Why does EU law mandate AI content watermarks that any motivated attacker can simply remove?

EU AI Act Article 50, in force since August 2025, requires that providers mark AI-generated content with machine-detectable signals. A September 2026 arXiv analysis reviewed every deployed text watermarking scheme and found that all of them are removable through paraphrase, translation, or lightweight fine-tuning with at most moderate effort. Image and audio watermarks are similarly defeatable by post-processing that is imperceptible to the human eye. Open-source generators have no incentive to embed marks, and the technical literature has reached rough consensus that no scheme survives a motivated adversary at scale. Compliance teams are satisfying the legal checkbox, but downstream systems that depend on the provenance signal for fraud detection, disinformation labeling, or liability tracing receive no reliable signal in practice.