New AI transparency rules: What they mean for content teams using Sulu.ai

• The EU AI Act's transparency rules apply from 2 August 2026.
• Most everyday AI tasks—translation, alt text, grammar polishing and SEO metadata—remain out of scope.
• Human review and clear editorial responsibility are the key to most disclosure exceptions.
• AI-powered user interactions and deepfakes require additional transparency.
• Sulu.ai's structured editorial workflows already align well with these requirements.

The AI Act sorts AI systems by risk, and Article 50 is the slice about transparency. 
Stripped to the essentials, Article 50 asks for honesty in a few specific situations:

• Tell people when they are dealing with an AI, for example a chatbot on your site.
• The outputs of generative AI, meaning synthetic image, audio, video, or text, have to be markable in a machine-readable way as artificial.
• If you publish a deep fake, or AI-generated text that informs the public on matters of public interest, you disclose it.
• Standard editing is carved out, and so is content that has gone through genuine editorial review.

Providers or deployers: which are you?

One distinction runs through all of it. The law separates providers, who build and place an AI system on the market, from deployers, who use it. If you run a content team, you are almost always the deployer. This matters, because it decides which duties are yours and which belong to the vendor behind your tools. The heaviest technical one, marking generated content so it is detectable as artificial, sits with the provider, not with you.

The rules apply from 2 August 2026. Content you generated before then does not need labelling retroactively, though republishing it after that date does. Providers of generative systems that already exist get until 2 December 2026 to bring their marking into line. And because these guidelines are non-binding, expect them to be refined as real enforcement experience builds.

The practical shape is reassuring, because most everyday AI work is simply out of scope. The guidelines carve out short outputs like captions, alt-text, and UI labels, AI translations of text, grammar correction and light stylistic polishing that does not change meaning, and minor image adjustments like cropping or colour correction. So generating alt-text for accessibility, translating a page, or letting an assistant tighten your grammar does not, on its own, create a labelling duty. That covers a large share of what content teams reach for AI to do.

Where you do have duties

Two situations do put real obligations on you.

The first is deep fakes. If you publish AI-generated or manipulated image, audio, or video that could plausibly pass as authentic, you must disclose that it is artificial, with a marker a person can actually notice, not a hidden metadata tag. A synthetic avatar of your CEO addressing customers needs it. An AI product shot that makes the real product look better than it is can qualify too. Content that is obviously unreal, or clearly artistic and satirical, gets a lighter touch.

The second is text on matters of public interest. AI-generated or manipulated text meant to inform the public about politics, public health, consumer safety, the environment, or financial and scientific topics open to debate needs a disclosure. Worth knowing for marketing teams: ordinary advertising copy and product descriptions are generally not caught, unless they make claims about health, consumer safety, or sustainability.

The exception that rewards good editing

Here is the part that matters most, and it is genuinely good news. The disclosure duty for public-interest text falls away when two things are true: the text has gone through genuine human review, and a named person or organisation holds editorial responsibility for it. Read that again, because it is essentially a description of a healthy editorial workflow. The catch is that the review has to be real. Fact-checking the substance counts; a cursory spellcheck or a rubber-stamp sign-off does not. And there is a sharp edge to watch. If an AI system touches the content again after editorial sign-off, the exception breaks, and you owe a label.

When you do have to disclose, the standard is simple. The label has to be clear, shown at the first exposure, meaning the top of the text or the start of the video, and accessible to people with disabilities. The test a reader applies is easy: would a normal person notice it without hunting for it?

This is worth getting right. Non-compliance with Article 50 can draw fines up to 15 million euros or, for a company, up to 3 percent of total worldwide annual turnover, whichever is higher. It is not, however, worth panicking over. If you already run content through real review before it goes live, you are most of the way there. The teams that will find August hard are the ones where AI has quietly started publishing with no human in the loop.

We build AI into a platform you publish from, so the question we actually care about is what changes for you when you use Sulu.ai. The short answer is less than you might fear, because the law rewards the workflow Sulu is built around.

Your everyday tasks mostly stay out of scope

Start with the everyday jobs. Writing assistance, page and media translation, alt-text, SEO titles and descriptions. These land squarely in the carved-out zone we just covered. They help you publish, they do not fabricate a false picture of reality, and on their own they do not create a labelling duty. That is most of what Sulu.ai does, day to day, already out of scope.

Your editorial workflow is the exception

Where it gets interesting is the public-interest exception, because that is less about the AI and more about the platform around it. Transparency there is decided by who reviews, who signs off, and who is named as responsible. That belongs where editing and publishing actually happen, not in a separate compliance tool bolted on the side. Sulu gives you a central place to manage content, with editorial workflows and role-based access that decide who reviews and approves what you publish. Draft with Sulu.ai, have an editor review the substance and sign off inside that workflow, name someone as responsible, and you are inside the exception. The rule of thumb writes itself: let Sulu.ai assist up to the review, then let a person own the final call.

Why your model choice matters

Your model choice matters here too, because the marking duty sits with the provider behind your tools. Sulu.ai lets you choose your own model, so you can favour providers whose marking and detection meet the rules and who sign the Commission's Code of Practice on Transparency of AI-Generated Content, published in June 2026. Because every supported model draws on the same usage credits, moving to a compliant provider does not change your workflow or your pricing. As detection and watermarking standards mature, that freedom to move becomes more valuable, not less.

A note on Intelligent Search

Most of Sulu.ai helps editors before they publish, but Intelligent Search works differently. It answers visitors' questions directly, making transparency part of the user experience. Users should know they're receiving an AI-generated response, while provider-side marking obligations remain the responsibility of the model provider. There is one built-in relief, straight from the guidelines: an answer shown only to the person who asked is not published to a wide audience, so the public-interest labelling rule does not attach to it. 

That said, we still recommend making it clear that visitors are interacting with AI. That's how Intelligent Search is presented on our own website, and we believe it's good practice regardless of whether a legal obligation applies.

 We are building Search on grounded generation, where answers come from your own content and carry references back to the source pages. Grounding improves accuracy and trust. It does not replace transparency, but the two work well together.