If your product generates images, audio, video or text, two dates on the 2026 calendar should be in your planning: 2 August 2026, when Article 50 of the EU AI Act starts to apply, and 2 December 2026, when the transition period for machine-readable marking of AI-generated content ends. This article explains what each date means and what "labeling" actually requires. It's compliance tooling and documentation, not legal advice.
Two dates, two different things
It's easy to conflate the two deadlines. They're distinct.
- 2 August 2026 — Article 50's transparency obligations begin to apply. This is the general start date for the whole article, including the deepfake and public-interest-text disclosure duties in 50(4).
- 2 December 2026 — the end of the transition period specifically for the machine-readable marking of AI-generated content under 50(2), for systems already on the market. In other words, the *human-facing* disclosure duties land in August; the *machine-readable* marking requirement has a short runway into December.
The May 2026 Digital Omnibus agreement postponed only the high-risk obligations of the AI Act to 2027/28. Article 50 was not moved. These dates stand.
What Article 50(2) requires
Providers of AI systems that generate synthetic audio, image, video or text content must mark the outputs, in a machine-readable format, as artificially generated or manipulated. The emphasis on machine-readable is what makes this different from a visible watermark: the marking has to be detectable by software, not just visible to a person.
There's no single mandated format in the text, but the practical toolkit looks like:
- Embedded metadata — provenance information attached to the file or response.
- HTML meta tags — for content delivered on the web, tags that flag the content as AI-generated.
- JSON-LD — a structured-data block describing the content's origin.
- HTTP headers — server-level signals for API-delivered content.
Industry work on content provenance (such as C2PA) and the voluntary EU Code of Practice on the transparency of AI-generated content point in the same direction: attach durable, machine-readable provenance to what your system produces. The Code of Practice is voluntary — a way to operationalise the duty, not an additional legal requirement.
What Article 50(4) requires
Separately, deployers must:
- Disclose deepfakes — image, audio or video content that has been generated or manipulated to resemble real people, places or events.
- Disclose AI-generated or AI-manipulated text published to inform the public on matters of public interest.
There are narrow exceptions — where the content is part of an evidently artistic, creative, satirical or fictional work, or where the text has undergone human editorial review and a person holds editorial responsibility. These carve-outs are fact-specific, which is why a simple checker can only flag "this may apply" for public-interest text rather than deciding it for you.
A practical timeline
- Now → August 2026: Inventory every system that generates content. For each, decide which duties apply (50(2) marking, 50(4) deepfake/public-interest disclosure). Add human-facing disclosure where required.
- By 2 August 2026: Have your human-facing disclosures live — deepfake labels, public-interest text labels, and any interaction notices from 50(1).
- By 2 December 2026: Have machine-readable marking in place for AI-generated outputs, including for systems that were already on the market.
- Ongoing: Re-check whenever you add a content type or swap a model, and keep a dated record of what you shipped and when.
Common pitfalls
- Assuming a visible watermark is enough. 50(2) is specifically about *machine-readable* marking. A logo in the corner of an image doesn't satisfy it on its own.
- Treating August and December as the same deadline. The human-facing duties don't wait for December.
- Forgetting the evidence. If you can't show *when* you started marking content, you can't easily demonstrate you met the transition. Keep a log.
How the marking formats fit together
You don't have to pick one format. The three signals serve different consumers and are complementary:
- HTML meta tags are read by browsers, crawlers and preview systems when your content is a web page. They're the lowest-effort way to flag a generated page.
- JSON-LD is structured data that search engines and other tools already parse. Embedding an AI-provenance block alongside your existing structured data means the signal travels with the page.
- HTTP headers work for content delivered through an API or CDN, where there's no HTML to annotate — the marking rides on the response itself.
Emitting all three maximises the chance that whatever system encounters your content can detect the marking, which is the point of "machine-readable".
Where C2PA fits
You'll hear C2PA (the Coalition for Content Provenance and Authenticity) mentioned alongside AI content labeling. C2PA is a standard for cryptographically signed content provenance — think of it as a tamper-evident manifest embedded in a media file. It's a strong tool for durable provenance, especially for images and video, and it aligns with the direction Article 50(2) points in. It's also more involved to implement than metadata tags. The pragmatic path for most teams: start with the lightweight machine-readable signals now, and treat signed provenance like C2PA as a later, deeper investment where it matters (for example, high-stakes media). Article 50(2) asks for machine-readable marking; it does not mandate C2PA specifically.
Frequently asked questions
Does a visible "AI-generated" caption count? A visible caption is good practice and helps humans, but 50(2) specifically calls for a *machine-readable* format. A caption alone isn't machine-readable in the required sense — pair it with metadata.
What about content generated before August 2026? The marking duty is about outputs your system produces. For systems already on the market, the machine-readable marking requirement has its transition until 2 December 2026. Focus your effort on making sure new outputs are marked and that your generation pipeline applies marking going forward.
We generate text, not images — does 50(2) still apply? Yes. Synthetic text is explicitly within scope. The same machine-readable marking approach applies; for web-delivered text, meta tags and JSON-LD are the natural fit.
Who's responsible — us or the model provider? As with the rest of Article 50, the duty follows the system and its intended purpose. If your product is the one generating and delivering the content to users, the marking duty is yours.
A quick recap of the two dates
To keep them straight: 2 August 2026 is when Article 50 starts to apply, including the human-facing disclosure duties for deepfakes and public-interest text. 2 December 2026 is the end of the transition window for the *machine-readable* marking of AI-generated content under 50(2), for systems already on the market. New systems should build marking in from the start. If you remember nothing else: human-facing labels first, machine-readable marking shortly after, and a dated record of both.
How DiscloseKit helps
DiscloseKit generates the machine-readable marking snippets for you — HTML meta tags, a JSON-LD block, and HTTP header examples — per product and per content type, with a marking API for automating it at scale. Every snippet generated and every disclosure shown is written to an append-only evidence log with a timestamp, so the "when did you start?" question has a clean answer. And the free checker maps your content types to the exact obligations and deadlines above.
Run the check to see which of the August and December duties apply to your product.