Quantifying re-identification risk requires moving past static definitions. Traditional laws often view metadata as non-sensitive unless it contains a name or ID. However, AI changes this math. Modern LLMs act as high-dimensional pattern matchers. They can take seemingly anonymous logs—like timestamps, IP fragments, or geolocation pings—and cross-reference them with public datasets to unmask individuals.
To measure this, we should use Differential Privacy metrics. Specifically, we can calculate the $epsilon$ (epsilon) privacy loss. This mathematical value measures how much a single individual's presence in a dataset changes the output of an AI model. A higher $epsilon$ suggests a higher risk of leakage. We can also apply k-anonymity or l-diversity tests to see how many distinct individuals share the same metadata fingerprints. If an AI can narrow a person down to a tiny subset of the population using only metadata, the risk is high.
The OpenAI-Australia incident shows that the legal distinction between sensitive and non-sensitive data is currently broken. If a machine can turn a 'non-sensitive' log into a personal identifier, the law must treat that log as sensitive data from the start. We need to transition from labeling data based on what it is to labeling it based on what it can become when processed by an AI.
