The EU AI Act explained simply: what's banned, what changes on December 2 and how it protects you

A plain-English guide to the European Union's AI Act in 2026: banned uses, mandatory labels for chatbots and deepfakes, the new ban on 'nudify' apps and the timeline after the Digital Omnibus.

Illustration of a head with circuits representing artificial intelligence
Image: DeltaWorks on Pixabay

Europe was the first to pass a full law on artificial intelligence: the AI Act. It was approved in 2024, but it kicks in in stages, and part of the timeline was changed in 2026. The result: plenty of people don’t know what already applies and what doesn’t.

Here it is without the legal jargon, focused on what matters to you as a user.

Note: this is a general explainer, not legal advice. If your business uses or sells AI, talk to a specialist lawyer.

The idea in one sentence

The law sorts AI by the risk it poses to people. The more harm it could do, the more rules it has to follow:

Risk level Examples What the law says
Unacceptable Scoring citizens on their behaviour, manipulating vulnerable people Banned
High AI that screens CVs, approves loans, grades exams or helps the police Allowed, with strict checks
Transparency Chatbots, fake images and videos (deepfakes) Must tell you it’s AI
Minimal Spam filters, video games, recommendations Almost no obligations

Banned since February 2025

Since February 2, 2025, the EU has banned AI systems that, among other things:

  • Score people on their social behaviour (so-called “social credit”).
  • Manipulate people with deceptive tricks or by exploiting their age, disability or financial situation.
  • Read emotions at work or in school (except for medical or safety reasons).
  • Scrape faces from the internet or CCTV in bulk to build facial-recognition databases.
  • Try to predict whether someone will commit a crime based only on their profile or personality.
  • Identify people by their face in real time on the street, with very narrow exceptions for police (for example, searching for a kidnapping victim).

Breaking these bans carries the law’s biggest fines: up to €35 million or 7% of a company’s worldwide turnover, whichever is higher.

2025: rules for ChatGPT, Gemini and co.

Since August 2, 2025, the companies behind large general-purpose AI models (the ones powering ChatGPT, Gemini, Claude or Llama) have their own obligations: document how their models work, respect copyright and publish a summary of the data used to train them. The most powerful models must also assess and reduce their risks.

August 2026: they have to tell you it’s AI

Since August 2, 2026, the transparency rules apply:

  • If you’re talking to a chatbot, it must be clear it isn’t a human (unless that’s obvious).
  • Deepfakes (fake images, audio or video that look real) must be labelled as AI-generated or AI-manipulated.
  • AI-generated content must carry a technical marker (a kind of watermark machines can detect). Tools already on the market before August get extra time, until December 2, 2026.

Want to check a suspicious photo yourself? See our guide on how to tell if an image is AI-generated.

The “Digital Omnibus”: the 2026 timeline change

In 2026 the EU passed a package of changes called the Digital Omnibus on AI (Regulation (EU) 2026/1744). It was published in the Official Journal on July 24 and has been in force since July 27, 2026, according to law firm Lewis Silkin. It does three main things:

1. It delays the high-risk AI rules

Companies asked for more time and got it:

Type of high-risk AI Original date New date
AI for hiring, credit, education, policing, public services… August 2, 2026 December 2, 2027
AI built into regulated products (toys, medical devices…) August 2, 2027 August 2, 2028

In other words, the rules that would protect you most when an AI decides on your job or your loan don’t apply yet. Critics such as Tech Policy Press argue the delay leaves those systems without oversight for longer.

2. It bans “nudify” apps

This is the biggest change for ordinary people. The law adds to its list of bans AI systems that create intimate images of people without their consent (apps that “undress” photos) or child sexual abuse material. According to Gibson Dunn, the ban also covers tools where that output is foreseeable and easy to get because they lack reasonable safeguards.

The ban applies from December 2, 2026.

3. It cuts red tape for smaller companies

Small and mid-sized businesses get lighter paperwork and priority access to the regulatory “sandboxes” countries set up to test AI under supervision.

Full timeline

Date What starts to apply
February 2, 2025 Bans, and the duty to train staff in AI
August 2, 2025 Rules for large models (ChatGPT, Gemini…)
August 2, 2026 Chatbot disclosure and deepfake labels
December 2, 2026 Ban on nudify apps; watermarks for older tools
December 2, 2027 High-risk AI (hiring, credit, education…)
August 2, 2028 High-risk AI inside regulated products

Does it apply outside the EU?

Yes, in a way. The law covers any company that offers AI to people in the EU, wherever it’s based. That’s why US companies like OpenAI and Google have to comply in Europe. If you live in the UK or the US, it doesn’t protect you directly, but it often shapes how global apps are designed. Each EU country has its own watchdog; in Spain, for example, it’s the AESIA agency.

What you can do

  • If you come across an unlabelled deepfake or an app that creates fake nudes, report it to the platform and, if it involves you or someone close to you, to the police.
  • If a chatbot pretends to be a person without telling you, you can complain to the company.
  • Be wary of striking videos and audio even when they look real: voice-cloning scams keep growing. Here’s how to protect your family from AI voice scams.

In short

The EU AI Act already bans the most dangerous uses, requires a warning when you’re talking to a machine or watching a deepfake, and from December 2 will ban apps that create fake nudes. What has been delayed are the rules for AI that decides on your job, your credit or your education: those arrive at the end of 2027.

Sources