EU AI Act Watermarking: What Content Businesses Need to Check in 2026

As of August 2026, watermarking is no longer a conference slide. Article 50 of Regulation (EU) 2024/1689 — the EU AI Act — applies from 2 August 2026. The Commission’s guidelines on those transparency duties and the voluntary Code of Practice on Transparency of AI-generated Content are the pages to keep open. A 2 August 2026 thread from @AndrewCurran_ treated the date as the start of mandatory labelling and text detectability. Treat social posts as a pointer; the legal text is the source.

Two duties get mixed up in sales calls. Providers of generative systems must mark synthetic audio, image, video, or text in a machine-readable way and make it detectable (Article 50(2)), unless the system is only doing standard editing or does not substantially change the input. Deployers must disclose deepfakes and must label AI-generated or manipulated text published to inform the public on matters of public interest, unless a human reviewed it and someone takes editorial responsibility (Article 50(4)). If you use the system under your own authority, you may be the deployer. If a client legal entity controls your use, the client may remain the deployer. A provider may embed a machine-readable signal, but edits can affect detectability. Small teams should record who controls the system, reviews the output, and holds final publication responsibility.

Do not assume your location or client list settles the question. Scope can depend on the EU market, an EU deployer, or output used in the EU. A vendor may apply marking more broadly, so check its policy rather than assuming every export contains a signal or survives editing. The Commission has also said systems already on the market before 2 August 2026 may have a short marking grace period (widely described as until 2 December 2026 — confirm on the live guidelines). Content published before 2 August 2026 is not, on the Commission’s own factsheet, required to be marked retroactively.

Four places it hits a freelancer

Direct content sales. A blog pack, newsletter, or product page drafted in a covered system may contain a machine-readable signal, but editing can affect detection. That is not the same as a public “AI wrote this” badge. Article 50(4)’s visible label is aimed at public-interest text without human review, not every marketing paragraph. Still: a buyer who runs a detector, or a platform that starts checking, may detect a provider mark you never mentioned. Put the tool, the human pass, and who owns the claims in the SOW.

SaaS or social-account retainers. If you post as the brand, you are closer to a deployer of the output the public sees. Deepfake-style images or cloned voice need a clear label under Article 50(4). A weekly LinkedIn calendar of lightly edited AI drafts is a different risk: the Act’s public-interest test may not fit a product launch post, but the platform or the client’s counsel might still want a disclosure. Do not strip marks or sell a “de-watermark” script. The Code of Practice tells providers not to offer circumvention tools.

Your own site or newsletter. You are the publisher. If the piece is newsy and you did not actually edit it, the Article 50(4) exception for human review and editorial responsibility is the one that matters. A reviewed, sourced article with your name on it is a different object from a raw model dump. Keep the review trail.

Code and documentation. A public README or docs page may be covered as text, but the Commission places source code outside Article 50(2)’s marking obligation. Do not assume model-written code comments are covered. Intermediate model “thoughts” that nobody is meant to read are outside the Commission’s description of synthetic content people perceive. Confirm current guidance if you ship model output straight into a public docs site.

What you should actually do

1. Ask the vendor, in writing, whether outputs are marked and how a client can detect them. Link their trust or transparency page in the reply. If they cannot say, do not promise “undetectable.”
2. Disclose the stack in the contract: which tools, what you review, what you will not strip. A one-line “AI-assisted, human-edited, freelancer remains responsible” clause beats a surprise detector screenshot.
3. Keep a human pass that is real. The exception is editorial control, not a glance. Verify, qualify, or remove claims you cannot source. Leave a comment or checklist.
4. Do not run “remove watermark” utilities on client work. If a client asks, refuse and document the request. This is risk control: Article 50(2) places the duty on the provider, not a separate preservation duty on freelancers or clients.
5. Separate public-interest publishing from ordinary commercial copy. If you write explainers that look like news for an EU audience, assume Article 50(4) until counsel says otherwise.

Who should worry, and who can wait

High impact: people selling unreviewed word-count packs into the EU; affiliate or news-like sites; anyone shipping deepfake-style video or cloned voice; retainers that post as the client without a named editor.

Lower impact: a specialist who interviews the client, cites sources, and signs the piece; internal drafts that never publish; code comments in a private repo; work that is only light edit of the client’s text (closer to the standard-editing exception — confirm on the facts, do not self-certify).

Geography alone does not settle scope. Check where the system is offered, who controls it, and where its output is used. A vendor may mark more broadly, while your duties depend on role and use.

A five-minute check before you send

1. Open the vendor’s current transparency or detection page. Note the date.
2. If the vendor offers an official detector, treat its result as a limited spot check. One paragraph cannot prove that the delivery is marked or unmarked.
3. Confirm the file you will send still matches what you reviewed, not an earlier raw export.
4. Add one sentence to the delivery note: tool name, “human-edited,” “marks not removed.”
5. If the piece is public-interest-shaped and the audience is in the EU, record the substantive review and the person or entity ultimately responsible. Add a visible disclosure or seek legal advice when required.

This is not legal advice. It is the minimum a freelancer or small content team can do in August 2026 without pretending the watermark is someone else’s problem. The Act’s official pages will move; reread them when a client asks, and do not argue from a screenshot of a thread.