Why this object hangs off that object

Article 4 prescribes no form. The Commission confirms that no certificate is required, no obligation to measure knowledge exists, no training is mandatory and no governance structure is prescribed.

Hangs off: Article 4: AI literacy

Placed against the official source | guidance

The duty to take measures also covers contractors, service providers and sometimes clients. Supervision lies not with the AI Office but with national market surveillance authorities, enforcing since 2 August 2026.

Hangs off: Article 4: AI literacy

Placed against the official source | guidance

An AI agent that makes bookings, handles correspondence or concludes contracts must identify itself as AI and state on whose behalf it is acting, including towards the person instructing it.

Hangs off: Article 50: transparency

Placed against the official source | guidance

For deep fakes in evidently artistic, creative, satirical or fictional work, the disclosure must not hamper the work, but it remains mandatory. Where the content mixes in an informative or commercial character, the standard label applies.

Hangs off: Article 50: transparency

Placed against the official source | guidance

Signing the code of practice on transparency of AI-generated content is voluntary and not signing is not in itself non-compliance. Those who do not sign must demonstrate compliance by other means, including a gap analysis against the code.

Hangs off: Article 50: transparency

Placed against the official source | guidance

Whoever publishes a deepfake must apply a label perceivable by humans. The watermark or metadata supplied by the provider does not discharge that duty.

Hangs off: Article 50: transparency

Placed against the official source | guidance

The notice that a person is dealing with an AI system must be given at the latest at the time of the first interaction. That moment applies per natural person, not once per system.

Hangs off: Article 50: transparency

Placed against the official source | guidance

No label is needed where the AI text has been substantively reviewed by a human and someone holds editorial responsibility. A spellcheck or a written editorial policy is not enough, and any AI intervention after sign-off voids the exception.

Hangs off: Article 50: transparency

Placed against the official source | guidance

A line in the terms and conditions, a hidden metadata mark or a vague word like 'assistant' will not do. The guidelines name five methods that are insufficient when used alone.

Hangs off: Article 50: transparency

Placed against the official source | guidance

Only the machine-readable marking under Article 50(2) has a transitional period for systems already on the market. Legacy content need not be labelled retroactively, but legacy text you publish now must be.

Hangs off: Article 50: transparency

Placed against the official source | guidance

The exception for obvious AI interaction must be interpreted restrictively. It applies only where almost no doubt remains for an average member of the intended and reasonably foreseeable audience.

Hangs off: Article 50: transparency

Placed against the official source | guidance

Spellchecking, formatting and minor image corrections fall outside the marking obligation. An AI summary, rewrite or face replacement falls inside it.

Hangs off: Article 50: transparency

Placed against the official source | guidance

The labelling duty for AI text applies only where three elements coincide: the text is published, is intended to inform the public, and concerns a matter of public interest.

Hangs off: Article 50: transparency

Placed against the official source | guidance

Not every modification makes you a provider. The indicative threshold is a modification using more than a third of the original model's training compute, and your obligations are then limited to the modification.

Hangs off: Article 53: GPAI model providers

Placed against the official source | guidance

Annex I contains no list of high-risk products but an exhaustive list of harmonisation legislation. The AI Act does not extend the scope of that product legislation and does not change a product's risk profile.

Hangs off: Articles 43-49: conformity assessment, CE and registration

Placed against the official source | guidance

For products under Section A of Annex I the full set of high-risk requirements applies. For Section B only Article 6(1), Articles 102 to 109 and Article 112 apply.

Hangs off: Articles 43-49: conformity assessment, CE and registration

Placed against the official source | guidance

An AI system only becomes high-risk under Article 6(1) when two conditions are met at the same time: the system is itself a regulated product or a safety component of one, and that product must undergo third-party conformity assessment.

Hangs off: Articles 43-49: conformity assessment, CE and registration

Placed against the official source | guidance

A system falling within Annex III can still stay outside high-risk if one of four grounds is met and the system does not perform profiling. The provider makes that call itself, must document the assessment before placing the system on the market, and must register the system in the EU database.

Hangs off: Articles 43-49: conformity assessment, CE and registration

Placed against the official source | guidance

An AI system can be high-risk in two ways. Either it is itself a product, or a safety component of a product, covered by the Annex I product legislation and that product must undergo third-party conformity assessment (Article 6(1)). Or it falls within one of the use cases listed in Annex III (Article 6(2)). The two routes have their own criteria and their own application dates.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

If you market a system broadly without consistently limiting its application, high-risk use cases will be read into its intended purpose. Merely stating in the terms of service that high-risk uses are excluded is insufficient where the rest of your presentation in fact enables or promotes such uses.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

Classification answers one question: which Chapter III obligations apply. It says nothing about whether the use itself is lawful. Prohibited practices, data protection, consumer law, product safety and national law continue to apply in full.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

You cannot avoid classification by splitting a high-risk function into separate modules. The draft assesses the combined configuration.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

The exception is a provider self-assessment, but not a free pass. The draft sets out four mandatory components of the assessment and links registration and supervision to it.

Hangs off: Articles 43-49: conformity assessment, CE and registration

Placed against the official source | guidance

A system that falls within an Annex III use case can still escape high-risk classification if it meets one of four conditions. The draft guidelines make clear this is not a broad escape route.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

Human involvement does not change the intended purpose and therefore has no effect on classification under Article 6(2). Human oversight is a compliance requirement for high-risk systems, not an escape from the classification.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

Human oversight is a compliance requirement for high-risk systems, not a way to escape classification. Human involvement only counts when determining what task the system performs.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

The second condition requires a completed human activity with a result that the system only refines. A materially different outcome does not qualify.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

Intended purpose determines whether a system is high-risk. That purpose is set not only by the technical documentation but also by the instructions for use, promotional materials, sales materials and statements by the provider. Reasonably foreseeable misuse falls by definition outside the intended purpose.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

Before classification comes into play, the system must meet the definition of an AI system in Article 3(1). Not every software application and not every automated decision-making system falls within the AI Act.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

Systems assessing only legal persons fall outside the relevant Annex III use cases. Self-employed people and sole traders do count as natural persons, however.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

The preparatory task under Article 6(3)(d) precedes the assessment. Once the system evaluates the specific case or makes a recommendation, the exception is gone.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

As soon as the system profiles, the Article 6(3) exemption is ruled out. The draft guidelines give three cumulative elements against which you test this.

Hangs off: Annex III: high-risk AI

Placed against the official source | guidance

For classification purposes only the definition of safety component in Article 3(14) AI Act counts. Definitions of the same term in sectoral product legislation are not relevant here.

Hangs off: Articles 43-49: conformity assessment, CE and registration

Placed against the official source | guidance

An AI system is a safety component via two alternative routes. Either it is intended to fulfil a safety function, or its failure or malfunctioning endangers health, safety or property.

Hangs off: Articles 43-49: conformity assessment, CE and registration

Placed against the official source | guidance

The AI Act allows you to add AI-specific risks to your existing risk management and quality management systems, so that you can work within a single compliance framework.

Hangs off: Article 9: risk management system

Placed against the official source | guidance

Article 6(1) applies regardless of whether the AI system is embedded in the product or placed on the market independently. A software update, add-on or remote service can therefore be high-risk in its own right.

Hangs off: Articles 43-49: conformity assessment, CE and registration

Placed against the official source | guidance

Distributors, importers, deployers and other third parties can take on provider obligations. That happens when you put your own name or trademark on a system, when you make a substantial modification to a high-risk system, and when you change the intended purpose of a non high-risk system so that it becomes high-risk.

Hangs off: Articles 43-49: conformity assessment, CE and registration

Placed against the official source | guidance

The fact that a manufacturer may opt for internal control based on harmonised standards does not affect the high-risk classification. Classification follows the level of enhanced scrutiny required by the product legislation, not the procedure chosen.

Hangs off: Articles 43-49: conformity assessment, CE and registration

Placed against the official source | guidance