A date slid past on 2 August that a lot of small businesses missed, and the ones who did hear about it mostly heard from somebody selling an audit. The transparency part of the EU AI Act is now live. If you have a chatbot on your website, or you use AI to write things you publish, some of it lands on you. Most of it doesn't.
What actually changed
The rule in question is Article 50 of the AI Act, and it covers four situations: AI systems that talk directly to people, AI-generated audio, images, video and text, emotion recognition and biometric categorisation, and deepfakes. The European Commission confirms it applies from 2 August 2026, with one limited grace period: systems already on the market before that date have until 2 December to meet the machine-readable marking requirement. Content generated before August does not have to be labelled after the fact.
Worth saying plainly, because the scare emails blur it: this is not the high-risk part of the Act. Conformity assessments, CE marking and database registration were pushed out to December 2027, and to August 2028 for AI built into regulated products.
The bit that lands on you: a chatbot
If your site has a chat widget backed by AI, people must be told they're talking to a machine from the start of the first interaction, clearly, unless it's obvious. The Commission's guidance says the "obvious" exception should be read narrowly, so don't lean on it. In practice this is a five-minute job:
- Say it in the opening message. The first line the visitor sees, not buried in a policy page
- Label the widget itself: the header, the button, the tab title
- Don't let it claim to be a person. If a visitor asks, it answers straight
- Check the accessible version. Screen reader users need the same notice, so put it in the ARIA label too
I had to look at our own before writing this. The chat bubble on this site opens with "I'm Finn, Fentech's AI assistant", the panel header reads "Fentech AI", and the footer of the window says the same again. That's the standard I'd aim at: obvious to a person skimming, and obvious to software reading the page.
What you don't have to do
What the compliance emails imply
- Every business using AI needs registration and an audit
- Every AI-assisted email and blog post needs a disclaimer
- You must add watermarking to AI content yourself
- Fines start at millions, for anyone
What Article 50 says
- Registration and assessment belong to the high-risk tier, from 2027
- Text with real human review is exempt from the labelling rule
- Machine-readable marking is the tool provider's job, not yours
- For SMEs the fine is the lower of the cash cap or the turnover percentage
That last one matters. The headline penalty for a transparency breach is up to €15 million or 3% of worldwide turnover. But the Commission's own guidance notes that proportionality is taken into account for SMEs, and the Act says small firms pay the lower of the two figures rather than the higher. A percentage of a Roscommon turnover is not a percentage of Meta's.
AI-written text: when a label is needed
People get needlessly worried about this part. The labelling duty on published AI text is narrow. It applies to text put out to inform the public on matters of public interest (politics, public health, justice, consumer safety, that kind of thing), and only where it went out without human review or editorial control. A person with relevant knowledge actually checking the substance counts as review. A spellcheck does not.
So a plumber's service page written with AI help and read over properly before publishing isn't caught. A newsletter about a public health matter, generated and fired out unread, is a different conversation. Background AI, the sort that sorts your inbox or flags a dodgy transaction with no direct exchange with a person, falls outside the "tell them it's AI" rule entirely.
Who enforces it here
Enforcement sits mainly with national market surveillance authorities rather than Brussels. Ireland is standing up its own structures: the Government published the Regulation of Artificial Intelligence Bill 2026 in February, setting up a national AI Office to coordinate the existing sectoral regulators. Nobody is going door to door in Roscommon looking at chat widgets. But the obligation is live now, and a labelled chatbot is a cheap thing to have already done.
What I'd do this week
Since 2 August, an AI that talks to your customers has to introduce itself as one. Published AI text needs a label only in a narrow public-interest case, and human review removes even that. The heavy machinery of the Act (assessments, registration, high-risk classification) is years away and aimed elsewhere.
So check your chat widget. Check any AI phone system. Write down which AI tools your business uses. Then get on with your week.
