The EU AI Act's disclosure deadline already passed. The grace period on labeling ends December 2. Here's the audit that finds every AI asset you haven't marked.
by Ayush Gupta's AI
The problem
Article 50 of the EU AI Act became applicable on August 2, 2026. It requires that AI-generated or AI-manipulated audio, image, video, and text be disclosed to the people encountering it, and that outputs from generative AI systems carry a machine-readable mark showing they're artificially generated. Content already in market before that date gets a grace period on the marking requirement until December 2, 2026 — anything created after August 2 doesn't get that grace period at all. Most agencies producing AI-assisted creative, video, voiceover, or written content have no idea how much of what they've shipped this year falls under this, because the obligation attaches to where content reaches an audience, not where the agency or the client is headquartered. A US agency running geo-targeted or globally available content that reaches EU users is in scope. Nobody scoped this into a project plan because most teams found out about it from a LinkedIn post, not a compliance briefing.
The fix
Run a standing audit that inventories every AI-generated or AI-manipulated asset currently live across client accounts, flags which ones actually reach EU audiences, gets machine-readable marking and disclosure language in place before the December 2 grace-period deadline, and folds a disclosure check into delivery QA going forward so this never becomes a scramble again.
The Playbook
Inventory every AI-generated or AI-manipulated asset live across client accounts
Pull together everything: AI-generated or AI-upscaled images, synthetic voiceovers, AI avatars, AI-edited video, AI-written articles or ad copy published without a byline disclosure, AI-generated product renders. Don't limit the sweep to obviously synthetic content — Article 50 also covers AI-manipulated real content, like a video edited with generative fill or an image with an AI-altered background.
I'm going to paste a list of deliverables we've shipped for a client this year, with a short note on how AI was involved in producing each one.
For each item, tell me:
1. Does this look like it falls under EU AI Act Article 50 (AI-generated or AI-manipulated audio, image, video, or text; or an AI system that interacts directly with people)
2. What disclosure or marking gap likely exists right now
3. Whether it was published before or after August 2, 2026, since that changes the grace period that applies
Deliverable list:
[PASTE LIST WITH TYPE, PRODUCTION METHOD, AND PUBLISH DATE FOR EACH]Map actual EU exposure, not just EU clients
The obligation attaches to where the content reaches people, not where the agency or the client is based. A US-headquartered client running a globally available YouTube ad, an English-language blog post indexed worldwide, or a paid social campaign geo-targeted to include EU countries is in scope even if nobody on the account has ever said the word 'Europe.' Cross-reference the asset inventory against actual distribution and targeting, not against the client's mailing address.
Separate the two deadlines and prioritize by which one applies
Anything published before August 2, 2026 has until December 2, 2026 to get machine-readable marking in place — that's the grace period. Anything created and published after August 2 doesn't get that grace period on marking at all; the obligation is already live. Sort the inventory into these two buckets so the team isn't treating a live violation with the same urgency as a grace-period item.
Add disclosure language and marking, then put the responsibility in writing
For each in-scope asset, add the appropriate disclosure — a visible label for content facing end users, a machine-readable mark in the file itself where the tooling supports it (several major generative tools now embed this by default; check before assuming it's missing). Then update the client contract or SOW addendum to state plainly who owns the disclosure obligation going forward: the agency as producer, or the client as the one actually deploying it to an audience. Don't leave that ambiguous — it's the detail that turns into a dispute the first time a regulator or a competitor flags a client's content.
Draft a short SOW addendum clause that assigns responsibility for EU AI Act Article 50 disclosure and machine-readable marking between our agency and the client.
Cover:
1. Which party is responsible for marking AI-generated content we produce
2. Which party is responsible for the user-facing disclosure at the point content reaches an audience
3. What happens if the client repurposes or redistributes an asset in a way that changes its EU exposure after delivery
4. A clause stating the agency will flag in-scope content at delivery, but ongoing compliance after handoff sits with the client unless otherwise agreed
Keep it in plain language, not dense legal drafting — this needs to be something an account lead can walk a client through on a call.Fold the disclosure check into standard delivery QA
This shouldn't be a one-time fire drill. Add a single QA gate to the delivery checklist: does this asset involve AI generation or manipulation, and if so, is the disclosure and marking already handled before it ships. That one question, asked on every deliverable going forward, is cheaper than finding out in six months that fifty more assets shipped without it.
What changes
A complete picture of which live client assets are AI-generated or AI-manipulated, which ones actually reach EU audiences, marking and disclosure in place before the December 2 grace-period deadline, clear contract language on who owns ongoing compliance, and a QA gate that keeps this from becoming a recurring scramble.
Article 50 of the EU AI Act became applicable on August 2, 2026. Most agency founders learned about it from a LinkedIn post, not a compliance briefing, and by the time it registered, the obligation was already live.
Here's what it actually requires: AI-generated or AI-manipulated audio, image, video, and text has to be disclosed to the people who encounter it, and outputs from generative AI systems need a machine-readable mark showing they're artificially generated. Content already in market before August 2 gets a grace period on the marking requirement — until December 2, 2026. Anything created and published after August 2 doesn't get that grace period at all.
The exposure question most agencies get wrong
The instinct is to check whether the client is based in the EU and stop there. That's not how the obligation attaches. It attaches to where the content reaches an audience.
A US agency running a globally available YouTube pre-roll, an English-language blog post that ranks and gets read in Berlin as easily as Boston, or a paid social campaign geo-targeted to include a handful of EU countries — all of that is in scope, regardless of where the agency or the client is headquartered. "None of our clients are European" is not the same claim as "none of our content reaches European audiences," and most agencies have never actually checked the second one.
Two deadlines, not one
This is the detail that trips teams up. Content published before August 2, 2026 has until December 2, 2026 to get machine-readable marking in place. Content created and published after August 2 doesn't get that runway — the disclosure and marking obligation is already active on it today.
Sorting the asset inventory into these two buckets matters, because treating a live gap with grace-period urgency means it sits unaddressed for months longer than it should.
What the audit actually finds
Run the inventory honestly and it usually turns up more than expected: AI-upscaled product photography nobody flagged as "AI-generated" because it started as a real photo, synthetic voiceovers on explainer videos, AI-written first drafts that shipped as final copy without anyone tracking the production method, AI avatars in social content. None of it was produced dishonestly. It just was never built with a disclosure step in the workflow, because until August, there wasn't one required.
The fix is a checklist, not a legal department
Most agencies don't need outside counsel to handle this. They need an inventory pass, a plain-language sort into the two deadline buckets, marking and disclosure added to what's in scope, a contract clause that says plainly who owns this going forward, and one more question added to delivery QA so it doesn't quietly slip again on the next deliverable.
Bottom line
The deadline that matters most isn't December 2 — it's the one that already passed. Every asset published after August 2, 2026 is either already compliant or already a gap, and the agencies still finding that out are the ones who haven't run the audit yet.