Machine-Readable AI Labeling with C2PA, IPTC & Watermarks Explained

Last updated: July 21, 2026

Machine-readable AI labeling, technically carried by C2PA, IPTC metadata, and invisible watermarks, is the part of the EU AI Act that raises the most questions. Unlike a visible EU icon or text such as "AI generated," this marking is invisible to humans: it sits inside the file itself and is read by software. This article explains the three methods in detail, what Article 50(2) of the EU AI Act specifically requires, why a combination makes sense, and how to write the markings into your images and videos step by step.

Key Takeaways

• Mandatory from August 2, 2026: providers of synthetic content must mark their AI outputs as artificially generated in a machine-readable way under Article 50(2) of the EU AI Act.
• Three methods are established: IPTC/XMP (DigitalSourceType), C2PA / Content Credentials (signed manifest), and an invisible watermark (anchored in the pixels).
• No single technique meets all four criteria, "effective, interoperable, robust, reliable." The combination is the safe path.
• Transition period: for systems placed on the market before August 2, 2026, the machine-readable marking has been postponed to December 2, 2026 under the Digital Omnibus (provisional, as of May 2026). The core deadline and the visible deployer obligation (paragraph 4) remain at August 2, 2026.
• Deployers (e-commerce, marketing, D2C) are primarily subject to the visible obligation under paragraph 4. The machine-readable layer is a strong additional signal for them.
• Sanctions: violations of the transparency obligations can be fined up to 15 million euros or 3% of global annual turnover.

Table of Contents

1. What Article 50(2) requires for machine-readable AI labeling
2. IPTC DigitalSourceType: the metadata foundation
3. C2PA / Content Credentials: signature and provenance
4. Invisible watermark: survives recompression
5. The three methods compared
6. Practical scenarios and edge cases
7. Guide: writing metadata yourself
8. Compliance checklist
9. FAQ

What Article 50(2) of the EU AI Act Requires for Machine-Readable AI Labeling with C2PA

Article 50 of the AI Act is titled "Transparency Obligations for Providers and Deployers." Paragraph 2 addresses providers of AI systems that generate synthetic image, audio, video, or text content. They must mark their outputs so that they are machine-readable and identifiable as artificially generated or manipulated. This is the first time machine-readable AI labeling with C2PA and related methods has been anchored in law.

The legal text deliberately does not prescribe a specific technology. However, it requires that the solution be implemented in a way that is "effective, interoperable, robust, and reliable, to the extent technically feasible." These four criteria are exactly where the difficulty lies: a single method rarely fulfills them completely. That is why the European Commission's guidelines on the transparency obligations explicitly name three established approaches, IPTC metadata (XMP), C2PA / Content Credentials, and watermarks, that can be used individually or in combination.

The key deadlines at a glance:

The core deadline therefore remains August 2, 2026 (24 months after entry into force on August 1, 2024). The Digital Omnibus only postpones the machine-readable obligation (paragraph 2) for legacy systems; the visible deployer labeling (paragraph 4) is not postponed. You can find a complete overview of deadlines and fines under AI Act deadlines and fines.

A note on roles: most companies in e-commerce, marketing, or D2C are legally deployers and are primarily subject to the visible labeling obligation under paragraph 4. The machine-readable obligation under paragraph 2 primarily applies to the providers of the generators (such as OpenAI, Google, Adobe). In practice, it is still wise to cover both levels. The exact distinction is explained in the article Provider or deployer, who has which obligations. If you want to understand the overall legal picture first, you will find the basics in the pillar article on AI labeling obligations under the EU AI Act.

IPTC DigitalSourceType: The Metadata Foundation (XMP)

The simplest and most widely supported route is through IPTC Photo Metadata. The relevant field is called DigitalSourceType and is stored in the XMP block of the file, the same standard used by stock photo agencies, editorial systems, and programs such as Adobe Photoshop, Lightroom, or Capture One.

For AI content, two values matter most:

• trainedAlgorithmicMedia, the content was generated entirely by a trained AI model (classic "text to image," such as a fully generated product image).
• compositeWithTrainedAlgorithmicMedia, an existing image was partially altered or supplemented with AI (such as a real product photo with an AI-generated background or AI retouching).

This distinction matches the logic of the optional EU icons, which separate "fully AI-generated" from "partially AI-modified." More on this under EU icons for AI content.

Concrete example: a fashion retailer generates a catalog image entirely with AI. Set correctly, the field then reads DigitalSourceType = trainedAlgorithmicMedia. If the same retailer takes a real photo and only replaces the background with AI, compositeWithTrainedAlgorithmicMedia is the correct value. When checking a third-party file, the field therefore signals not just whether AI was involved, but also to what extent.

The advantage: XMP is an open, documented standard read and written by countless programs. An AI image with a correctly set DigitalSourceType signals "artificially generated" across platforms.

The disadvantage: XMP metadata is not signed and easy to remove. A simple screenshot, a format conversion, or a re-upload on many social media platforms wipes it out completely. IPTC is therefore a necessary but not a tamper-proof foundation if you want to write AI image metadata.

C2PA / Content Credentials: Signature and Provenance

C2PA (Coalition for Content Provenance and Authenticity) goes considerably further. Instead of just setting a field, C2PA attaches a cryptographically signed manifest to the file, known as Content Credentials. It records, in a traceable way, what the content was created with, when, and which editing steps followed. This is referred to as provenance: a verifiable chain of origin.

The decisive difference from plain IPTC is the digital signature. If the file is manipulated afterward, the signature breaks, and a verification tool detects the discrepancy. A C2PA viewer, such as the open verification portal from the Content Authenticity Initiative, can read out and verify the manifest. This approach best meets the "robust" and "reliable" criteria from Article 50(2).

What a C2PA manifest typically contains:

• the creator or the generating AI tool (software agent),
• the timestamp of creation,
• the chain of editing steps ("assertions"),
• an indication of AI generation (trainedAlgorithmicMedia, analogous to IPTC),
• the signature of the issuing certificate.

C2PA is backed by a broad industry alliance (including Adobe, Microsoft, and leading camera manufacturers) and is considered the de facto standard for content provenance. Major generators such as the Adobe Firefly models already embed Content Credentials by default.

The weakness: the C2PA manifest can also be lost if a platform strips all metadata during upload or re-encodes the image. The signature protects against unnoticed manipulation, not against complete removal. That is exactly why the third technique comes into play.

Invisible Watermark: Survives Recompression

An invisible (robust) watermark embeds the label directly in the image's pixel data, not in a metadata block. It is not perceptible to the human eye, but a matching detector can read it out. Well-known methods for an invisible watermark on AI content include SynthID (Google) or classic frequency-domain watermarks.

The big advantage: such a watermark survives recompression, scaling, and moderate editing, and is not lost even when a platform removes all metadata. Where IPTC and C2PA fail due to metadata stripping, the pixel watermark holds up. It is therefore the most resilient of the three layers, particularly relevant for content distributed via Meta and Instagram ads or marketplaces such as Amazon and eBay, where metadata is regularly lost during upload.

The limits: watermarks are more proprietary and less interoperable; without the matching detector, the information cannot be read out. Heavy editing, tight cropping, or renewed AI processing can weaken or destroy the signal. As a standalone solution, it is therefore also insufficient. For video content, additional nuances come into play, covered in the article Labeling AI videos.

The Three Methods Compared, and Why the Combination Wins

None of the three techniques fulfills all four requirements (effective, interoperable, robust, reliable) on its own. Only working together do they cover each other's weaknesses.

The logic of combining methods:

• IPTC/XMP provides the universally readable base marking.
• C2PA adds tamper resistance and a verifiable chain of origin.
• The watermark ensures the label survives even when all metadata disappears.
• A visible EU icon or the notice "AI generated" additionally fulfills the deployer obligation under paragraph 4 toward the human viewer. The official EU icon set is optional; the labeling obligation itself is not.

This is exactly the layered strategy meant when Article 50(2) demands "effective, interoperable, robust, and reliable." For providers, the combination is the safest path to compliance; for deployers, it is a strong additional signal that documents their own diligence.

Practical Scenarios and Edge Cases

The theory sounds clear, but in practice the details decide. Three typical cases:

Scenario 1, AI product image for the online shop: a D2C retailer generates a hero image entirely with AI. As a deployer, they only owe visible labeling insofar as it involves a deepfake or content of public interest; a plain product image usually does not fall under that. Even so, the machine-readable marking plus an optional EU icon is recommended, to avoid cease-and-desist risk and platform downranking. Details on the retail context: AI labeling in e-commerce and marketing.

Scenario 2, deepfake in advertising: if a realistic-looking person is generated by AI or their voice is cloned, paragraph 4 applies directly: the deployer must visibly disclose the content as AI generated, from August 2, 2026, with no postponement. What exactly counts as a deepfake is explained in Deepfake labeling under the EU AI Act.

Scenario 3, art and satire: a lighter standard applies to obviously artistic, creative, or satirical content. Disclosure may be done in a way that does not impair the work, for example in accompanying text rather than as an intrusive overlay. The machine-readable layer is unaffected by this and can be embedded without issue.

Edge case, metadata stripping: the most common practical mistake is assuming that an IPTC field, once set, "will somehow survive." It does not. As soon as the file passes through a social platform's processing pipeline, the base marking is often gone. Only the pixel watermark and, depending on the platform, the C2PA manifest survive. Anyone relying exclusively on IPTC risks ending up with no machine-readable marking at all at the point of delivery.

Guide: Writing Metadata and Watermarks Yourself

The theory is well documented everywhere, but a tool that actually writes the labeling into the file has largely been missing. This is exactly where our free AI labeling tool comes in. Here is how you proceed:

1. Upload the file. Drag and drop your AI image or video into the tool. Beforehand, you can use "Check file" to test whether a label is already present.
2. Choose the source type. Determine whether the content is fully AI-generated (trainedAlgorithmicMedia) or only partially AI-modified (compositeWithTrainedAlgorithmicMedia). The tool sets the correct IPTC DigitalSourceType in the XMP block.
3. Add C2PA and a watermark. Optionally enable a C2PA manifest and an invisible watermark to gain tamper resistance and resilience against metadata stripping.
4. Place a visible EU icon. Position the official EU icon on the image or video via drag and drop to additionally satisfy the deployer obligation under paragraph 4.
5. Secure your evidence. The process is documented in an audit-proof way in the labeling register, including a SHA-256 hash, timestamp, and CSV export. This way, you can prove, if needed, that and when you labeled the content.
6. Scale up. For larger volumes, use batch processing (up to 50 files as a ZIP) or the REST API.

The evidence export is the decisive advantage over purely client-side generators: the register, batch processing, and API all run server-side, so you get an audit-proof compliance log.

Compliance Checklist: Machine-Readable Labeling

• Role clarified. Are you the provider or the deployer for this content? (See Provider deployer obligations.)
• Source type correct. trainedAlgorithmicMedia versus compositeWithTrainedAlgorithmicMedia assigned correctly.
• IPTC/XMP set as the universal base marking.
• C2PA manifest added when tamper resistance is needed.
• Invisible watermark embedded when the content passes through platforms that strip metadata.
• Visible labeling added wherever paragraph 4 applies (deepfakes, content of public interest).
• Evidence documented (hash, timestamp, register export).
• Deadlines observed. August 2, 2026 as the core deadline; December 2, 2026 only for legacy systems (paragraph 2, provisional).

You can find a complete, role-spanning version in the AI labeling checklist.

Frequently Asked Questions (FAQ)

What is C2PA / Content Credentials, and do I need it?

C2PA is an open industry standard that attaches provenance and editing data to a file as a cryptographically signed manifest ("Content Credentials"). Whether you strictly need it depends on your role: for providers of synthetic content, such a robust, tamper-proof marking is the safest way to meet Article 50(2). Deployers benefit from it as a strong additional evidence signal.

From when does the machine-readable labeling obligation apply?

The core deadline is August 2, 2026. For systems placed on the market before this date, the machine-readable marking has been postponed to December 2, 2026 under the Digital Omnibus (provisional, as of May 2026). The visible deployer labeling under paragraph 4 is not affected by this.

Does AI labeling even need to be machine-readable?

For providers of synthetic content, Article 50(2) requires a machine-readable marking. Deployers in e-commerce or marketing primarily owe the visible labeling under paragraph 4; for them, the machine-readable layer is voluntary but recommended as evidence of due diligence.

Who is responsible for the machine-readable marking?

The obligation under paragraph 2 applies to the provider of the AI system that generates the content (for example, the provider of the image generator). The deployer, who uses the content, is responsible for the visible disclosure under paragraph 4. In practice, you should cover both levels.

How do I add C2PA or XMP metadata to my image?

The easiest way is through a tool that writes the metadata directly into the file. With our tool, you set the IPTC DigitalSourceType, optionally add a C2PA manifest as well as an invisible watermark, and document the process in the register, without any coding knowledge and in batches too.

What is the difference between a visible label and an invisible watermark?

A visible label, such as the EU icon or the text "AI generated," is aimed at the human viewer and fulfills the deployer obligation under paragraph 4. An invisible watermark is not perceptible to humans, sits in the pixel data, and is read out by software. It serves machine-readable labeling and also survives the removal of metadata.

Does the labeling survive uploading to social media?

That depends on the method. Many platforms remove all metadata during upload, so IPTC/XMP and often the C2PA manifest are lost. An invisible watermark is anchored in the image itself and survives recompression and metadata stripping much better. That is exactly why combining several methods is recommended.

Is the IPTC field alone enough for compliance?

Generally not. IPTC/XMP is universally readable but unsigned and easy to remove. It only conditionally meets the "robust" and "reliable" criteria from Article 50(2). Only the combination with C2PA (tamper resistance) and a watermark (resilience) forms the effective solution required by law.

What happens if labeling is missing or incorrect?

Violations of the transparency obligations under Article 50 can be fined up to 15 million euros or 3% of global annual turnover, whichever amount is higher. There is also the possibility of competition law cease-and-desist notices. More on this under Penalties and cease-and-desist notices for missing AI labeling.

How can I check whether a third-party image carries a C2PA label?

Through a C2PA viewer or the "Check file" function in our tool. Both read out the manifest and show whether a valid, unmanipulated signature is present. If the manifest is missing, it may either never have been set or been stripped during upload; a watermark would then be the more reliable source.

Sources

• EU AI Act, Article 50, Transparency Obligations: https://artificialintelligenceact.eu/article/50/
• European Commission, guidelines on the transparency obligations: https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems
• European Commission, EU icons for labeling AI-generated content: https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content
• EU AI Act, implementation timeline: https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act
• Sidley, EU AI Act Transparency Obligations (compliance by August 2, 2026): https://datamatters.sidley.com/2026/06/24/eu-ai-act-transparency-obligations-preparing-for-compliance-by-2-august-2026/
• Gibson Dunn, EU AI Act Omnibus Agreement: https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/

Further reading: AI labeling, glossary of key terms and the overview Labeling AI images.

This article does not replace legal advice and reflects the status as of July 21, 2026. The Digital Omnibus is still in the legislative process at this time; the postponement rule for legacy systems is provisional. For your specific case, please consult qualified legal advice.