Short answer
Since August 2, 2026 the EU AI Act requires that people be told when they are talking to an AI system, and that AI-generated text, images, audio and video carry machine-readable marking. The high-risk obligations were postponed to December 2027 and August 2028. If your website assistant or AI-drafted content reaches people in the EU, label the bot, keep your vendor's watermarking on, and keep a simple inventory of the AI systems you run.
Most owners heard about the EU AI Act the way they hear about most regulation: a headline, a vague sense of dread, and then nothing. That is a reasonable reaction, because until this month almost nothing in the Act touched an ordinary company. On August 2 that changed, in a small and manageable way. This is the plain-English version.
What took effect on August 2, 2026
Article 50 of the Act, the transparency chapter, is now live. It has three parts that matter to a normal business:
- Chatbot disclosure. People interacting with an AI system must be told they are dealing with an AI, unless that is obvious from context. If your website has an assistant that answers questions, it should say so.
- Machine-readable marking of AI output. Providers of generative systems must mark the text, images, audio and video their systems produce so that detection tools can identify it as AI-generated. Systems already on the market before August 2 have until December 2, 2026 to comply.
- Deepfake labelling. Realistic AI-generated or manipulated images, audio or video of people, places or events must be disclosed as such.
The legal summaries from Gibson Dunn and Latham & Watkins are the clearest we have found if you want the primary detail.
What was postponed
The reform agreed in May and confirmed over the summer moved the expensive parts of the Act:
| Obligation | Old date | New date |
|---|---|---|
| Chatbot disclosure and content marking (Article 50) | August 2, 2026 | Unchanged (marking grace period to December 2, 2026) |
| High-risk uses listed in Annex III: recruitment, credit scoring, education, law enforcement | August 2, 2026 | December 2, 2027 |
| AI embedded in regulated products: medical devices, machinery, vehicles | August 2, 2027 | August 2, 2028 |
The simplified regime for smaller companies, with lighter documentation, reduced fines and access to regulatory sandboxes, now extends to firms with up to 750 employees and 150 million euros in revenue. In other words, nearly every company that reads this blog.
Does this reach a US company?
The Act applies to providers who place AI systems on the EU market wherever they are based, and to companies outside the EU when the output of their AI is used inside it. If your assistant chats with a prospect in Dublin, or your AI-drafted newsletter lands in inboxes in Berlin, the honest working assumption is that the transparency rules apply to you. The good news is that complying with them costs almost nothing, so the sensible move is to do it everywhere rather than try to geofence honesty.
The watermark question, explained
The marking rule is the one that generates the most confusion, because it sounds like your company has to stamp something on every email. It does not. The obligation sits with the provider of the generative system. Anthropic, for example, announced in August that Claude models launched on or after August 2 carry a statistical text watermark: a subtle bias in low-stakes word choices that a reader cannot see but a detector with the right key can. It adds no hidden characters, survives copy and paste, may survive light editing, and is likely to disappear under a heavy rewrite or translation. Anthropic says it will publish detection details and offer a detection API.
What this means for you is simple. You do not need to add anything. You should not try to strip anything. And you should stop assuming that AI-drafted text is undetectable, because increasingly it is not, at least to anyone with access to the vendor's detector. That is one more reason to be straightforward about where AI is used in your customer communication.
The short list worth doing this month
- Label the assistant. A single line such as "You are chatting with an AI assistant. Ask for a person any time." satisfies the disclosure rule and, in our experience, raises trust rather than lowering it.
- Make the human handoff real. Disclosure is hollow if there is no way to reach a person. Route "talk to someone" to a real inbox or number.
- Keep an inventory. One page listing each AI system you run, what it does, which vendor is behind it and who owns it internally. Regulators ask for this first, and it is also the document that tells you what you actually have.
- Check the high-risk list once. If you use AI to screen job applicants, score credit, or make decisions about access to education or essential services, you are in the high-risk category and should start planning for December 2027 now. Most back-office automation, from invoice extraction to inbox triage to quote drafting, is not on that list.
- Do not fake it. Do not present AI-generated reviews, testimonials, headshots or voice as human. That is where enforcement, and reputational damage, will concentrate.
How this fits with human-in-the-loop design
Every system we ship already assumes a person approves anything customer-facing or financial before it leaves the building. The transparency rules reward exactly that posture: if a human reviews and sends, the disclosure question is easy and the deepfake question never arises. Compliance is a side effect of good design, not a separate project.
Frequently asked questions
We are a US company with no EU office. Does the EU AI Act apply to us?
The Act covers providers who put AI systems on the EU market and, in some cases, companies outside the EU whose AI output is used inside it. If people in the EU can chat with your assistant or receive your AI-generated content, treat the transparency rules as applying and confirm with counsel. For most small companies the practical burden is small: a disclosure line and honest labelling.
Do I have to watermark the emails and proposals my team drafts with AI?
The marking obligation sits with the provider of the generative system, not with the business using it. Anthropic, for example, now applies a statistical watermark to text from Claude models launched after August 2. What you should avoid is deliberately stripping or hiding those marks, and presenting AI-written material as human-written where that would mislead someone.
What was postponed?
The obligations for high-risk systems listed in Annex III, such as recruitment screening and credit scoring, moved from August 2, 2026 to December 2, 2027. AI embedded in regulated products such as medical devices and machinery moved to August 2, 2028. The chatbot and content-marking rules were not postponed.
What is the penalty for getting the transparency rules wrong?
Fines for breaching obligations other than the outright prohibitions can reach 15 million euros or 3% of worldwide annual turnover, whichever is higher. Regulators have said they will start with guidance and codes of practice, and the Act's simplified regime for smaller companies now reaches firms with up to 750 employees.
