Pull to refresh
Logo
What Changed
Following Why Sign Up
EU starts enforcing its AI Act

EU starts enforcing its AI Act

Rule Changes

Two-year grace period ends; Brussels can now fine the makers of large AI models

August 2nd, 2026: Enforcement and transparency duties begin

Overview

Updated Aug 2

For two years, Europe's AI rulebook came with a grace period. On August 2, 2026, that period ended.

The European Commission can now fine the makers of large AI models up to 15 million euros or 3% of global revenue, whichever is larger. Its AI Office can demand training documents, test the models, and order changes. Chatbots must tell users they are machines, and AI-made images, audio, video, and text must be labeled.

Why it matters

Any AI chatbot used in Europe now has to tell you it is not human, and the deepfakes in your feed must carry a label.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

3%
Top fine as share of global revenue
Or 15 million euros, whichever is larger, for breaking the model rules.
€3.3B
Estimated yearly EU compliance cost
Figure cited by trade group DIGITALEUROPE while lobbying for delays.
450M
People in the EU single market
The user base any AI system deployed in the bloc must now cover.
125+
AI Office staff
Only some of them supervise general-purpose models across 100-plus duties.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Log in Sign Up

People Involved

Henna Virkkunen
Henna Virkkunen
European Commission Executive Vice-President for Tech Sovereignty, Security and Democracy
Overseeing AI Act enforcement and the simplification package
 Joel Kaplan
Joel Kaplan
Chief Global Affairs Officer, Meta Platforms
Leads Meta's refusal to sign the voluntary code

Organizations Involved

European AI Office
European AI Office
EU regulatory body
Lead enforcer over general-purpose AI models

The Commission unit that supervises the makers of large AI models across the EU.
DIGITALEUROPE
DIGITALEUROPE
Industry trade association
Lobbied for lighter and delayed rules

A Brussels trade group that pushed for the AI Act to be simplified and delayed.

Timeline

August 2024 December 2026

7 events Latest: August 2nd, 2026 · 3 months ago
Tap a bar to jump to that date

  1. Marking grace period ends

    Upcoming Legal

    Generative AI systems already on the market before August 2, 2026 must meet the machine-readable marking requirement by this date.

  2. Enforcement and transparency duties begin

    Latest Legal

    The AI Office can now investigate model makers and propose fines up to 15 million euros or 3% of global revenue. Chatbots must disclose they are AI, and synthetic content must be labeled.

  3. Council clears the Digital Omnibus

    Policy

    The Council of the EU gives final approval to the amendments, delaying high-risk AI duties to December 2, 2027 while keeping the August 2026 deadlines.

  4. Commission proposes Digital Omnibus

    Policy

    Brussels proposes easing and delaying parts of the Act after industry pushback over cost and complexity.

  5. Model rules apply; code signatories listed

    Legal

    Duties for general-purpose AI providers apply to new models. The Commission publishes the list of firms that signed the voluntary Code of Practice.

  6. First bans take effect

    Legal

    Prohibited AI practices are banned and AI literacy duties begin, the earliest hard deadline under the Act.

  7. AI Act enters into force

    Legal

    The EU's Artificial Intelligence Act becomes law, starting a staggered rollout of duties over several years.

Scenarios

1

Commission opens its first formal case against a major model maker

Possible Resolves by Aug 2, 2027

Discussed by: TechPolicy.Press and Reuters coverage of AI Office powers

The AI Office uses its new authority to open a formal investigation or propose a fine against a named general-purpose AI provider. A likely trigger is a refusal to hand over documentation, a failed model evaluation, or a missed transparency duty. Meta, the most prominent holdout on the voluntary code, is the name most often floated as an early test case.

2

EU delays or softens the transparency rules again

Unlikely Resolves by Aug 2, 2027

Discussed by: The Register and IAPP coverage of the Digital Omnibus

After delaying the high-risk rules in the June 2026 Omnibus, Brussels goes further and postpones or weakens the transparency and labeling duties under continued industry pressure. This would need a fresh Commission proposal that clears both the Council and Parliament. Supporters would frame it as simplification; critics as a retreat.

3

A major provider limits or pulls a model from the EU

Possible Resolves by Aug 2, 2027

Discussed by: Analytics Insight and TECHi coverage of exposure risk

Rather than meet a specific obligation, a large model maker restricts a product in Europe or delays an EU launch and blames the AI Act, echoing earlier feature holdbacks by Meta and Apple over other EU rules. A public statement tying the decision to the Act would confirm it. Most firms are instead aligning global products with EU standards, which cuts against this path.

4

Enforcement stays light through the first year

Likely Resolves by Aug 2, 2027

Discussed by: Coverage noting the AI Office's limited staffing

The AI Office, stretched across more than a hundred duties with a small model-supervision team, spends the first year on documentation reviews and guidance rather than fines. Firms adjust quietly, and no formal penalty lands. This is the base case if the office prioritizes cooperation over confrontation while it scales up.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

May 2018

GDPR takes effect (2018)

The EU's General Data Protection Regulation began applying across the bloc after a two-year grace period. It set fines up to 4% of global revenue and forced companies worldwide to change how they handle personal data.

Then

Firms rushed to add consent banners and privacy notices, and early enforcement was slow as regulators staffed up.

Now

GDPR became a global template; several countries copied it, and large fines against US tech firms followed years later.

Why this matters now

The AI Act follows the same playbook: a grace period, revenue-based fines, and reach far beyond Europe's borders. GDPR's slow start also hints at how AI enforcement may open.

March 2004 to March 2013

Microsoft EU antitrust fines (2004-2013)

The European Commission fined Microsoft over how it bundled software and shared technical information, with penalties eventually topping 2 billion euros. Brussels ordered product changes for the EU market.

Then

Microsoft appealed, complied in stages, and shipped modified products for Europe.

Now

The case showed Brussels would take on the largest US tech firms and reshape their products to meet EU rules.

Why this matters now

It is a reminder that the Commission has both the will and the track record to force changes on dominant American technology companies, the same firms now facing the AI Office.

Sources

(10)
+4
European Commission EU Artificial Intelligence Act (Future of Life Institute) Analytics Insight SpaceDaily / AFP Digital Watch Observatory The European Post The Register Tech Policy Press Jones Day EU Artificial Intelligence Act (Future of Life Institute)