A new visibility test for everyday AI

From 2 August 2026, the European Union’s AI Act applies a new set of transparency obligations intended to make certain uses of artificial intelligence identifiable to people. The change is likely to be noticeable not because it bans familiar automated services, but because it asks providers and organisations deploying them to reveal when AI is involved in situations where that knowledge can affect trust, judgement or personal autonomy.

The result may be a subtle shift in how Europeans encounter technology. A person may be told that a customer-service assistant is automated, that a system is analysing emotions or assigning a biometric category, or that an apparently authentic image, video or audio clip has been artificially generated or manipulated. The rules also address machine-readable markings that can help platforms and other systems identify synthetic content.

That does not mean every use of an algorithm, or every product with an AI feature, will suddenly carry a conspicuous label. The legal scope is narrower and more structured than the broad claim that all AI in daily life must be disclosed. Yet the rules will expose how often AI is embedded in interactions and content that people may previously have regarded as ordinary digital services.

Four situations, rather than a universal AI warning

Article 50 of the AI Act focuses on specific transparency risks. Providers of AI systems designed to interact directly with people must make clear that the person is interacting with AI, unless that would already be obvious to a reasonably informed and attentive person in the context. A conventional text chatbot may therefore need an explanation; an overtly robotic automated interface may not need a repetitive warning.

Providers of generative systems also have to make synthetic audio, images, video and text detectable through machine-readable markings, where technically feasible and effective. This is primarily an infrastructure obligation. The marking may help a platform, journalist, researcher or verification tool identify the origin of content without necessarily placing a large visual notice before every viewer.

Separate duties fall on organisations that deploy certain systems. They must inform people who are exposed to emotion-recognition or biometric-categorisation systems. They must also visibly disclose deepfakes: realistic artificial or manipulated audiovisual material that could be mistaken for authentic depictions of people, places, objects or events.

A further provision concerns AI-generated or manipulated text published to inform the public on matters of public interest. The obligation is not a blanket label for all drafted text. It applies where the material has not undergone human review or editorial control and where a person or organisation does not accept editorial responsibility for it. That distinction attempts to preserve accountability in publishing while addressing the possibility of automated information products presented as independent reporting.

The practical meaning of “AI everywhere”

The regulation will nevertheless draw attention to AI’s expanding role. Customer support, voice interfaces, image editing, content production, fraud prevention and workplace monitoring are among the settings in which consumers or workers may come across disclosures. A notice can turn a background technical choice into a visible fact: the helpful agent is not human, the image is not a straightforward record of an event, or a call may be subject to an assessment of emotional cues.

This visibility matters because the social significance of AI often lies in context rather than in the mere use of software. A recommendation engine suggesting a song, for example, is not automatically the same type of encounter as a system purporting to be a human adviser. Likewise, basic editing assistance that does not substantially change supplied material or its meaning is treated differently from content that fabricates a credible event.

The rules are therefore designed to target deception, confusion and opacity rather than treat AI as inherently suspect. They aim to give people a cue at the point where an automated system could influence how they interpret another person, content or a consequential interaction.

The risk of disclosure fatigue

The principal challenge is whether transparency remains meaningful once it becomes routine. Europe has already experienced how well-intentioned online notices can deteriorate into a low-attention ritual. Cookie-consent banners made data practices more visible, but their prevalence also encouraged many users to click through them without reading.

AI disclosures could face the same problem. If a label is vague, ubiquitous or poorly timed, it may become visual clutter rather than useful information. A generic statement that “AI may be used” tells a customer little about whether an automated system is simply routing a request, generating an answer, assessing voice patterns or making a decision with a tangible effect.

The stronger approach is proportionality. Labels should explain the relevant fact in plain language and appear when the person can act on it: before the first substantive chatbot interaction, at the beginning of an emotion-analysis process, or alongside a deepfake likely to be mistaken for genuine footage. Visual disclosure must be understandable, while machine-readable provenance tools need to work reliably across services and platforms.

The Commission has published guidelines and a voluntary code of practice to help establish common approaches. These materials matter because the Act sets legal objectives, but firms still need workable answers to questions about technical marking, formats, interfaces, accessibility and evidence of compliance.

Enforcement will shape the first real-world outcome

The legislation provides for penalties of up to €15 million or 3 percent of a company’s worldwide annual turnover for breaches of the relevant transparency obligations, with the higher amount applicable. But a statutory ceiling is not the same as immediate, uniform enforcement.

National market-surveillance authorities will be central to oversight, while the European Commission’s AI Office has responsibilities for systems within its competence and the European Data Protection Supervisor oversees EU institutions. Different national capacities, varying sector practices and unresolved edge cases may initially produce uneven results across the bloc.

There is also a limited transition for certain generative systems already on the market before 2 August 2026. Their providers have until 2 December 2026 to meet the machine-readable marking and detectability requirement. Content created before the August deadline does not require retroactive labelling, although the Commission encourages disclosure where possible.

The first phase should therefore be viewed as the beginning of an operational test, not the finished form of European AI governance. Regulators will need to identify material non-compliance without reducing transparency to box-ticking. Companies will need to map systems that may have been adopted piecemeal by marketing, support, communications and operations teams. And users will need notices that answer a practical question: what exactly is the AI doing here?

Transparency is a foundation, not a verdict

The new rules do not determine whether a particular AI system is accurate, fair, safe or appropriate. A disclosed deepfake can still mislead, and a clearly identified chatbot can still provide poor service. Conversely, disclosure does not make an AI use improper.

Its purpose is more basic: to reduce the chance that people are unknowingly dealing with a machine or interpreting synthetic material as authentic. If the rules are applied with precision, they could build a more legible digital environment without turning every online activity into a sequence of warnings. If disclosures become indiscriminate, the signal may be lost in the noise.

Europe’s immediate discovery, then, may not simply be the scale of AI adoption. It may be the difficulty of designing transparency that people can actually notice, understand and use.

Sources