VAARHAFT
EU AI ACT · ARTICLE 50
02AUG2026

Article 50 makes unlabelled AI images a compliance violation — from this date.

The transparency obligations of the EU AI Act apply from 2 August 2026. AI-generated and AI-manipulated images must be labelled — machine-readable by providers, visibly disclosed by deployers. Violations carry fines of up to 15 million euros or 3% of worldwide annual turnover.

€15 M

maximum fine per violation — or 3% of worldwide annual turnover, whichever is higher (Art. 99(4))

50 (2)+(4)

the two paragraphs that govern AI images: machine-readable marking and visible deepfake disclosure

62 %

success rate of humans detecting AI-generated images — barely better than chance (Microsoft AI for Good Lab, 2025)

72 h

is all it takes to audit your complete image inventory — the plan is below

The obligations

What applies from 2 August 2026

Article 50 distinguishes between those who build generative AI systems and those who use them. Both roles carry their own duties — and most companies working with AI images are deployers without knowing it.

ART. 50 (2)
Providers

Machine-readable marking

Providers of AI systems that generate synthetic image, audio, video or text content must mark outputs in a machine-readable format so they are detectable as artificially generated or manipulated — effective, interoperable and robust, as far as technically feasible.

ART. 50 (4)
Deployers

Visible deepfake disclosure

Anyone publishing AI-generated or AI-manipulated images, audio or video that depict real persons, objects, places or events must disclose this clearly. Artistic and satirical contexts enjoy a lighter regime — commercial communication does not.

ART. 99 (4)
Enforcement

Fines from day one

National market surveillance authorities can penalise violations of Article 50 with fines of up to 15 million euros or 3% of total worldwide annual turnover. There is no grace period — the obligations are enforceable from 2 August 2026.

EXEMPTIONS
Scope limits

What does not fall under it

Assistive functions and standard editing that do not substantially alter the input — colour correction, cropping, moderate retouching — remain outside the labelling duty. Where exactly the line runs is set out in the Commission's transparency guidelines of 20 July 2026.

Who is liable

Most companies are deployers — and liable themselves

ART. 50 (2)

Provider

Develops a generative AI system or substantially modifies one. Owes the machine-readable marking inside the output itself — watermarks, metadata, provenance signals.

Model vendors · image-generation platforms · companies fine-tuning their own models

ART. 50 (4)

Deployer

Uses AI systems under its own authority and publishes the output. Owes the visible disclosure — regardless of which tool generated the image or which agency delivered it.

Marketing teams · e-commerce · insurers · media houses · clients of agencies

The role is assessed per use, not per company — one organisation can be provider and deployer at the same time.

The legal framework

Four dates led up to enforcement

10 JUN 2026

Code of Practice on Transparency of AI-Generated Content

The AI Office's code of practice — the practical yardstick for what counts as adequate marking and disclosure.

20 JUL 2026

Commission guidelines on transparency obligations

The adopted guidelines interpret Article 50: scope, exemptions, and how machine-readable marking and visible disclosure interact.

24 JUL 2026

Digital Omnibus on AI in the Official Journal

Regulation (EU) 2026/1744 is published — the final adjustments to the framework, days before applicability.

02 AUG 2026

Article 50 applies

The transparency obligations and the fine regime become enforceable in all 27 member states.

The 72-hour audit

Audit your image inventory in 72 hours

Three days from kick-off to a remediation list. The Fraud Scanner classifies your inventory at scale; our team turns the findings into concrete labelling actions.

01DAY 1

Inventory & scan

We compile your image inventory — website, campaigns, product images, social media — and run it through the Fraud Scanner in bulk.

02DAY 2

Classification

Every image is classified: AI-generated, AI-manipulated or authentic — including its current marking status, metadata and provenance signals.

03DAY 3

Report & actions

You receive a prioritised report: which images need labelling, which should be replaced, and where your publishing process needs a checkpoint.

The tooling

Detection that works without watermarks

Markings can be stripped and metadata rarely survives the upload chain. Compliance needs detection that judges the image itself — not just its label.

Fraud Scanner

Classifies AI-generated and manipulated images at scale — with heatmaps and reports your compliance documentation can reference.

Explore the Fraud Scanner

Image Trust Layer

The detection technology underneath: trained on current and older generators, robust against compression, cropping and re-uploads.

How the technology works

Frequently asked questions

Yes. The obligations attach to the role — provider or deployer — not to company size. Fines are calibrated to turnover, but the duty to label applies from the first published AI image.

C2PA is a strong basis for the machine-readable marking under Article 50(2). It does not replace the visible disclosure deployers owe for deepfake-type content under Article 50(4) — and metadata is often stripped on the way to publication, so verification remains your job.

Standard editing — colour correction, cropping, moderate retouching — counts as an assistive function and is exempt. Substantial manipulation of what the image shows falls under the labelling duty. The Commission's guidelines of 20 July 2026 draw the line in detail.

Whoever publishes content under their own authority is the deployer — usually you, not the agency. Contracts can allocate the work internally, but not the regulatory responsibility towards the authority.

National market surveillance authorities can investigate, order corrective measures and impose fines of up to 15 million euros or 3% of worldwide annual turnover — whichever is higher.

No. This page summarises the regulatory framework for information purposes. For an assessment of your specific setup, involve your legal counsel — we are happy to contribute the technical evidence.

This page is provided for information purposes and does not constitute legal advice.

Get started

Talk to our experts

Get a detailed insight into how you can protect yourself from image‑based fraud