How to Comply with EU AI Act Article 50

On August 2, 2026, the transparency obligations in Article 50 of the EU AI Act became enforceable across all 27 member states. For the first time in any G7 jurisdiction, a company that fails to tell users they are talking to a chatbot, or that fails to mark AI-generated content, faces fines of up to €15 million or 3% of global annual turnover, whichever is higher. The deadline is real. Anthropic activated invisible watermarking across all of Claude’s output worldwide in the same week, and the European Commission’s final Guidelines on Article 50 were released on July 20, 2026, just thirteen days before enforcement began. For any team that generates images, video, audio, or text and serves EU users, the question is no longer whether to disclose, but what minimum implementation actually meets the requirements.

Key Takeaways

• Article 50 sets four separate obligations: chatbot disclosure, machine-readable marking, deepfake labeling, and emotion/biometric notification, each with a different responsible party.
• The Code of Practice requires at least two active marking methods at the same time because no single watermarking technique meets all four legal criteria of being effective, interoperable, solid, and reliable.
• C2PA signed metadata alone can be removed by a single screenshot, so the Code pairs it with an imperceptible in-pixel watermark like SynthID.
• Systems already on the EU market before August 2, 2026 have a grace period on the marking requirement until December 2, 2026; new systems must comply immediately.
• Signing the Code of Practice provides a presumption of conformity, and about 190 organizations had signed by the end of July 2026.

What Article 50 Actually Requires

Article 50 includes four distinct transparency obligations, with different responsible parties for each. This causes the most common compliance errors, since the same company can be a “provider” for one product and a “deployer” for another.

The first obligation, under Article 50(1), applies to providers of AI systems designed to interact directly with people. Chatbots, voice assistants, and autonomous agents must inform users they are interacting with AI at the first point of contact, unless it is clear from the situation. The Commission’s draft guidelines, published May 8, 2026, narrowed the “obvious” exemption significantly: only interactions where a reasonably well-informed, observant person would clearly recognize AI involvement are exempt. Most consumer-facing deployments do not qualify.

The second, under Article 50(2), requires the most technical effort. Providers of generative AI systems must embed machine-readable markings in AI-generated audio, images, video, and text, making them identifiable as artificially generated. This is the watermarking requirement, and it received a targeted deferral under the May 2026 Digital Omnibus: systems already on the EU market before August 2, 2026 have until December 2, 2026 to comply, while systems launched on or after August 2 must comply immediately.

The third obligation, under Article 50(4), applies to deployers. Anyone publishing AI-generated or manipulated content depicting real or realistic people, objects, or events must visibly disclose that it is a deepfake, regardless of deceptive intent. The fourth, under Article 50(3), requires deployers to inform people when they are exposed to emotion recognition or biometric categorization systems.

According to Al Jazeera’s breakdown of what came into force, these provisions require disclosure, not a ban on the technology, and violations can lead to administrative fines of up to €15 million or 3% of global annual turnover.

The Detectability Question: Is C2PA Enough?

The short answer is no. C2PA Content Credentials alone do not meet Article 50(2), and the reason is stated in the Code of Practice itself.

The Act requires that machine-readable markings be effective, interoperable, solid, and reliable, four criteria that must be met simultaneously. The Code of Practice published on June 10, 2026 acknowledges that no single active marking technique currently satisfies all four at once. C2PA metadata is cryptographically signed and interoperable, but it is removed when content is captured by a screenshot, uploaded to social media, or converted to another file format. An imperceptible watermark embedded in pixels or tokens survives recompression but weakens under cropping and adversarial manipulation.

The result is a required multi-layer approach. Under the Code, providers must implement at least two active layers of machine-readable marking at the same time: C2PA signed metadata plus an imperceptible watermark. Legal analysis firm Bird & Bird called this a “crucial reality check” for providers, noting that the search for a single technical solution to satisfy Article 50(2) may be over, as reported by TechTimes.

The Commission’s guidelines add another point that small teams often overlook: technical feasibility under Article 50(2) is an “objective notion” not dependent on individual providers’ resources. A startup cannot claim exemption based on cost. Midjourney, which had not implemented C2PA in its outputs as of the Code’s publication, was identified as facing direct regulatory risk under this obligation.

The verification process is where the two layers differ in practice. C2PA credentials are verified with open-source tools such as libc2pa, the Rust SDK maintained by the Content Authenticity Initiative, which also offers a public web verification tool. The imperceptible watermark is verified through a detection API such as Google’s SynthID, integrated into Search, Chrome, and Pixel devices. A visible label or disclaimer does not fulfill the marking requirement, because Article 50(2) specifically requires automated detection without human interpretation.

How OpenAI, Google, and Adobe Answered

The three largest companies took different approaches to meet the same goal, and the differences reveal what the Commission expects.

Google has the most extensive in-house system. SynthID embeds an imperceptible pattern directly into pixels, audio samples, and text tokens across Imagen, Veo, Lyria, and Gemini. It is the most widely used imperceptible watermarking method, and Google has also open-sourced SynthID-Text so third-party model developers can integrate it. In May 2026, Google added SynthID verification directly into Search and Chrome, allowing users to check synthetic images via Google Lens.

OpenAI chose the interoperability approach. In May 2026 it announced embedding Google’s SynthID watermarks into ChatGPT image outputs, layering them on top of C2PA Content Credentials already attached to DALL-E and Sora generations, and previewing a public verification portal. This is notable because OpenAI had abandoned its own watermarking plans in September 2024 after a survey found almost 30% of ChatGPT users would use it less if watermarking were added. The EU obligation changed that calculation.

Adobe relies on the C2PA side. Firefly signs every output with C2PA credentials by default, and Adobe is the founding backer of the C2PA standard itself. The company’s approach depends heavily on signed metadata rather than in-pixel watermarking, which is why Adobe’s compliance under Article 50(2)’s two-layer requirement is under ongoing review, though the company has not publicly described a SynthID-class watermark for Firefly.

Anthropic, while not one of the original three, provides a clear example. As Search Engine Journal documented, Anthropic signed the Article 50(2) Code of Practice and now embeds invisible text watermarks in Claude output plus C2PA signed metadata for generated .svg, .png, and .jpg files, applied worldwide rather than only in the EU. The company’s own limitations section is frank: a detected mark indicates Claude may have processed content, not that Claude authored it, and proofreading, translation, or summarization can produce a mark on human-written text.

Minimum Implementation by Modality

For a product team releasing an image, video, or text generator in 2026, the minimum compliance path is now clear. The table below matches each modality to the two required layers and verification tool, based on the Code of Practice and vendor implementations described above.

Text is the weakest part of this system, and product teams should plan accordingly. SynthID-Text works by biasing token selection so a detector can later identify the pattern, but paraphrasing through a second model erases the statistical signal. Anthropic’s watermark travels with copied text and may survive some editing, but the company has not published its detection method, and independent researchers such as GPTZero’s Alex Cui have argued that text watermarks can be defeated by heavy paraphrasing. For any product where text provenance matters, treat watermark as one signal among several, not a guarantee.

The practical steps for a team building this in 2026 are clear. Sign the Code of Practice first; it was open for signature until July 22, 2026 for the initial list, and it shifts the evidentiary burden toward regulators. Then embed C2PA credentials at generation time using libc2pa or Content Authenticity Initiative libraries. Then add the imperceptible watermark, either by adopting SynthID through a partner model or by integrating an open-source method. Finally, connect verification into your own upload and moderation pipeline so you can document, with logs, that a given piece of content was marked and that the mark remained.

Labeling UI Requirements

The marking requirement under Article 50(2) concerns machine-readable detection. The labeling requirement under Article 50(4) concerns what a human sees, and the two are often confused. They are separate obligations with separate responsible parties.

The Code of Practice introduced standardized icons for labeling AI-generated content: “AI” in English, with localized versions including “KI” in German and “IA” in French, provided without attribution requirements. It specifies placement rules by modality: persistent on-screen labels for video, visible markings for images, and audible disclaimers for audio. It also distinguishes between “fully AI-generated” and “AI-assisted” content, each with different disclosure requirements.

Engadget’s report on the new rules clarifies the visible-label requirement further. If a machine helped create “authentic-looking” deepfake content, the content must carry an “AI” mark. If image, audio, video, or text is generated entirely by AI, it must carry an “AI-Generated” label. The Commission’s examples include fully AI-generated deepfake videos with politicians or fictional events, fully AI-composed music or art, and AI-generated news summaries.

Two exemptions apply to real products. Personal content like group chats is exempt, as are “evidently artistic,” satirical, and fictional works. Content generated before August 2, 2026 does not have to be labeled retroactively, though the Commission encourages voluntary labeling of older material where possible.

The evidence requirement is the part most teams under-invest in. A working notice is not enough; as CMSWire’s coverage of customer experience notes, teams need logged proof that disclosure appeared in a specific customer’s actual interaction, not just that it exists somewhere in the product. A vendor’s claim that its platform “supports AI Act compliance” does not relieve the deployer’s obligation to verify disclosure appears correctly in its own customer journey.

Limitations and Trade-offs

The realistic view on Article 50 compliance in 2026 is that it sets a minimum standard, not a maximum, and the minimum has known gaps. Independent experts have documented where the two-layer approach fails, and these issues are practical.

The screenshot attack is the cheapest and most common. IEEE Spectrum showed in 2024 that a single screenshot removes both C2PA metadata and IPTC watermarks from AI images, defeating detection in about two seconds. The same vulnerability remains two years later: Reuters’ July 2026 test found Meta’s Muse Image detector missed 55% of its own cropped images. The crop-and-recompress vulnerability that Meta highlighted in its 2024 threat model remains open, as we covered in our analysis of Meta’s watermarking risks.

Unsecured open-source models create a permanent pool of unmarked content. Stability AI has not integrated SynthID into Stable Diffusion, and Meta’s Llama models do not embed SynthID-Text by default. Any provenance strategy relying solely on watermarking inherits that coverage gap, because older unmarked models remain available.

There is also a subtle integrity risk that most verification pipelines miss. A CVPR 2026 workshop paper described what it calls Integrity Clash: a condition where a digital asset carries a cryptographically valid C2PA manifest asserting human authorship while its pixels simultaneously carry a watermark identifying it as AI-generated, with both signals passing their checks independently. Any verification process that checks C2PA and watermarking separately, without comparing results, is vulnerable. We explored this in detail in our guide to verifying AI model outputs.

The regulation does not require that marking survive every attack. Compliance and security are related but not the same. A platform that deploys C2PA plus SynthID detection complies with Article 50(2). It is not protected from all attacks. The distinction matters because regulators will ask about compliance, but users will ask about trust, and trust breaks when an attacker succeeds. For a broader view of how these technologies fit within the 2026 regulatory environment, see our guide to AI watermarking and provenance.

For security or platform engineers, the practical advice is to design defenses against the attacks you can block, not those you cannot. The screenshotter and metadata remover can be stopped with C2PA plus watermarking. The regenerator, who passes AI output through their own model, defeats all current defenses at once. Multiple pieces of evidence, verified provenance, checked embedded marks, and a retained “inconclusive” category is the standard in production, and no negative result can prove content came from a camera.

Related Reading

More in-depth coverage from this blog on closely related topics:

• Meta AI Content Watermarking Risks in 2026
• Best Hardware for Large AI Models
• iOS 26.5 update: new features, bugs
• Choosing the Best Local AI Inference Tools
• Tailscale Peer Relay Feature Guide

Sources and References

Sources cited while researching and writing this article:

• What came into force with the EU’s AI Act this week – and what didn’t
• EU AI Act Chatbot Disclosure and Deepfake Labeling: July 22 Signatory Deadline
• Anthropic To Mark Claude Text & Files Under EU AI Act Code
• EU announces new rules on AI transparency
• Does Your Chatbot Meet EU AI Act Disclosure Rules?