Since August 2, 2026, the EU enforces labeling AI generated content under Article 50 of the AI Act. The obligation doesn't stop at the big AI vendors. It reaches anyone who publishes AI-made images, audio, video, or text to an EU audience. I spent the past two weeks rewriting disclosure captions across my own test accounts to see exactly what a compliant label has to say, and this guide is the checklist that came out of it.
Short answer: As of August 2, 2026, the EU enforces labeling AI generated content reaching EU users under AI Act Article 50. Providers must mark synthetic output in a machine-readable format; deployers (that's most creators and businesses) must clearly disclose deepfakes and unreviewed AI text on public-interest topics. Fines reach €15 million or 3% of global turnover.
I already covered the legal mechanics of Article 50 — who it covers, the two enforcement dates, the deepfake definition — in my full write-up on the EU’s AI content labeling rules. This guide skips the legal reading and gets straight to the part most people actually need: what to write, where to put it, and which tools do part of the work for you.
What you'll need
You don't need a compliance officer for this. You need three things: the AI tool or platform you're publishing from, a place to put a caption or on-screen text (a social post, a video description, an article byline), and roughly ten minutes per content type to write your disclosure once and reuse it. If you're publishing on your own site, check whether your CMS strips image metadata on upload — most do, which is exactly why the visible caption matters more than the invisible watermark. Nothing here requires a paid tool, though a few make the process faster, which I cover further down.
Step-by-step: how to label AI-generated content correctly
1. Sort your content into two buckets
Article 50 treats "this was made with AI" differently from "this looks like a real person, place, or event but isn't." A cartoon illustration made with Midjourney needs a light disclosure. A photorealistic video of a real politician saying something they never said is a deepfake, and it needs a clear, hard-to-miss one. Decide which bucket each piece falls into before you write anything.
2. Check whether the tool already labels the file for you
When I tested ChatGPT, DALL·E, and Gemini image outputs this month, all three embedded C2PA Content Credentials metadata by default. You can confirm this yourself at Content Credentials’ verification page by uploading the original file — the coalition behind the standard lists OpenAI, Google, Microsoft, and Adobe among its core members. That metadata satisfies the provider's obligation under Article 50(2). It does not satisfy yours as a deployer if what you're publishing is a deepfake or unreviewed AI text on a public-interest topic. That disclosure is separate, and it's on you.
2a. Know the grace period, if you built on an older system
Not every system had to comply on day one. Under the European Commission’s Article 50 guidance, tools already on the market before August 2, 2026 get until December 2, 2026 to finish the technical marking requirement specifically. That grace period covers the marking mechanic only — a deployer's duty to disclose deepfakes and unreviewed AI text on public-interest topics still applies from August 2, 2026, regardless of which underlying model or platform you used.
3. Write a caption disclosure, not just metadata
Metadata is invisible to a reader scrolling their feed. Put a short, plain-language line where a person will actually see it: under the image, in the video description, or as the first line of an article byline. I go through exact wording in the next section.
4. Add it before you publish, not after
I tested this the hard way on a scheduled post: adding the disclosure line after a post is already live doesn't undo the exposure window where it ran undisclosed. Build the caption into your publishing template so it's there from the first view.
5. Re-check anything you plan to re-share
When I ran a Gemini-generated image through a screenshot-and-repost cycle — the exact way most people share images on social platforms — the embedded metadata didn't survive. My caption did, because I'd typed it as text rather than relying on the file. Treat every re-upload as a fresh publish that needs its own visible disclosure.
Example disclosure text you can copy
These are the lines I've been using across the posts I ran while testing this month. Swap the bracketed tool name and adjust to your platform's character limits.
- AI image, non-deepfake: "Created with [tool name] AI image generation."
- AI video or image resembling a real person, place, or event (deepfake disclosure): "This was created or altered with AI and does not depict a real event."
- AI voice/audio: "This audio was generated using AI voice synthesis ([tool name])."
- AI-written article, no human editorial review: "This article was drafted with AI assistance and has not been reviewed by a human editor." (If a human editor did review it, Article 50's text-disclosure requirement doesn't apply — say so instead: "Drafted with AI assistance, reviewed and edited by [name].")
- Obvious satire/art exemption: A one-line "AI-generated satire" note is enough — Article 50 allows a lighter disclosure for content that's evidently artistic, fictional, or satirical.
Common mistakes to avoid
The mistake I see most often is assuming a tool's built-in watermark covers you. It covers the provider, not you as the deployer, and those are two different obligations under Article 50.
Second is burying the disclosure in a bio or terms page. The European Commission’s own guidance calls for labels to be "clearly and distinguishably" presented at the point someone encounters the content. A link three clicks away doesn't count.
Third is skipping disclosure on re-shares because "the original already had a watermark." I tested this directly: metadata rarely survives a screenshot or re-encode, so your caption is often the only disclosure left standing after a repost.
Fourth is over-labeling low-stakes content and under-labeling deepfakes. A stylized AI illustration and a synthetic video of a real person don't need the same weight of disclosure — treating them the same either annoys your audience or under-protects you.
Comparison: what needs a label, and how visible it has to be
| Content type | Who's on the hook | Disclosure needed | How visible |
|---|---|---|---|
| AI illustration/art, no real person | Provider (metadata) | Optional but encouraged | Low — metadata is enough |
| AI product photo, stylized, non-deepfake | Deployer | Light disclosure | Caption or alt text |
| Deepfake image/video of a real person, place, or event | Deployer | Mandatory, clear disclosure | High — visible at first exposure |
| AI voice clone / synthetic audio | Deployer | Mandatory | Audible or in description |
| AI-written text on public-interest topics, unreviewed | Deployer | Mandatory | Clear, near the byline |
| AI-written text, human-reviewed | N/A | Not required by Article 50 | — |
| Satire, art, fiction (evidently so) | Deployer | Lighter disclosure | Low |
Tools that make this easier
A few tools take some of the manual work out of this. My ChatGPT image generation guide covers how OpenAI's built-in Content Credentials show up in exported files. That's worth knowing before you assume you need a separate watermarking step. Canva’s AI tools also attach Content Credentials automatically to AI-generated design elements. If you're already building graphics there, you likely have provider-side metadata without extra setup. For checking whether text reads as AI-written before you decide it needs a disclosure line, my guide to spotting AI writing walks through the same detection signals I use when reviewing drafts. I track which vendors keep their compliance claims honest under testing on my AI tool ratings and AI tool reviews pages. I log regulatory changes like this one as they happen on AISagely’s AI news hub.
Frequently Asked Questions
Does the EU enforcing labeling of AI generated content apply outside the EU?
Yes. The obligation follows your audience, not your address. If your AI-generated content reaches someone in the EU, Article 50 applies to you regardless of where you or your company are based.
Is a tool's built-in watermark enough to comply?
Not on its own. Provider-side metadata satisfies the provider's obligation, not a deployer's separate duty to disclose deepfakes or unreviewed AI text on public-interest topics. You still need your own visible disclosure for those.
Do I need to label AI content I posted before August 2, 2026?
No. Per the European Commission’s guidance, content generated before the applicability date doesn't need retroactive labeling, though labeling it voluntarily is encouraged.
What happens if I don't disclose a deepfake?
Fines can reach €15 million or 3% of global annual turnover, whichever is higher, per the same Commission guidance, with proportionality considerations for smaller entities. Enforcement runs through national market surveillance authorities.
Is the easiest way to comply just adding a caption?
For most creators, yes. A one-line, clearly visible caption in plain language covers the deployer-side disclosure requirement for the vast majority of everyday AI content — you only need the heavier deepfake-specific wording when the content resembles a real person, place, or event.