EU AI Act Article 50 provenance signals and technical limits

Know which provenance signals a file carries, which transformations strip them and how to keep Article 50(2) marking intact through your publishing pipeline.

What the law requires

Article 50(2) expects marking that is robust and reliable as far as technically feasible. Provenance lives in metadata and manifests, and routine compression, resizing, screenshots and reposts remove it. Knowing where a signal breaks is the difference between a mark that exists and a mark that works.

EU AI Act Article 50(2)MACHINE READABLE MARKINGPROVIDER / externally testable
EU AI Act Article 50(5)CLARITY ACCESSIBILITY TIMINGBOTH / externally testable

What proves it

isGenAI records the observed mechanism and its tested persistence for each product, with the transformation suite and result stored as evidence. No universal robustness percentage is ever claimed; every number traces to a tested artifact.

Official source text, public technical observations, vendor-submitted evidence and external inference are always kept apart. A missing public signal is recorded as NOT OBSERVED, a technical fact, never as a legal finding.

Observations from public scorecards

0 published observations for this topic under isgenai-methodology-2026-08-16.v1, each a dated technical observation with linked evidence. Legal applicability stays with counsel.

0Products with public scorecards
0Relevant controls verified
0Partially observed
0Not observed in tested flow

Observations publish the moment they clear the evidence gate. Output marking is verified on generated artifacts and disclosure inside the real user workflow, never from a marketing page alone.

Related public findings

Findings appear here as soon as a published scorecard records one for this topic.