Germany: Adopted BSI guide on explainable Artificial Intelligence in an adversarial context

On 6 January 2025, the German Federal Office for Information Security (BSI) adopted a white paper serving as a guide addressing the explainability of artificial intelligence (AI) in adversarial contexts. The document focuses on the limitations of Explainable Artificial Intelligence (XAI), particularly post-hoc methods used to interpret black box AI models. The white paper identifies three challenges, namely the disagreement problem, manipulation risks, and fairwashing. Solutions to these problems include standardising explanation methods, employing robust audits such as white-box or outside-the-box access), and developing new manipulation-resistant techniques. Furthermore, the white paper proposes detection strategies, such as outlier analysis and statistical comparisons, to identify inconsistencies and prevent deceptive practices in AI assessments. The objective of the white paper is to inform the development of reliable assessment procedures and digital consumer protection measures in line with the requirements of the European Union AI Act.