AI Content Disclosure Requirements: The Marketer's Calendar for August 2, 2026

If you run marketing in or into the EU, you have two dates on the calendar and one job between them. Article 50 of the EU AI Act, the transparency block, applies on August 2, 2026, and the machine-readable marking grace period for synthetic content closes on December 2, 2026. That's it. The high-risk obligations got pushed to December 2027 and August 2028, but the rules that touch marketers — chatbot disclosure, synthetic media labeling, deepfake notices — are live this summer, per Gibson Dunn's June 2026 omnibus analysis.

Most of what you're reading on this is either a 4,000-word law-firm alert written for general counsel, or a marketing blog telling you everything is fine. Neither is useful. So here is the operator version: what the ai content disclosure requirements actually cover, what they almost certainly don't, and a calendar you can copy into your project tracker on Monday.

TL;DR

• Article 50 applies August 2, 2026; synthetic-content watermarking grace ends December 2, 2026.

• Triggers: AI chatbots, generated images, AI voice, deepfakes, and AI-written news pieces on public-interest topics.

• Probably untriggered: human-edited AI-assisted blog drafts, ad copy variants, internal briefs.

• Providers of generative systems must mark outputs in a machine-readable format under Article 50(2).

• Action this week: inventory every customer-facing AI surface and assign a disclosure owner.

1. The two dates that matter: August 2 and December 2, 2026

The EU AI Act has a staggered application schedule, which is where most of the confusion comes from. The November 2025 omnibus agreement pushed several high-risk system deadlines to December 2, 2027 and August 2, 2028, but it did not move the Article 50 transparency obligations. Those still apply on August 2, 2026, per Gibson Dunn's summary of the agreement.

Four months later, on December 2, 2026, the grace period for the synthetic-content marking requirement ends. After that date, providers of generative AI systems whose outputs reach the EU market need machine-readable marking in place, not just plans for it.

A marketer reading those two dates should hear two different jobs. August 2 is a disclosure problem: are users told? December 2 is a plumbing problem: is the file itself marked?

The Article 50 timeline marketers actually need to track. High-risk obligations land on a different team.

2. What Article 50 actually covers

The text of Article 50 names four trigger conditions, and they map cleanly onto things marketers do every day.

AI systems that interact with people. If your chatbot, voice agent, or AI receptionist talks to a human, the human has to know they're talking to a machine — unless it's obvious from context. Obvious from context is doing a lot of work in that sentence, and the safe read is: if a reasonable visitor could mistake the agent for a person, disclose.

Synthetic media generation. Images, audio, and video generated or meaningfully manipulated by AI must be detectable as such. This is the watermarking clause, Article 50(2), and it sits on the provider of the generative system, not on you as the deployer. But you still inherit the downstream obligation if you're publishing the asset.

Deepfakes. Anything that resembles a real person, object, place, or event and could mislead a reasonable viewer needs a visible label. This is the strictest tier. A face-swap testimonial video, an AI-cloned celebrity voiceover, a synthesized founder clip — all need a clear notice that the content is artificially generated.

AI-generated text on matters of public interest. If your content addresses topics like elections, public health, or major social debate and is AI-generated without meaningful human editorial review, it needs a disclosure. This is the clause most overread by marketing blogs.

3. What it probably doesn't cover (and where the line blurs)

Here's the part the law firms won't say plainly because their job is to err toward warning. A human-edited AI-assisted blog post — where a marketer used a model to draft, then rewrote, restructured, fact-checked, and published under their byline — is very unlikely to trigger Article 50's text-disclosure clause. The clause carves out content that has undergone human review or editorial control and where a natural or legal person holds editorial responsibility.

In practice, that means most of what we help clients with at Entropy's content marketing practice — AI-assisted research, drafting, ad copy variants, internal briefs, product descriptions — sits outside the text-disclosure trigger as long as a real editor is in the loop and the topic isn't public-interest.

Where does the line blur? Three places worth flagging:

Programmatic landing pages generated at scale with no per-page editorial pass start looking less like "human-edited" and more like "AI-generated text published without review." If you're spinning up 800 location pages a week from a template, get legal eyes on whether your editorial layer counts.

AI-narrated explainer videos sit awkwardly between synthetic media and chatbot interaction. The voice is synthetic, which triggers Article 50(2)'s marking requirement on the provider, and arguably the labeling requirement on you.

AI-generated images in display ads. The provider of the image model handles watermarking under Article 50(2). Your obligation is making sure the watermark survives your compression pipeline, which it often doesn't. We've written separately about why AI detector outputs are unreliable for editorial decisions — the same caveat applies to assuming an upstream watermark is still in the file you ship.

4. The machine-readable marking requirement, in plain language

Article 50(2) reads: providers of generative AI systems must ensure outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated" (official text).

For a marketer, three operator-level translations matter.

First, the obligation is on the provider of the generative system. If you generate images in Midjourney, audio in ElevenLabs, or video in Runway, the marking obligation runs to them. You inherit it by association when you publish.

Second, machine-readable means something a detector or platform can read programmatically — typically C2PA content credentials, SynthID-style invisible watermarks, or metadata embedded at generation time. A visible "Made with AI" caption is a separate, additional requirement for deepfakes and chatbots; it does not satisfy 50(2).

Third, watermarks degrade. Screenshots, re-encodes, social platform compression, and crop-to-square workflows strip metadata routinely. The compliance posture that survives audit is: generate with marking enabled, log the original asset, publish from the marked version, and document the chain.

This is separate from Google's stance on AI content for ranking — which we covered in our piece on whether Google penalizes AI-generated content — and shouldn't be conflated with it.

5. Who's exposed: a self-assessment by asset type

Run every customer-facing AI surface through three questions. Does it interact with a person? Does it produce synthetic media? Does it publish text on matters of public interest without human review? If any answer is yes, you have a disclosure obligation by August 2, 2026.

A quick walkthrough by asset class, based on what we typically see in client audits:

Website chatbot answering pre-sales questions — triggers AI-interaction disclosure. Add a clear notice at session start.

AI receptionist or voice agent handling inbound calls — triggers AI-interaction disclosure, and arguably synthetic media if the voice is cloned or hyper-realistic.

Product images generated by a diffusion model — triggers synthetic media marking on the provider; you should verify the watermark survives your CMS pipeline.

AI-cloned founder voiceover for a brand video — triggers deepfake labeling. Add a visible notice in the first frame and in the description.

Programmatic SEO landing pages generated by an LLM with light templating — likely triggers AI-generated text disclosure if no real editorial review exists per page.

Blog post drafted with an LLM, then edited and fact-checked by a human writer — almost certainly outside Article 50's text trigger, provided the topic isn't election-related or public health.

Ad copy variants generated by an LLM for A/B testing — outside the trigger; these are short-form commercial speech, not public-interest text.

Internal sales enablement docs written with AI assistance — outside scope; Article 50 governs user-facing systems.

6. A compliance calendar you can copy

Here is the calendar we're handing to clients this quarter. Adjust dates to your sprint cadence, not ours.

By July 1, 2026 — inventory every AI surface that touches an EU user. Chatbots, voice agents, generated images, generated video, AI-narrated audio, programmatic pages. Assign one disclosure owner per surface.

By July 20, 2026 — draft the disclosure copy. AI-interaction notice for chatbots. Visible label for deepfakes. Caption pattern for synthetic media in social posts. Get it in front of legal once, not five times.

By August 2, 2026 — Article 50 applies. Disclosures live on every surface in your inventory. Run a same-day audit and screenshot proof for your records.

By October 1, 2026 — confirm with every generative AI vendor in your stack that machine-readable marking will be enabled by their December 2 deadline. Get it in writing. If a vendor can't confirm, start the swap.

By November 15, 2026 — test that watermarks survive your full publishing pipeline end-to-end. Generate, edit, compress, publish, re-fetch, verify the marking is still detectable.

By December 2, 2026 — synthetic content marking grace period closes. Every generated image, audio file, and video you publish should leave your pipeline marked.

Q1 2027 and onward — quarterly audit. New tools enter the stack constantly; the inventory rots faster than you'd think.

The high-risk obligations under the broader Act, now pushed to December 2, 2027 and August 2, 2028 per the omnibus agreement, will land on a different team in your company. Article 50 lands on marketing.

FAQ

When do AI content disclosure requirements take effect under the EU AI Act?

Article 50 transparency obligations apply on August 2, 2026. The machine-readable marking requirement for synthetic content has a grace period that closes on December 2, 2026. Separate high-risk system obligations were postponed in the November 2025 omnibus agreement to December 2, 2027 and August 2, 2028.

Does Article 50 require disclosure on AI-assisted blog posts?

Probably not, when a human editor materially reviews and takes editorial responsibility for the piece. The text-disclosure clause targets AI-generated content published on matters of public interest without meaningful human review. Commercial blog posts edited by a real writer with a byline sit outside the trigger in nearly all cases.

What counts as machine-readable marking for synthetic content?

Article 50(2) requires outputs be marked so a detector can programmatically identify them as artificially generated or manipulated. In practice this means C2PA content credentials, invisible watermarks like SynthID, or embedded provenance metadata. A visible "Made with AI" caption is a separate labeling obligation and does not satisfy the machine-readable clause.

Who is responsible for watermarking generative AI outputs?

The obligation sits on the provider of the generative AI system — the company that builds and offers the model. Marketers using those tools inherit downstream exposure when publishing. The practical compliance posture is to verify your vendors will meet the December 2, 2026 deadline and confirm that marking survives your editing and publishing pipeline.

Do AI chatbots on marketing sites need a disclosure notice?

Yes, from August 2, 2026, AI systems that interact with people must inform users they are interacting with a machine, unless it is obvious from context. For most pre-sales and support chatbots that mimic conversational tone, the safe read is to add a clear notice at session start naming the system as AI.

Pick the most exposed surface in your stack — usually the chatbot or the generated-image pipeline — and assign one owner this week. That's the move. If you want a second set of eyes on your inventory before August, get in touch.