Data Re-identificationin in the Age of AI

In this talk, I focused on a simple question: Can data that looks anonymous still be linked back to a real person?

We often assume that removing names, email addresses, or other direct identifiers is enough to protect privacy. In practice, this is not always the case. Information such as location traces, timestamps, purchases, search histories, movie ratings, or patterns of online behavior may seem harmless on their own. When several of these pieces are combined, however, they can reveal much more about an individual. This process is known as data re-identification.

Re-identification is not a new problem, but AI makes it easier to perform at a much larger scale. AI systems can process large amounts of data and connect information from different sources much faster than manual analysis. Public profiles, social media posts, photos, location data, and chat logs can all provide small clues about identity. The way we use AI tools may also create additional data traces, as people often paste text, upload documents, or share logs that contain personal or confidential information. As a result, AI can both make existing data easier to connect and create new sources of information that may contribute to re-identification.

Real-world cases show how re-identification can happen in practice. The AOL search-log release in 2006 showed that supposedly anonymous search histories could be linked back to a real person. The Netflix Prize dataset also demonstrated that anonymous movie ratings could be matched with information from other sources. Location data creates similar risks because repeated phone traces can reveal homes, workplaces, and daily routines.

These examples show that removing names alone does not necessarily make data anonymous. What matters is whether the remaining information can still be linked back to a person. This idea is also reflected in the GDPR: if an individual can still be identified using information that is reasonably available, the data may still be considered personal data.

There is no single technique that can completely solve this problem. Different approaches can help reduce the risk, including data minimization, pseudonymization, aggregation, k-anonymity, differential privacy, federated learning, and re-identification risk testing.

The main takeaway from the talk is that privacy depends not only on what one dataset contains, but also on what can be learned by combining it with other information. As AI makes linking and inference easier, evaluating re-identification risk becomes increasingly important whenever data is collected, shared, or released.