Ackatec
All insights

EU AI Act December 2 Deadline: How U.S. Software Teams Should Mark AI-Generated Output

Generative AI features on the EU market before August 2, 2026 must mark output as AI-generated by December 2. Who is in scope and how to prepare securely.

Author
Ackatec
Published
Reading time
7 min read
  • AI governance
  • EU AI Act
  • Compliance
  • AI security
  • Penetration testing

On December 2, 2026, a four-month grace period in the EU AI Act runs out. Providers of generative AI systems that were already on the EU market before August 2, 2026 must, by that date, mark the text, image, audio, and video their systems produce in a machine-readable way that shows it was artificially generated or manipulated. The transition comes from the Digital Omnibus on AI (opens in a new tab), Regulation (EU) 2026/1744, which added it as new Article 111(4) (opens in a new tab) of the AI Act.

The short answer: if your company sells or offers a product under its own name that generates synthetic content, and that output is used in the EU, decide now whether you are a provider under Article 50(2). If you are, put a marking and detection approach in place before December 2, and treat the signing keys, detection service, and anti-tampering terms as security controls, not just compliance paperwork.

This post explains the obligation for security and compliance leaders. It is not legal advice, so confirm your role and scope with counsel.

What does Article 50(2) of the EU AI Act require?

Article 50(2) (opens in a new tab) says providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video, or text content must ensure the outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated." The solutions must be effective, interoperable, robust, and reliable as far as technically feasible.

There is a narrow exception for systems that perform an assistive function for standard editing or do not substantially alter the deployer's input or its meaning. A grammar fixer may qualify. A feature that drafts new text, generates images, or clones a voice is unlikely to.

Why is December 2, 2026 the date that matters?

Article 50 has applied since August 2, 2026 under Article 113 (opens in a new tab). A generative system placed on the market after that date had to comply from launch. The Omnibus gave only systems already on the market before August 2, 2026 extra time: Article 111(4) says those providers "shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026." Recital 38 of the Omnibus describes this as a four-month transitional period.

The Omnibus pushed the high-risk AI rules to December 2027 and August 2028, but after December 2 every in-scope generative provider must mark output, whenever its system launched.

Does the EU AI Act apply to U.S. companies?

It can. Article 2 (opens in a new tab) applies the Act to providers placing AI systems on the EU market whether or not they are established in the Union, and to providers and deployers in a third country where the output produced by the AI system is used in the Union.

The role question deserves care. Under Article 3 (opens in a new tab), a provider is an organization that develops an AI system, or has one developed, and places it on the market or puts it into service under its own name or trademark, whether paid or free. A SaaS company that wraps a third-party foundation model in a branded writing assistant, image tool, or support bot may therefore be the provider of that AI system, even though it did not train the model. Do not assume your model vendor's marking covers your product; confirm what it marks and whether that survives your own processing.

What happens if you miss the deadline?

Article 99 (opens in a new tab) sets the ceiling for breaches of Article 50 transparency obligations at up to EUR 15 million or 3 percent of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and startups, the cap is whichever of the two is lower. Member States set the actual penalty rules.

Commercial pressure may arrive sooner. Enterprise buyers with EU exposure may start asking about AI content marking in security questionnaires, and a vague answer can slow a deal.

How do you meet the marking obligation in practice?

The law does not prescribe a single technique. The European Commission published a voluntary Code of Practice on Transparency of AI-generated Content (opens in a new tab) on June 10, 2026. In July 2026, the Commission and the AI Board assessed the code (opens in a new tab) as adequately covering Article 50(2), (4), and (5). The Commission also notes that adherence is not conclusive evidence of compliance. Providers that do not sign must be able to show their own measures are adequate.

Germany's Federal Network Agency, the Bundesnetzagentur, summarizes the provider commitments (opens in a new tab). For signatories, they include:

  • Signed metadata. Where the file format supports metadata, record whether content is AI-generated or AI-manipulated, digitally signed and time-stamped in a tamper-evident way.
  • Imperceptible watermarking. Embed a watermark that is hard to separate from the content, except for very short text, either during generation or after it.
  • Optional fingerprinting or logging. Keep a record that can later confirm whether your system produced a given piece of content.
  • No stripping. Retain existing metadata markings, and prohibit intentional removal or tampering in your terms and conditions or documentation.
  • Detection. Offer a detection solution, such as a public specification, a software library, or an API, with a mechanism for each marking technique you use, and let users download a signed detection result.

Where are the security risks in AI content marking?

Marking is a trust system, and trust systems attract attackers. Several commitments in the code are really security controls, so this deadline belongs on the security team's agenda as well as legal's.

Protect the signing keys

Signed metadata is only as trustworthy as the private key behind it. If an attacker steals that key, they can label content as coming from your system, or make AI-generated content look authentic. Store signing keys in a hardware security module or managed key service, restrict which services can call them, log every signing operation, and rehearse rotation. Treat key compromise as an incident scenario in your incident readiness planning.

Treat the detection service as an attack surface

A public detection API accepts untrusted uploads, so it inherits the usual risks of file parsing, resource exhaustion, and data exposure. The code expects detection services that require uploads to follow data minimization and purpose limitation, delete content immediately after detection, and apply security safeguards for confidentiality, integrity, availability, and resilience. An authorized penetration test of the detection endpoint, upload handling, and rate limits can show whether those claims hold before regulators or researchers test them for you.

Test whether markings survive your own pipeline

Many products resize, transcode, or compress generated content before the user sees it, and each step can drop metadata or weaken a watermark. Add regression tests that run real outputs through every export path and confirm the markings are still detectable at the end.

What should you do before December 2?

  1. Inventory every product feature that generates text, images, audio, or video, and note when each was first offered to EU users.
  2. Decide, with counsel, whether you are a provider or a deployer for each feature, and whether the assistive-editing exception applies.
  3. Ask each model vendor what it marks, how, and what detection it provides, then verify the answer with your own outputs.
  4. Choose a marking approach, either by signing the code or by documenting an equivalent approach you can defend.
  5. Put signing keys, the detection service, and export paths under normal security ownership, testing, and monitoring.
  6. Update terms of service to prohibit removing or tampering with markings, and keep evidence of what you shipped and when.

If you are also tracking other EU deadlines, our post on EU Cyber Resilience Act reporting covers the vulnerability-reporting duties that began this year, and testing AI agents before launch covers the attack paths that generative features open.

Getting ready without a separate compliance track

Treat December 2 as one governance and security exercise. AI governance and readiness work can build that inventory, ownership model, and review cadence, and compliance readiness can connect the evidence to the other frameworks your buyers already ask about.

If you are not sure whether your product is in scope, or whether your marking would hold up to scrutiny, talk with Ackatec. Preparing these controls reduces regulatory and security risk but does not guarantee any particular regulatory outcome.

References