Most businesses filed the EU AI Act under "something for big tech to worry about." Since 2 August 2026 that filing is wrong. The transparency obligations in Article 50 are now enforceable, and they do not care how large you are or whether your AI is high-risk. They apply based on what your system does.
If you run a chat widget, publish AI-drafted content, or generate images for your marketing, you are in scope. The obligations are modest. Ignoring them is not.
What Article 50 actually requires
Strip away the legal drafting and there are two duties that matter for an ordinary company:
Tell people when they are talking to a machine
If a person interacts with an AI system, they must be informed of that—unless it is obvious to a reasonably observant person in the circumstances. A chat widget that opens with a friendly first name and never says otherwise does not meet this. One that identifies itself as an assistant does.
Mark AI-generated content as AI-generated
Synthetic text, images, audio and video must be marked in a machine-readable format and be detectable as artificially generated. The machine-readable part is the half most businesses miss: a line of small print at the bottom of an article satisfies a human reader, not the requirement.
There is a transition for systems already on the market before August 2026, which have until 2 December 2026 to comply with the content-marking duty. That is the deadline actually in front of most companies now, and it is close.
What was postponed, and why it does not help you
You may have read that the AI Act was delayed. Partly true, and widely misread.
The obligations that were deferred are the ones on high-risk systems—Annex III standalone systems pushed to December 2027, with separate timing for systems embedded in regulated products. High-risk means recruitment, credit scoring, education access, critical infrastructure. If you are running a support chatbot, none of that was ever your deadline.
Article 50 was not postponed. For a company whose AI arrives through a subscription or an integrated tool, 2 August 2026 was not a quiet date—it was the moment the first concrete obligations became real. The delay headlines were about a different set of rules than the ones that apply to you.
What this means for three ordinary setups
A chat widget on your website
Disclose that it is an AI assistant. Not buried in a policy—in the interface, where a visitor sees it before typing. The good news is that this is also better product design: people calibrate their questions when they know what they are talking to, and they are far less annoyed by "I don't know, let me get someone" from an assistant than from what they assumed was a person.
If a human takes over the conversation, say that too. The handoff is already the most important moment in a support chat; making it explicit is now both a legal duty and the thing customers want.
AI-drafted articles or product descriptions
This is where the machine-readable requirement bites. Marking means embedded metadata, not a visible sentence—though a visible note is good practice and builds trust with readers.
Worth noting: content that a human has reviewed and edited with real editorial responsibility sits differently from raw model output published untouched. The distinction is not a loophole to lean on, but it is a strong argument for a review step you should want anyway. This is exactly why our AI Article Publisher holds every draft in a review queue and publishes nothing automatically—a design decision made for editorial quality that turns out to be the compliant one too.
AI-generated images in your marketing
The same marking duty applies. Most reputable image tools now embed provenance metadata by default—check whether yours does, and whether your own pipeline strips it. Image optimisation and resizing routinely discard metadata, which means a compliant image can become a non-compliant one on the way to your website without anyone touching it deliberately.
A short, honest checklist
You can work through this in an afternoon:
- List every place AI touches something a customer sees—chat, articles, product copy, images, email replies, translations.
- For each, ask: does a person know this is AI? Is the output machine-markable?
- Check whether your tools already embed provenance metadata. Several do; you may be closer than you think.
- Check whether your own processing strips that metadata. This is the most common silent failure.
- Write down who reviews AI output before it is published, by name. If the answer is nobody, that is the finding.
The framing that helps
Nearly everything Article 50 requires is something you would do anyway if you cared about the relationship with the person on the other side. Saying "this is an assistant" sets expectations correctly. Marking generated content protects you the day someone accuses you of passing off a machine as a person. Keeping a human review step is how you avoid publishing something embarrassing.
The companies that will struggle are the ones that deployed AI quietly, hoping nobody would ask. The ones that were open about it from the start have very little to change—which is the usual pattern with regulation that codifies what decent practice already looked like.
One closing caveat, and it matters: this is a practical summary written by engineers, not legal advice. The dates and obligations above reflect the position in September 2026, and this is an area that has already been amended more than once. If AI touches a regulated part of your business, have a lawyer read the actual text.
Not sure where AI touches your customers?
We'll walk your site and your workflows with you and map every point where a customer meets AI output—then tell you which ones need a change and which are already fine.
Tell us what repeats→Common questions
Does the EU AI Act apply to small businesses?
The Article 50 transparency obligations apply based on what an AI system does, not on company size. If you run a chat widget or publish AI-generated content, you are in scope regardless of headcount or turnover.
Wasn't the AI Act postponed?
Only partly. The deferrals apply to high-risk systems — Annex III standalone systems moved to December 2027, with separate timing for AI embedded in regulated products. Article 50 transparency was not postponed and has been enforceable since 2 August 2026.
What does it mean to mark AI-generated content?
Outputs must be marked in a machine-readable format and be detectable as artificially generated. A visible note at the bottom of an article satisfies a human reader but not the requirement. Systems already on the market before August 2026 have until 2 December 2026 to comply with this duty.