Bottom line: Tax authorities use real tax data not only for AI training but also during ongoing operations for validation, which requires data protection and EU AI Act compliance.
German tax authorities rely on actual citizen tax data not only in training for their AI models, but these data are also continuously drawn upon for comparative analysis during operations.
For developing AI systems for tax audits and administration, tax authorities use real tax data as training data. This data inventory is not, as is common practice in many AI projects, limited to training but is also shared during productive operations. The rationale: AI models require continuous comparative analyses with real tax data to ensure their accuracy and reliability.
For compliance officers, this creates a complex tension between data protection requirements and administrative purposes. The permanent use of real citizen data for AI model validation falls under GDPR provisions and is subject to the assessment standard of the EU AI Act if the systems are to be classified as “high-risk AI systems” – which is likely in the analysis of tax data for administrative audits.
Data protection compliance requires that authorities can demonstrate an appropriate legal basis and that data processing is necessary. At the same time, they must demonstrate that technical and organizational measures provide appropriate data protection. The question of whether anonymized or pseudonymized data would be sufficient is rejected by many authorities with reference to technical necessity – which increases compliance requirements.
Source: www.golem.de · Published June 2, 2026
Lumi AI News — AI-assisted curation pursuant to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.2.9.