# Romania - ANAF algorithm for the selection of tax audits
Romania, 2021-2023 | Severity: N2

## What Happened
Romania's tax authority, ANAF, deployed an automated risk-scoring and data analysis algorithm to select taxpayers (both individuals and corporations) for mandatory tax audits. The algorithm was criticized for operating as an absolute "black box," lacking transparency and public documentation regarding its risk indicators. Investigations revealed that the system frequently targeted small and medium-sized enterprises (SMEs) with high false positive risk scores due to minor transactional discrepancies, while failing to detect sophisticated, large-scale tax evasion schemes by multinational corporations.

## How MEG 1 acted
The automated risk-scoring system failed the Dynamic Accuracy Index (DAI) audit due to its lack of transparency, high error rate, and bias against small taxpayers, violating the requirements of procedural fairness.

## MEG 2 Analysis - Legal Framework
* **Agent level:** N2 (management) - automated decision-support and risk-profiling system used by a public authority to execute tax audit selections; directly impacts the financial and operational rights of taxpayers.
* **Cause (6.1):**
  * **(a) system defect** - design defect in the risk-scoring model that lacks transparency and exhibits systemic bias against specific taxpayer demographics (design defect).
  * **(b) autonomous decision error** - the system's decision to flag taxpayers for audits based on opaque, unverified risk scores.
* **Attachment of liability:** ANAF as the operator is fully liable for the unfair auditing selections caused by deploying an unverified, opaque system.

## Procedural Mechanism
* **6.2 (architectural human confirmation):** Taxpayer audit selection decisions must require architectural human confirmation. The AI can only suggest risk levels; a human tax officer must verify and document the justification before initiating an audit.
* **7.1 (independent forensics):** The risk indicators used, the taxpayer data, and the model's decision path must be securely logged in an unalterable ledger to allow for judicial and human rights audits.
* **9.4 (least privilege):** The profiling system was granted significant influence over tax audits without public validation or independent oversight layers, violating safe operational boundaries.