← All questions
Security & SafetyMedium· Copyright· Extractability· Measurement

If a model is probed and 0.07% of sampled documents are...

A weak answer treats the 0.07% figure as a static system property. A strong answer recognizes that extractability depends on the probe length (k). By re-running the probe at a more realistic k=2,000, you can gain a better understanding of the actual risk. This is a cheap experiment—costing roughly 3 GPU-hours—that provides the empirical data necessary to make an informed decision about shipping licensed content.

Code playground (Python)