AI Act: Brussels Has Not Banned AI - It Has Simply Asked Us to Pay Attention

Transparency: please tell me whether I am speaking to a person 

The most visible change is the transparency obligation in Article 50 of the AI Act. Where an AI system is designed to interact directly with natural persons, those persons must in principle be informed that they are interacting with AI, unless that is obvious from the circumstances. The idea is not particularly revolutionary. If a customer believes he is exchanging messages with Sophie from customer service while Sophie is in fact an algorithm having a very productive afternoon, the customer should be told. 

Fortunately, the legislator has not prescribed a three-page disclaimer in twelve-point Times New Roman. A clear and timely notice should normally do the job. For many businesses, compliance will therefore be more a matter of housekeeping than constitutional reform. Still, the rule matters. Transparency is not a decorative extra in the AI Act; it is one of the recurring themes of the entire Regulation. 

Deepfakes: excellent editing does not turn fiction into fact 

The same logic applies to so-called deepfakes. AI-generated or manipulated images, audio and video which may reasonably appear authentic must be disclosed as artificially generated or manipulated. The provision is easy to understand when one thinks of fabricated political speeches or a celebrity apparently endorsing a product they have never heard of. But it is equally relevant in ordinary commercial practice. 

Marketing departments can therefore continue to be creative. Virtual models, synthetic voices and AI-generated campaign material are not prohibited. The line is crossed when creativity becomes deception. A fictional spokesperson is perfectly possible; presenting that spokesperson as a real person without saying so is rather less comfortable. Lawyers have always had a soft spot for disclosure, preferably before the litigation starts. 

AI-generated text: not every sentence needs a warning label  

Article 50 also deals with certain AI-generated or manipulated text published for the purpose of informing the public on matters of public interest. This is one of the areas where headlines tend to be rather more dramatic than the legislation itself. The AI Act does not require every e-mail, brochure, LinkedIn post or legal memorandum touched by ChatGPT to carry a scarlet letter saying ‘AI was here’. 

For most professional users, AI will remain a drafting tool. A lawyer may ask it to suggest a structure, a business lawyer may use it to summarise a long agreement, and a marketing team may ask for a first version of a press release. If a human being then checks, edits and takes responsibility for the result, the practical position is very different from publishing fully automated content without meaningful editorial control. The machine may hold the pen for a moment; it should not be allowed to sign the advice. 

AI literacy: knowing enough to know when not to trust it 

Strictly speaking, the obligation concerning AI literacy has applied since 2 February 2025, but 2 August 2026 is a useful moment to revisit it because the surrounding compliance framework has now become much more tangible. Article 4 requires providers and deployers of AI systems to take measures, to their best extent, to ensure a sufficient level of AI literacy among staff and other persons dealing with AI systems on their behalf. 

There is no European licence to operate ChatGPT and, mercifully, no compulsory examination in neural networks. The required level of understanding depends on the context. An HR manager using AI to rank candidates should know more about bias, data quality and automated decision-making than a receptionist using an AI tool to improve the wording of an invitation. A lawyer using AI for legal research should at the very least know that a beautifully formatted case citation may still refer to a judgment that has never existed.  

That last point is worth emphasising. Generative AI has a rather charming habit of sounding most confident precisely when it is wrong. Experienced litigators will recognise the phenomenon. The difference is that, in court, counsel can usually see who is making the argument. 

Human oversight: the machine may assist, but it does not inherit the file 

The principle of human oversight runs like a red thread through the AI Act. AI can assist with contract review, recruitment, compliance checks, risk assessments or customer communication, but responsibility does not disappear merely because software has produced the first answer. For lawyers and in-house counsel this should feel familiar. Delegation is possible; abdication is not. 

A practical example is contract review. AI can flag an unusual indemnity clause in a few seconds, compare definitions across a document and even propose alternative wording. That is useful. It is also considerably faster than asking a junior lawyer to read the agreement at 6.30 p.m. on a Friday. But if the AI overlooks a change-of-control clause that destroys the transaction, explaining afterwards that ‘the system did not pick it up’ is unlikely to impress the client, the board or the insurer. 

Confidentiality: the oldest rule remains one of the best 

For professional users, there is another point which the AI Act does not magically replace: confidentiality. Lawyers, patent and trademark attorneys, HR departments and management teams routinely handle information that should not simply be pasted into whichever AI tool happens to be open in the browser. Trade secrets, litigation strategy, unpublished inventions, personal data and draft transaction documents deserve a little more ceremony. 

A sensible internal AI policy should therefore say which tools may be used, for what purpose, with which categories of information and subject to which checks. This is not bureaucratic enthusiasm. It is the digital equivalent of not leaving a confidential file on the train. The fact that the train now answers questions does not improve the situation. 

What should companies actually do? 

For most companies, the answer is reassuringly practical. They do not need to establish an AI ministry. They should identify the AI tools already being used, decide which uses are acceptable, train staff sufficiently, introduce a few clear rules on confidential and personal data, verify where transparency notices are required and make sure that important outputs are subject to meaningful human review. 

The temptation, particularly in larger organisations, will be to produce a forty-page AI policy, a twelve-page risk matrix and an annual e-learning module which every employee completes while answering e-mails. That may satisfy a spreadsheet somewhere, but it is not necessarily compliance. A concise policy that people understand and actually follow is usually more useful than a magnificent policy nobody has read since the steering committee approved it. 

The lawyer’s role 

For lawyers and in-house counsel, AI compliance will therefore be less about becoming computer scientists and more about asking familiar legal questions in a new technological setting. Who is doing what? On whose behalf? With which data? Who checks the result? What is disclosed to the outside world? What happens when the system gets it wrong? These are not exotic questions. They are the questions lawyers have always asked. Only the vocabulary has changed. 

This is also why the AI Act should not be approached as a stand-alone island. Depending on the application, GDPR, consumer law, employment law, intellectual property law, confidentiality obligations and sector-specific regulation may be at least as important. AI has not replaced the existing legal framework. It has merely acquired a rather substantial new neighbour.  

Conclusion 

The AI Act has not made artificial intelligence illegal, nor has it required every European company to hire a Chief Artificial Intelligence Officer before lunch. What it has done is put an end to the idea that using AI is legally neutral simply because the technology is easy to access. 

Since 2 August 2026, transparency obligations have become more concrete, AI literacy matters more in practice and human oversight is no longer something to be added to a PowerPoint slide at the end of a compliance presentation. Businesses should know where AI is being used, understand the risks and keep a human being meaningfully involved. 

Perhaps the best rule remains the simplest one. Treat AI as you would treat a very bright trainee: fast, useful, occasionally ingenious and perfectly capable of producing something that looks excellent at first sight. By all means use it. Just do not send the work out without checking it. Artificial intelligence may be artificial. Professional responsibility remains, for the foreseeable future, stubbornly human. 

And, in the interests of transparency: AI helped write this article. But the views, the style, the irony and any mistakes are still very much my own. 

De Clercq & Partners is happy to assist you in assessing how the EU AI Act applies to your organisation’s use of AI and in developing practical measures regarding transparency, AI literacy, confidentiality and human oversight. Please do not hesitate to contact us at info@dcp-ip.com