Article 50 II of the EU Artificial Intelligence Act mandates dual transparency for AI-generated content via human-understandable and machine-readable labels (arXiv:2603.26983; EUR-Lex 32024R1689). The cited preprint uses synthetic data generation and automated fact-checking as diagnostic cases, showing that post-hoc labeling cannot satisfy requirements due to non-deterministic LLM outputs and iterative workflows. Compliance issues are compounded by the assistive-function exemption failing to apply when systems assign truth values.

A 2023 study on LLM watermark robustness demonstrates that machine-readable marks are fragile under standard data processing and risk being learned as spurious features, corroborating the preprint's findings on synthetic data paradoxes (arXiv:2301.10226). The original paper correctly flags absent cross-platform marking formats and misalignment with probabilistic model behavior yet understates parallels to GDPR technical feasibility shortfalls that delayed enforcement. Missing guidance for heterogeneous user expertise further widens the gap.

These architectural deficiencies in the August 2026 enforcement timeline indicate that transparency cannot be retrofitted, potentially undermining the Act's accountability mechanisms across high-risk and general-purpose AI systems as noted in the primary source.