The European Union has begun enforcing a major new layer of its Artificial Intelligence Act, requiring businesses to disclose when people are interacting with AI and when certain content has been generated or manipulated by it.
The rules took effect on August 2, 2026, and represent the most visible part of the AI Act for ordinary users. Chatbots must identify themselves in situations where their artificial nature is not obvious. Realistic AI-generated audio, images and videos may require clear labels, while generative systems must embed machine-readable information that helps platforms and detection tools recognise synthetic content.
The legislation does not require a warning on every spell-checked email or AI-assisted photograph. Its obligations depend on who provides the technology, who deploys it and how the resulting content is presented to the public.
Chatbots Can No Longer Quietly Pretend to Be Human
Companies providing interactive AI systems must design them so that individuals are clearly informed when they are communicating with a machine rather than a person.
The requirement covers systems capable of genuine two-way interaction, including customer-service chatbots, AI agents and digital avatars. The disclosure must appear from the beginning of the first interaction and must be clear, distinguishable and accessible. Background automation and machine-to-machine systems that never communicate directly with a person generally fall outside this particular obligation. The European Commission’s Article 50 transparency guidance provides detailed examples of when a notice is required.
A disclosure may not be necessary when it is already obvious to a reasonably observant person that the system is artificial. However, the Commission says that exception should be interpreted narrowly because it removes information that could affect how much trust a user places in the conversation.
This could change automated customer support, recruitment platforms, travel booking services and AI-operated sales systems. A company may no longer be able to rely on a human-sounding name and realistic conversational style while leaving customers uncertain about whether a person is involved.
AI-Generated Content Needs Hidden Technical Markers
Providers of generative AI systems must also make synthetic outputs detectable by adding machine-readable marks.
These markers can include metadata, digital provenance information, watermarks or other technical signals that indicate whether text, audio, video or an image was generated or substantially manipulated by AI. The system must use methods that are effective, reliable, interoperable and robust as far as technically feasible.
The requirement is aimed at the technology provider rather than only the person publishing the content. A company developing an image generator, for example, must build detection support into its system before placing it on the European market.
The EU’s Code of Practice on Transparency of AI-Generated Content provides a voluntary framework for meeting these obligations. Signing the code is optional, but compliance with the underlying law is mandatory. Organisations that choose a different approach must show regulators that their alternative is equally adequate.
Deepfakes Must Carry a Visible Disclosure
Machine-readable metadata alone is not enough when the public encounters a realistic deepfake.
The organisation or professional using the AI system must provide a clear, perceivable label when an AI-generated or manipulated image, video or audio recording resembles a real or plausibly real person, object, place, organisation or event and could falsely appear authentic.
The notice must be understandable without special detection software. It may appear visually or be communicated through an audible disclosure, depending on the format. It should be provided no later than the audience’s first exposure to the content.
The rules are designed to address situations such as fabricated footage of political figures, cloned voices used in scams, altered recordings of public events and realistic images falsely presented as documentary evidence.
Artistic, fictional, satirical and creative works receive more flexibility. They may still need an appropriate disclosure, but it should not unnecessarily interfere with the audience’s enjoyment of the work.
Not Every AI-Assisted Image Needs a Warning
The definition of a regulated deepfake is narrower than “anything edited using AI.”
Standard production tools that remove noise, adjust colour, sharpen an image or assist with ordinary editing may not create a disclosure obligation when they do not substantially change the meaning of the original material.
Special effects and artificial backgrounds used in clearly fictional films are also unlikely to mislead an audience that does not expect the material to be authentic. The assessment depends on the resemblance, the message being communicated, the intended audience and the context in which the content appears.
This distinction prevents the rules from placing the same warning on a routine photo enhancement and a fabricated video appearing to show a government official announcing a policy that does not exist.
Providers are also exempt from machine-marking obligations in certain narrow situations, including source code, short strings of symbols and some outputs confined to closed industrial development environments.
AI-Written Public-Interest Content Faces Special Rules
Text receives a more targeted form of regulation.
AI-generated or AI-manipulated text must be clearly labelled when it is published to inform the public about matters of public interest and has not undergone meaningful human review or editorial control.
The Commission says this category can include politics, democratic processes, public administration, justice, law enforcement, fundamental rights, public safety, public health, environmental protection, consumer safety and significant economic, financial, scientific or cultural developments.
A newsroom or professional publisher may avoid the label when a qualified person has substantively reviewed the material, checked its sources and accepted editorial responsibility. Merely correcting grammar, approving the format or running a basic spell check is not sufficient.
The rule means that automatically generated news summaries, financial updates or public-health articles may require disclosure when no person has verified their substance. It does not automatically label every article in which a journalist used AI for research, transcription or routine editing.
Emotion Recognition Must Also Be Disclosed
The transparency regime extends beyond generative AI.
Organisations using AI to infer emotions or categorise people through biometric information must inform those exposed to the system. This applies whether the analysis occurs in real time or after a recording has been collected.
The provision could affect workplaces, educational platforms, retail environments, security systems and services that claim to evaluate attention, mood or emotional responses from voices, faces or body movements.
Disclosure does not necessarily make every use lawful. Other sections of the AI Act, data-protection rules and sector-specific laws may restrict or prohibit particular applications. The transparency requirement simply prevents covered systems from operating without the affected person knowing they are being used.
The EU Has Created Standard AI Labels
The Commission has introduced optional icons that publishers can use to identify synthetic material.
The designs distinguish between content created entirely by AI and human-made material that was partially modified. The icons are available in several black, white and transparent versions and can be accompanied by explanatory text.
The Commission’s official AI-label page says testing showed that the icons performed better when combined with a written description. Their use is optional, but the legal duty to provide an effective disclosure is not.
Consistent symbols could make AI warnings easier to recognise across websites and social platforms. However, different businesses may continue using their own formats as long as those disclosures satisfy the law.
The Rules Apply Beyond European Companies
The AI Act can affect providers established outside the EU when their systems are offered in Europe or their outputs are used within the bloc.
That gives the transparency rules global significance. American, British, Chinese and other international AI companies cannot necessarily avoid the requirements simply because their headquarters are elsewhere.
The EU has also expanded its AI Office in Brussels with 38 additional staff members and introduced confidential complaint and whistleblower channels. Regulators may request company documentation, interview employees and investigate whether AI products comply with the Act.
National market-surveillance authorities will handle much of the enforcement, while the AI Office oversees systems within its jurisdiction. The European Data Protection Supervisor is responsible when EU institutions use or provide the affected technology.
Violations Could Bring Multimillion-Euro Fines
Failing to meet the transparency obligations can result in fines of up to €15 million or 3% of a company’s total worldwide annual turnover, whichever legal calculation applies.
Regulators are expected to consider proportionality when dealing with smaller businesses and small mid-cap companies. EU institutions, bodies and agencies can face separate penalties of up to €750,000.
The obligations apply immediately to new systems. Generative AI systems already placed on the market before August 2 receive a limited transition period until December 2, 2026, but only for the requirement to embed machine-readable marks.
Content created before August 2 does not have to be relabelled retroactively, although the Commission encourages voluntary disclosure where practical.
Europe Is Entering Its “AI Label” Era
The transparency rules could become AI’s equivalent of the privacy notices and cookie banners that spread across the internet after earlier European regulation.
Supporters believe the disclosures will make impersonation, hidden automation and synthetic misinformation more difficult. Critics fear that excessive warnings could create label fatigue, causing users to ignore notices that appear too frequently or reveal little meaningful information.
The law attempts to avoid that outcome by focusing visible labels on direct AI interactions, deepfakes, biometric analysis and unreviewed public-interest content rather than every minor use of automation.
Its success will depend on whether technical markers survive editing and reposting, whether platforms can detect them consistently and whether companies provide labels that genuinely inform rather than merely satisfy a compliance checklist.
Europe’s message is nevertheless clear: AI may remain invisible inside software, but it should not remain invisible to the people whose decisions, beliefs and personal information it can influence.
