AI Update · 6 August 2026
Your chatbot now has to say it's a bot
On August 2, the EU's transparency rules for AI became enforceable. Chatbots must identify themselves and AI-generated images must be labeled. Most businesses here are not legally in scope, and it will still change how their tools behave.
On August 2, a set of rules in the EU AI Act became enforceable. The short version: an AI system that talks to a person has to say it is an AI system, and AI-generated images, video and audio have to be labeled as such.
The European Commission put it plainly in its announcement — chatbots and other interactive systems have to tell users they are dealing with AI, not a human. Deepfakes and synthetic media need visible labels and machine-readable markers underneath.
This is the first AI regulation that touches the exact tools a small business actually uses.
What actually took effect
Article 50 of the AI Act covers transparency, and it has two halves.
The first is disclosure. If a customer is chatting with a bot on your website, in your DMs, or on a phone call placed by an AI agent, they have to know. The disclosure has to arrive at first contact, and it has to be hard to miss. Legal analysis of the rule is consistent on that last part: burying it in terms and conditions or three menus deep does not satisfy anything.
The second is labeling. Images, audio and video that were generated or meaningfully altered by AI have to carry a marker. Not just a visible caption — a machine-readable one embedded in the file, so platforms can detect it automatically.
Worth noting what did not take effect. The heavier obligations for so-called high-risk AI systems were pushed back to December 2027 by an amendment the EU Council approved on June 29. Plenty of coverage in July treated that delay as the whole story. It was not. Transparency was carved out of the deferral and landed on schedule.
You are probably not in scope. It reaches you anyway.
If you run a clinic in Haifa or a shop in Nazareth serving customers here, this is European law and you are outside it. If you sell into Europe, or your bot talks to European customers, the picture changes and you should ask someone who does this professionally rather than take a blog post's word for it.
But scope is not the reason to pay attention. The reason is that your tools are about to change underneath you.
The companies that build your chat widget, your booking assistant and your image generator will not maintain one version for Europe and another for everyone else. They will ship compliance globally because it is cheaper. Expect a disclosure line to appear at the top of your chat widget. Expect an invisible watermark inside every image your generator produces, whether you asked for it or not.
That last one has a real consequence. Content credentials embedded in AI-generated images are already readable by some platforms. A product photo you generated and passed off as a shoot is now checkable.
The disclosure that does not count
If you decide to get ahead of this, there is a wrong way to do it that looks like the right way.
A line in your privacy policy does not count. A grey asterisk under the send button does not count. "Powered by AI" in your website footer does not count. The standard is that a reasonable person notices before they start talking.
What counts is the first thing the bot says. One sentence, in your voice, before the conversation starts.
Start this week
Three things, none of which take an afternoon.
Open your own chat widget as a customer. Read the first message. If it does not identify itself as automated, rewrite it. Something like: "Hi, I'm the automated assistant for [business]. I can answer most questions, and I'll pass you to a person whenever you want." That sentence does the disclosure job and sets expectations at the same time.
Give people a way out. The rules do not require a human handoff, but the trust does. A bot that cannot say "let me get someone" is a dead end with a friendly tone.
Decide your line on AI images. You do not have to stop using them. Decide where they are fine — backgrounds, social graphics, mood imagery — and where they are not. A generated photo of a dish you do not serve, or a treatment result you did not produce, was already a problem. It is now a detectable one.
The part that is not about compliance
Every business that has run a bot for a while learns the same thing: people work out that it is a bot within two messages. The ones who were not told feel handled. The ones who were told just ask their question.
Disclosure was always the better product decision. A regulator in Brussels has now made it the default. You can adopt it this week and skip the part where a customer discovers it on their own.
Written and illustrated by the Create AI pipeline — the same one we put to work for clients.