High-risk § 2.7.1

According to Article 6(3) AI Act, a provider of an AI system that would normally classify as high-risk pursuant to Article 6(2) AI Act may exempt that system from high-risk classification if any of the following conditions are met:

First, it is i mportant to clarify that the filter mechanism appli es only to AI systems that would normally be classified as high-risk pursuant to Article 6(2) AI Act. It does not apply to AI systems that classify as high-risk pursuant to Article 6(1) AI Act.

Second, the conditions listed in Article 6(3) AI Act are exhaustive, but alternative. Consequently, if the AI system falls within at least one of the conditions, it may be deemed not to pose significant risk of harm to the health, safety or fundamental rights of natural personsdespite the fact that certain risks might remain. There is no separate or independent assessment to determine whether the AI system poses a significant or any risk of harm besides those conditions. Whether an AI system may benefit from the filter mechanisms in Article 6(3) AI Act depends on a self-assessment by the provider. These conditions are explained in more detail in the subsections below . As Article 6(3) represents an exception from rules aimed at (among others) protecting fundamental rights, the conditions must be interpreted narrowly.The conditions for the filter must also be interpreted in the light of the first sub-paragraph of Article 6(3) AI Act that the system should not materially influence the outcome of the decision.

Third, the third sub-paragraph of Article 6(3) AI Act provides that an AI system referred to in Annex III shall always be classified as high-risk where the system performs profiling within the meaning of Article 4(4) of Regulation (EU) 2016/679² or Article3(4) of Directive (EU) 2016/680³ or Article 3(5) of Regulation (EU) 2018/1725⁴ (see more on profiling in Section 2.7.3 below).

Additionally, even if an AI system fulfils one of the conditions for the filter mechanism (for example, performs a narrow procedural task), such a system cannot benefit from that mechanism and will still be classified as high-risk if it forms part of a complex system where its combined intended purpose or joint outputs materially influence an individual decisionwithin a high-risk use case or where the AI system is part of complex interconnected systems, such as agentic AI systems (see para. 72 above).

Point (a) of Article 6(3) AI Act lists as the first condition under which the filter mechanism may apply that an AI system is intended to perform a narrow procedural task. As clarified in Recital 53 AI Act, such a system is considered to pose only limited risks. That recital lists as examples of such systems ‘an AI system that transforms unstructured data into structured data, an AI system that classifies incoming documents into categories or an AI system that is used to detect duplicates’.

This condition may therefore cover AI systems that are intended to categorise, change the format, structure or presentation of data, or change its metadata.

However, this condition does not apply to all categorisation systems. In particular, AI systems that perform a value judgement of datarelevant for decision-making, for example categorisation of input data as ‘useful’ or ‘less useful’ for the human assessment, or attributing a score or ranking to input data, are likely to have a broader impact on the assessment process and will therefore not be considered to perform only a ‘narrow procedural task’.

Point (b) of Article 6(3) AI Act lists the second condition under which the filter mechanism may apply, which is that the AI system is intended to improve the result of a previously completed human activity. Several cumulative elements must be present for this condition to apply:

• a human activity has been completed (e.g., human assessments or decisions);
• the completion of the activity led to a result; and
• the result is improved by the AI system (e.g., adjusted to achieve more clarity without changing the outcome or conclusion, etc.).

The fact that the human activity should be completed, and that the completion should lead to a result, means that the AI system should not replace, nor autonomously perform, the human activity. Recital 53 AI Act specifies that such a n AI system ‘provides only an additional layer to a human activity with consequently lowered risk’, referencing as an example an AI system that is applied ex post to improve a previously drafted text.

The notion ‘improve’ in this context means that the application of the AI system may lead to changes, but these changes should not revert or lead to a full revision or replacement of the previously completed human activity. The EU legislature deliberately chose the formulation ‘improve’, rather than ‘review’, to make this distinction and to highlight that the AI system must not be intended to provide a materially different result to the one previously performed by a human activity, but to verify or refine that activity. Any improvement introduced by the AI system should not change the rights, protection, legal or economic position of the persons who may be impacted by the system’s output or by decisions takenon the basis of that output.

Point (c) of Article 6(3) AI Act lists the third condition under which the filter mechanism may apply, which is that the AI system is intended to detect decision-making patterns or deviations from prior decision-making patterns and is not meant to replace or influence the previously completed human assessment. Recital 53 lists as an example of such an AI system a system that is applied to detect inconsistencies or anomalies in the grading pattern of a teacher.

The notion ‘decision-making pattern’ should be interpreted narrowly. Decision-making patterns are patterns or trends derived from past data that predict or analyse how decisions are taken over time on a specific subject matter.

Unlike the other conditions, which exclude AI systems from high-risk classification only if the system is restricted to procedural or preparatory tasksor improvements, the scenario listed in point

First, the human assessment must have been completed, meaning that all necessary steps in the evaluation process must have been taken.

Second, the AI system can only perform an ex-post comparative assessment, which involves determining whether a current decision aligns with previous patterns or decisions. It must not infer relevant criteria from previous decisions and propose a new assessment based on those criteria.

Third, the AI system must not be intended to be used to replace or influence the previously completed human assessment without proper human review. ‘Proper’ means that the human review applied after the AI system’s assessment should be meaningful and comprehensive, and it should consider, together with the result provided by the AI system, the various elements taken into account during the previously completed human assessment. Therefore, the output of the system can highlight discrepancies with previously made decisions, but cannot modify the input or alter the underlying reasoning. In order to allow proper human review, the system should not only flag, but also provide information on the discrepancies. It should however merely inform (e.g. through analytical tools used for quality assurance and statistics) and not replace the human assessment.

Point (d) of Article 6(3) AI Act lists the fourth condition under which the filter mechanism applies, which is that the AI system is intended to perform a preparatory task to an assessment relevant for the purposes of the use cases listed in Annex III.

The notion ‘preparatory’ in point (d) of Article 6(3) AI Actrefers to tasks that occur prior to the actual assessment process. This specification is important, since it differentiates the preparatory task exemption in point (d) from the narrow procedural task exemption in point (a) of Article 6(3) AI Act, which can occur during the assessment process so long as it is clearly defined and limited in scope. At the same time, if the narrow procedural task occurs prior to the assessment and is preparatory in nature, both points in letters a) and d) may apply.

As specified in Recital 53 AI Act, the preparatory nature of the task should be such that the ‘possible impact of the output of the system [is] very low in terms of representing a risk for the assessment to follow’. This aligns to the objectives of Article 6(3) AI Act, which sets out the conditions under which AI systems should not be considered tolead to a significant risk of harm, in particular because they do not materially influence the decision-making. Recital 53 AI Act specifies that an AI system that does not materially influence the outcome of decision-making should be understood as an AI system that does not have an impact on the substance, and thereby the outcome, of decision-making, whether human or automated.

Recital 53 AI Act provides as examples AI systems performing preparatory tasks such as indexing, searching, processing, and linking. Such tasks add structure to the inputs that will be used in the later assessment producing an outcome, but they do not provide themselves an assessment leading to an outcome.

The decisive factor in determining whether a task is ‘preparatory’ is the task’s role in the decision-making process and proximity to the final human decision. If a task is deemed preparatory, the filter mechanism in point (d) of Article 6(3) AI Act may be applicable, regardless of the specific field in which the system will be used.

If an AI system’s output is intended to inform a human operator’s assessment, it may be considered to perform a preparatory task only if the system’s output is a general input or factor, for example supplementary information that supports an operator’s decisionmaking process. - Where the AI system is intended to produce a specific recommendation or evaluation of the case , it plays a decisive role in the assessment or the decision and can therefore not be considered to perform a preparatory task within the meaning of point (d) of Article 6(3) AI Act.