signs off on it and reviews it, it isn't mandatory.
That said: using the icon alone doesn't guarantee compliance. The goal isn't to dodge fines. It's that nobody mistakes what's real for what's artificial.
4. AI Act implementation timeline
| Date | What applies |
|---|---|
| Feb 2025 | Prohibitions in force + basic AI literacy training |
| Aug 2025 | Rules for general-purpose models (GPT, Gemini…) |
| Aug 2026 | General application of the Regulation |
| Dec 2027 | High-risk rules for education, employment, biometrics and justice |
Official sources: Regulation (EU) 2024/1689 · European Commission — AI Act regulatory framework. September 2026.