The European Union will require AI watermarking and deepfake labelling from August 2nd, when the transparency provisions of the bloc’s AI Act take effect. People living in the EU must be told when they are directly interacting with AI, and providers must add machine-readable marks that help detect AI-generated or manipulated content.

Websites, customer service departments and government agencies using chatbots must tell users they are interacting with AI rather than a real person, unless that fact is already obvious, such as through a prompt like “ask our AI chatbot for more information.” The information must appear at the appropriate time, be clearly displayed and be easy to understand. New chatbots operating in the EU must comply from August 2nd, while existing chatbots have until December 2026.

Images, videos, sounds and text created or edited by an automated system must include a technical traceability mark (machine-readable origin signal). The mark may be embedded data, origin information or a technical watermark accessible to verification tools, and it must remain intact if the file is cropped or otherwise changed. The article states that most AI-generated photos, videos or articles do not need a visible “generated by AI” label for ordinary users.

Deepfakes (AI-made or altered realistic media), with concessions for personal, artistic, satirical and fictional uses, must be clearly marked as generated or manipulated by AI. A video that appears to show a politician, business leader or other identifiable person giving information, advice or comment must make the AI manipulation understandable to viewers. An audio deepfake like the US Joe Biden message that urged Americans not to vote in the 2024 Presidential election would need to begin with a disclaimer in “plain and simple natural language.”

Emotion recognition and biometric categorisation systems, including facial recognition, must inform people exposed to them from August 2nd. A third wave covering recruitment tools, credit scoring, education, law enforcement, border control and critical infrastructure starts on December 2nd, 2027. Safety components in products such as medical devices, toys and elevators come under the rules on August 2nd, 2028, and fines can reach €35 million or seven percent of global revenue.

The European Union will require AI watermarking and deepfake labelling from August 2nd, when the transparency provisions of the bloc’s AI Act take effect. People living in the EU must be told when they are directly interacting with AI, and providers must add machine-readable marks that help detect AI-generated or manipulated content.

Websites, customer service departments and government agencies using chatbots must tell users they are interacting with AI rather than a real person, unless that fact is already obvious, such as through a prompt like “ask our AI chatbot for more information.” The information must appear at the appropriate time, be clearly displayed and be easy to understand. New chatbots operating in the EU must comply from August 2nd, while existing chatbots have until December 2026.

Images, videos, sounds and text created or edited by an automated system must include a technical traceability mark (machine-readable origin signal). The mark may be embedded data, origin information or a technical watermark accessible to verification tools, and it must remain intact if the file is cropped or otherwise changed. The article states that most AI-generated photos, videos or articles do not need a visible “generated by AI” label for ordinary users.

Deepfakes (AI-made or altered realistic media), with concessions for personal, artistic, satirical and fictional uses, must be clearly marked as generated or manipulated by AI. A video that appears to show a politician, business leader or other identifiable person giving information, advice or comment must make the AI manipulation understandable to viewers. An audio deepfake like the US Joe Biden message that urged Americans not to vote in the 2024 Presidential election would need to begin with a disclaimer in “plain and simple natural language.”

Emotion recognition and biometric categorisation systems, including facial recognition, must inform people exposed to them from August 2nd. A third wave covering recruitment tools, credit scoring, education, law enforcement, border control and critical infrastructure starts on December 2nd, 2027. Safety components in products such as medical devices, toys and elevators come under the rules on August 2nd, 2028, and fines can reach €35 million or seven percent of global revenue.