If you have a chat box on your website, if you publish blog posts written with the help of an artificial intelligence tool, or if you use generated images in your ads, then starting Sunday, 2 August, you have a new duty: telling people clearly that a program is doing the work, not a person. The rule comes from the European artificial intelligence regulation, known as the AI Act, and in Romania market supervision falls to ANCOM. The good news is that for a small company the whole job fits into about an hour of work.
What applies from 2 August
The part coming into force now is called, in the text of the law, Article 50, and it asks for one thing: transparency. The date of 2 August 2026 stayed unchanged, even though other chapters of the regulation were postponed (Latham & Watkins).
Translated for a business owner, there are four situations (EU Artificial Intelligence Act):
1. Talking to an automated assistant. If a customer writes in a chat box on your site, gets an automatic reply on WhatsApp, or speaks to a phone robot, they must know the other side is a program. The exception: cases where it is obvious to anyone.
2. Computer generated content. Text, images, audio and video produced by artificial intelligence must be marked, so that another program can recognise that a human did not make them. The technical duty here falls mainly on the tool vendor, not on you.
3. Systems that read emotions or sort people by biometric data. If you use such systems, the people exposed to them must be informed.
4. Material that looks real but is not. If you publish a photo or a clip altered by artificial intelligence that looks authentic (a so called deepfake), you have to say it is artificial. The same logic applies to published text on matters of public interest, with one important exception: if a person has reviewed it and takes editorial responsibility, the marking duty falls away.
The notice has to appear at the first interaction at the latest, in a form that is easy to spot, not buried in the terms and conditions.
This is not a law only for technology companies
This is where most entrepreneurs get it wrong. It does not matter whether your company writes software or sells garage doors. What matters is whether you use these tools. "Any company that uses an artificial intelligence system becomes, in the eyes of the law, a deployer with clear obligations", explains lawyer Marius Stanciu (Digi24).
So the list includes the accounting firm with an automated chat on its site, the online shop that generates product descriptions, the clinic publishing articles written with an AI assistant, and the agency putting a generated image into an ad.
In Romania the roles were split by the decision of 12 March 2026: ANCOM is the main market surveillance authority, ASF and BNR cover the financial area, ANSPDCP covers sensitive domains, and ADR is the notifying authority (Digi24).
What to do in practice, in this order
Step 1: a one page inventory. Write down every place where your company touches artificial intelligence: chat on the website, automated messages, blog texts, product descriptions, images in ads, synthetic voices in clips.
Step 2: one sentence next to the chat. Something simple and visible from the very first message, along the lines of "You are talking to an automated assistant. For a colleague from our team, type «operator»." You do not need a lawyer for this, only an honest line placed where people see it.
Step 3: a note on generated material. For images and clips created by a computer, add a visible mention. For blog posts written with AI help, the cleanest solution is also the healthiest one editorially: a person reads them, checks the numbers and takes responsibility. That gets you out of the marking duty and, on top of that, you publish accurate things.
Step 4: one question for your suppliers. Write to the vendor of your chat solution or text generation tool and ask directly: who handles automatic content marking, you or us? Having the answer in writing helps later.
Step 5: put a date in the calendar. Review the list once every three months. These tools change faster than internal procedures do.
How big the fines are, and why not to panic on Monday morning
The figures in the regulation are large: up to 15 million euro or 3 percent of total annual worldwide turnover for breaching the transparency obligations (Latham & Watkins), up to 35 million euro or 7 percent for prohibited practices, and up to 7.5 million euro or 1 percent for false or misleading information given to the authorities (Digi24).
Three details matter enormously for a small Romanian company, though.
First: ANCOM has announced that, together with the other authorities, it is preparing the national law that sets out the responsible institutions, the cooperation mechanisms and the sanctioning procedure. Until that law enters into force, the authorities cannot apply sanctions for breaches of AI Act obligations (Economica.net).
Second: the simplified compliance framework, originally designed for small companies, was extended to companies with up to 750 employees and 150 million euro in annual revenue, with simplified guidance, reduced fines and standardised documentation templates (Latham & Watkins).
Third: for generative tools already on the market before 2 August 2026, the technical content marking duty only applies from 2 December 2026 (Latham & Watkins), and the heavy chapter, the one on high risk systems, was moved to 2 December 2027.
The practical conclusion: you have time to do the job properly, but no reason at all to postpone it, because the part that concerns you is cheap and quick to settle.
Care with your data, the part everybody forgets
Transparency is only half the story. The other half is about what you send into these tools, and there you answer to GDPR, not to the AI Act.
The rule of thumb: never paste into a public chat anything you would not put on a poster at your company entrance. No client lists with names, phone numbers or national ID numbers. No signed contracts. No payslips. No passwords or access keys. No medical files or documents from your lawyer.
What to do instead: replace real names with "Client A" and "Supplier B", delete addresses and identification codes, keep only the figures you actually need for the calculation. If you want a draft reply to a complaint, describe the situation in general terms without revealing who the person is. The result is just as good, and the risk disappears.
Two more useful things. Check in the tool settings whether your texts are used to train the model, and switch that option off if it exists. And write half a page for your team about what is allowed and what is not, because most data leaks do not come from bad intent, they come from being in a hurry.
One concrete step for this weekend
Open your company website, look at the chat box and at the last three pieces you published. If any of them was made by a program and nothing on the page says so, add a sentence right now that does. Ten minutes of work, one obligation ticked off and, more importantly, a customer who understands who they are talking to.
Sources
- Latham & Watkins, AI Act Update: EU Resolves to Change Rules and Extend Deadlines
- EU Artificial Intelligence Act, The Transparency Rules of Article 50
- Digi24, Din august 2026, AI Act devine pe deplin aplicabil în UE. Ce trebuie să știe firmele din România și ce amenzi riscă
- Economica.net, ANCOM anunță că regulamentul UE privind inteligența artificială trebuie pus în aplicare de România începând cu data de 2 august
- AI Act Blog, Article 50 transparency obligations: the AI Act deadline of 2 August 2026 that has not been postponed