- Standardising Human Input Space in AI Systems to Safeguard Fundamental Personal Expression
Artificial Intelligence
Making the input conditions established by the provider explicit and traceable reduces the information asymmetry between providers and deployers. Many deployers in the European market are SMEs. Under the AI Act they must use high-risk systems in accordance with the instructions for use and assign human oversight, but they can only do so if the conditions under which the validated operational domain holds are documented and verifiable. Expressing that within the existing standards, rather than as a separate family of requirements, keeps the burden proportionate. The effects I would expect over time are lower assessment costs for SMEs deploying AI and fewer barriers to adoption caused by uncertainty about what a system can be relied upon to do. Documenting the human determinations made upstream also keeps the system traceable to the choices that shaped it, which is the condition under which the fundamental rights behind those choices, including freedom of expression, remain meaningful.
- New standard development
-
Within CEN/CENELEC JTC 21 WG4, contribution to prEN 18229-2, AI Trustworthiness Framework, Part 2: Transparency. Contributions were also made to prEN 18229-3 (Human Oversight) and to the CEN/CLC/JTC 21 WG1 Technical Report JT021052, Overview and Architecture of Standards in Support of the EU AI Act.
- Designing AI Transparency: Information Needs as a Condition for Fundamental Rights Protection
Artificial Intelligence
The methodology is particularly relevant to SME deployers of high-risk AI systems. To rely on an output or exercise oversight, a deployer needs to know the input, operating and contextual conditions under which the provider’s validation and testing evidence applies to the case at hand. Where these reliance conditions are not stated, or not traceable to the provider’s accuracy and robustness determinations, the SME may over-rely on the system or incur additional technical assessment costs. Assessing transparency by the tasks it enables, rather than by mere documentary completeness, gives SMEs instructions for use they can act on and makes the information comparable between suppliers. The effects I would expect are lower onboarding costs for SME deployers, a clearer basis for comparison at procurement and for the allocation of responsibility between provider and deployer, and more effective human oversight in organisations without dedicated AI expertise.
- New standard development
-
Custom funding impact (2nd Open Call 2029)This fellowship supports my work at CEN/CLC/JTC 21/WG 4 (Foundational and societal aspects) and, for cross-WG coordination, WG 1/SAG.
I contributed to the the development of the following standards:
- prEN 18229-4, AI trustworthiness framework - Part 4: Transparency (numbered Part 2 until the renumbering of the series at the October 2026 plenary).
- prEN 18229-6 (Accuracy, formerly Part 4) and prEN 18229-7 (Robustness, formerly Part 5), to the CEN/CLC/JTC 21 WG4 Technical Report “Safeguarding Fundamental Rights in Trustworthy AI” (work item JT021053), and to WG1/SAG through the cross-WG concept note WG1 N631.
Read more
