By Valentina Remonato, StandICT.eu 2029 Fellow, UNI delegate to CEN/CLC/JTC 21 (WG1/SAG and WG4)
What does a person need to understand before relying on the output of a high-risk AI system or exercising human oversight of that system? This question was the starting point for a methodology connecting deployer tasks and information needs with the provider’s technical determinations across the AI Trustworthiness Framework (prEN 18229 series). Transparency, in this approach, should be assessed by the decisions and actions it enables.
Article 13(1) of the AI Act establishes a functional transparency requirement: high-risk AI systems must be designed and developed so that their operation is sufficiently transparent to enable deployers to interpret outputs and use them appropriately. The type and degree of transparency must also support compliance with the relevant provider and deployer obligations under Chapter III, Section 3. The information listed in Article 13(3) constitutes minimum content for the instructions for use, subject to the stated applicability qualifiers, not necessarily sufficient transparency. The distinction is between documentary completeness and functional sufficiency: addressing every listed heading does not, by itself, establish that the deployer can use the information to perform the relevant task.
The method begins by identifying what the deployer, or a person acting on its behalf, needs to do: interpret an output, monitor the system, select an oversight measure, intervene or decide not to rely on a result. These tasks are identified from the intended purpose, intended domain of use, conditions of use and intended category of deployer. The provider does not need to know the internal organisation of every future customer, but the method uses the anticipated deployer characteristics and context of use that the provider must consider under Article 9(5).
For each task, the question is what the deployer must understand to perform it in an informed manner. This is the deployer information need. It is distinct from the transparency mechanism: the need identifies what must be understood; the mechanism determines how that understanding is enabled.
Consider a deployer deciding whether to rely on a particular output. Human oversight measures must enable the natural persons assigned oversight, as appropriate and proportionate, to decide not to use a system or to disregard, override or reverse its output in a particular situation (Article 14(4)(d)), but aggregate performance metrics alone cannot establish whether the validation and testing evidence applies to the case at hand. The methodology addresses this gap through reliance conditions: the input, operating and contextual conditions relevant to assessing whether that evidence applies. Meeting them supports the relevance of the evidence but does not guarantee that an output is correct; where a condition is not met, or cannot be assessed, the deployer needs to be able to recognise that limitation or uncertainty. The provider traces these conditions to its determinations under the Accuracy and Robustness parts of the framework, selects appropriate transparency mechanisms and verifies whether they enable the deployer to fulfil the identified information needs.
The same methodology was adapted for a contribution to the proposed Technical Report Safeguarding Fundamental Rights in Trustworthy AI, continuing the analysis to the human oversight function the information supports and the fundamental rights safeguard to which that function contributes. This application is analytical rather than normative. Within JTC 21, the work was also presented in cross-Working Group discussions and developed into contributions concerning Accuracy, Robustness and the Fundamental Rights Technical Report.
The work builds on the task-based transparency architecture developed during my first StandICT.eu fellowship, which grew out of an attempt to determine whether preserving a recognisable space for human agency at the input stage of AI systems could contribute to trustworthy and human-centric AI. The second fellowship extended that architecture across the Trustworthiness Framework.
Under Article 40(1) of the AI Act, the presumption of conformity arises from conformity with harmonised standards or relevant parts thereof whose references have been published in the Official Journal of the European Union, and applies only to the extent that those standards or parts cover the applicable requirements. A standard limited to checking for the presence of the information listed in Article 13(3) may therefore not fully cover the functional transparency requirement in Article 13(1). The methodology proposes one verifiable way to specify existing Article 13 requirements, without introducing new legal obligations.
The standardisation activity described in this article was carried out within CEN/CLC/JTC 21, in WG1/SAG and WG4. This activity is a beneficiary of EU funding via StandICT.eu 2029 Programme. StandICT.eu 2029 has received funding from the European Union’s Horizon Europe programme under Grant Agreement No. 101213612.
The views expressed are those of the author.