AI watermarking and deepfake labelling will soon be required by law in the EU as the transparency part of the bloc’s sweeping AI Act comes into force on August 2nd.
Anyone living in the EU will become better informed about when they are dealing with AI, under new rules that come into force on August 2nd.
Under the terms of the AI Act: “AI providers will have to design AI systems to inform users when they are directly interacting with AI and they will have to add machine-readable marks to enable the detection of AI-generated or manipulated content”.
Chatbots
Any website, customer service department, or government agency using a chatbot must inform users that they are interacting with an AI – rather than a real person – unless there is already no doubt that AI is involved (for example, if users are given the message ‘ask our AI chatbot for more information’ or similar).
Such information must be clearly displayed at the appropriate time, and easily understood. It is the chatbot provider’s responsibility to design it in such a way that the user receives this information.
It should be noted this is for all new chatbots operating in the EU from August 2nd. Existing ones have until December 2026 to comply with the legislation. Many already do.
Pictures, video or text created by AI
New rules also apply to content created or edited by AI.
Any image, video, sound, or text produced by an automated system must be identifiable as such by a technical mark that will allow their artificial origin to be traced.
This traceability mark could take several forms: data embedded in the file, information about its origin, or a technical watermark, accessible to verification tools. Crucially, this information must remain intact even if the original is cropped or otherwise altered.
Verification tools must be permitted to retrieve this information – but, crucially, it does not have to be clear to the user. So, most photos, videos, or articles generated by AI do not have to be accompanied by a visible “generated by AI” label.
This principle also applies to original content modified with AI.
Deepfakes
Deepfakes – with certain concessions for personal, artistic, satirical and fictional use – must be clearly marked as generated or manipulated by AI.
Viewers of a video that apparently features a politician, business leader, or other identifiable person offering information, advice or comment, must be able understand that they are seeing something that is AI-generated or manipulated.
Under the AI Act, an audio deepfake – similar to one in the US where a deepfake Joe Biden message urged Americans not to vote in the 2024 Presidential election – would need to begin with a disclaimer stating in “plain and simple natural language” that it is AI-generated.
Furthermore, emotion recognition and biometric categorisation systems – for example, facial recognition – must, from August 2nd, inform people who are exposed to them.
Next phase
A third wave of the AI act, taking in recruitment tools, credit scoring, education, law enforcement, border control, and critical infrastructure – which involve much stricter regulations – comes into effect on December 2nd, 2027.
Systems used as safety components in certain products, such as medical devices, toys, or elevators, will be subject to these regulations starting August 2nd, 2028.
Fines of up to €35 million or seven percent of global revenue can be imposed for non-compliance.
