- Accuracy and evaluation methods in the context of NLP systems
Artificial Intelligence
- New standard development
- Towards EU AI Act Compliance: Data Quality and Bias in AI systems
As co-founder of an AI start-up and a consortium partner in research projects involving other SMEs, I am very aware of the importance of clear and simple guidelines that can help smaller organisations comply with regulations. SMEs often lack the resources to identify appropriate methods for compliance through extensive research. This project should therefore have a positive impact on SMEs ability to govern and use data correctly, and to produce and assess bias in AI systems, thus simplifying compliance with the AI Act requirements.
The bias and datasets standards support the European Commission’s standardisation request two on data and data governance. Such a standard is beneficial to consumers, who can rely on quality products and be protected from discrimination, as well as to regulators and auditors, who will benefit from the fact that AI system developers adopt more systematic approaches to data governance and bias risk management. My work on other standards, such as 23282, supports the SR on accuracy, while 12792 supports the standardisation request 4 on transparency. Transparency helps downstream providers and system integrators build safe systems, helps deployers use systems properly, and assists auditors and regulators in carrying out their oversight duties. Finally, just as I have benefited from the experience of other standards experts in learning about standardisation, I feel it is my turn to give back. Whether as project editor, as head of delegation attending plenaries, or as expert participating in working group discussions, I am happy to share knowledge and experiences with those who are new to standardisation.
- New standard development
- Standardising bias management for AI systems in support of the AI Act
Artificial Intelligence
The standard prEN 18283 on bias management is one of the first standards to provide detailed requirements for detecting and managing bias in AI systems. It is intended to benefit more than just the AI Act, and to provide guidance to other developers and deployers of algorithmic systems as well.
Such guidance is useful to industry, and in particular to SMEs, who may not have the expertise or resources to conduct their own research into appropriate methods for data quality and bias management. The immediate goal of the standard is to support compliance with AI Act requirements. More broadly, the objective is to simplify the practical implementation of data quality and bias controls - thus improving the overall quality and reliability of AI products produced in the EU, and made available on the EU market. This makes quality and fairness a USP of EU AI products, while at the same time benefiting EU consumers and society in general, by ensuring that AI systems have been trained and tested according to best practice, and with appropriate bias mitigation measures and Fundamental Rights protections in place.
- New standard development
-
Custom funding impact (2nd Open Call 2029)This short-term fellowship supported my contribution to the development of prEN 18283, "Concepts, measures and requirements for managing bias in AI systems". This new standard being developed by CEN/CENELEC JTC 21 "Artificial intelligence" within the WG3 "Engineering aspects".
Value of Research
Title & Organisation Name: Co-founder and CTO, leiwand AI gmbh
Country: Austria
