From today, the way you interact with every digital tool in Europe is changing radically. If you are talking to a chatbot or watching a modified video, the platform must now tell you with absolute clarity. At the same time, the Ευρωπαϊκή Επιτροπή ξεκινά το πρέσινγκ στους κολοσσούς της τεχνολογίας, θέτοντας σε ισχύ τις πρώτες αυστηρές υποχρεώσεις του AI Act.
- Chatbots and the content they produce artificial intelligence (deepfakes, audio, image) are now required to have clear labeling and machine-readable watermarks.
- The European AI Office gains control power over general purpose AI models (GPAI), enforcing transparency in training data and respect for copyright.
- Heavy bans for high-risk systems are postponed to 2027 and 2028, while the success of the measure depends on the readiness of national authorities.
How your daily life is changing with new chatbots
Until yesterday, you could chat with an online support bot or a digital assistant without knowing whether the other side was a human or an algorithm. With the activation of the provisions of the European Regulation on Artificial Intelligence (AI Act) , this ends. Every interactive system is obliged to make it clear to you from the very beginning that you are interacting with software.
The measure is not limited to text. Images, audio clips and videos that have been created or transformed with artificial intelligence (the so-called deepfakes) must carry a visible or audible mark, as they may mislead the public as to their authenticity.
Behind the visible indication lies a more important technical requirement: the integration of “machine-readable metadata”. This means that files will carry digital watermarks, allowing platforms such as YouTube, X and Facebook to automatically detect synthetic content. To this end, the Commission has published the first list of more than 180 organisations that have joined the Transparency Code of Practice , which specifies technical standards for companies.
GPAI: The tight squeeze on AI giants
The European Union is now openly targeting General Purpose AI (GPAI) models. We're talking about the underlying technologies — such as OpenAI's GPT-4, Google 's Gemini , or Anthropic's Claude — on which hundreds of other applications are built.
The newly created European Artificial Intelligence Office (EU AI Office) is acting as a central controller. Providers of these models are now required to submit full technical documentation, comply with European copyright law and publish a detailed summary of the material they used to train their systems.
For models classified as “systemic risk” — that is, those trained with massive computing power and capable of affecting entire societies — the obligations are tightened. Their makers are required to assess and mitigate risks related to:
- Massive cyberattacks and infrastructure breaches.
- Psychological manipulation of users on a large scale.
- Involvement in chemical, biological or nuclear (CBRN) incidents.
- Cases where the system may act outside of human control.
The schedule: What's in effect today and what's postponed
If you think that the entire AI Act is implemented from today, the reality is different. The European Union has chosen a gradual implementation in order not to disrupt the market.
The design provides specific stages for the following requirements:
- August 2 (Today): Transparency rules for chatbots, deepfakes and GPAI model provider obligations apply.
- December 2, 2026: Strict bans are in place for systems that produce sexually explicit material without consent or child sexual abuse material.
- December 2, 2027: Rules for "high-risk" systems are activated (e.g. resume evaluation algorithms, justice system, credit rating).
- August 2, 2028: The rules apply to AI systems that are integrated into already regulated products (medical devices, automotive industry, aeronautics).
| AI Category | Basic Obligations | Competent Authority | Effective Date |
|---|---|---|---|
| Chatbots & Deepfakes | Conversation tagging, metadata watermarking | National Authorities / EU AI Office | Immediately (August 2) |
| GPAI Models (General Purpose) | Documentation, copyright policy, training data summary | EU AI Office (Commission) | Immediately (August 2) |
| Non-consensual Explicit AI | Total ban on creation & distribution | National Authorities / Courts | December 2 2026 |
| High-Risk AI Systems | Risk assessments, human oversight, audit logs | National Authorities of Member States | December 2 2027 |
The reality: Application gaps and open source
Here's the big detail that most press releases miss. The European AI Office is staffed by a scientific team of 60 independent experts. When you consider the volume of models released every day, that number is a drop in the ocean.
Furthermore, oversight is split down the middle: the AI Office oversees the giants and large platforms of the Digital Services Act, but for all other systems, responsibility falls to national authorities. In Greece — as in several other EU countries — the competent bodies do not yet have the necessary staffing or technical infrastructure to carry out meaningful code audits.
The second big thorn is open-source models. If you download a model locally to your computer, you can remove any watermarking mechanism with a few lines of code. The law requires companies to put a watermark, but it cannot prevent a malicious user from bypassing it by running the model offline.
Our opinion at TechNoid
At TechNoid, we believe that the AI Act is a necessary initiative at the institutional level, but suffers from excessive bureaucracy in practice. Requiring chatbots to declare their identity is a win for the consumer. However, the idea that digital watermarking will solve the problem of deepfakes is oversimplified. As open source software evolves, watermarks will be removed in seconds. The result? Legitimate European businesses are burdened with expensive compliance costs, while malicious actors will continue to operate unhindered outside European borders.
Frequently Asked Questions about the European AI Act Regulation
How will I know if I'm talking to a chatbot or a human?
The application or website is now required to display a written notice at the start of the conversation, explicitly informing you that the system is based on Artificial Intelligence.
Will I see special marking on all AI images and videos?
Not necessarily with a visible watermark. Systems are required to incorporate mainly machine-readable metadata so that platforms (social media) can automatically recognize and flag the content.
What are General Purpose AI (GPAI) models?
They are flexible models like GPT-4 or Gemini, which were not designed for a single job, but can perform dozens of different tasks and be integrated into other applications.
How are creators' copyrights protected?
GPAI development companies are required to publish summaries of training data and have a policy of compliance with EU intellectual property law.
Where can I file a complaint if I detect a violation?
The European Commission is activating a special complaints portal for citizens and businesses, as well as a secure channel for whistleblowers (employees of technology companies).
Why are the rules for high-risk systems delayed?
The EU has given a transition period until December 2027, so that businesses and national authorities can adapt their infrastructure without disrupting the operation of critical services.
Can the AI Act stop deepfakes completely?
Not completely. Despite strict rules, the use of open source models at the local level makes it technically difficult to universally ban content tampering.
