personal data

In depth

The decisive criterion

The decisive criterion is identifiability, and it is assessed in the light of the means that can reasonably be brought to bear to trace back to the person. This gives the notion a relative character that is underestimated: one and the same file can be personal for whoever holds the correspondence table and not be personal for a recipient who has no way of reaching it. Indirect identification is the most frequent source of error, because innocuous attributes taken separately become identifying once combined, especially when the population concerned is small. A municipality, a job title and an age bracket are sometimes enough to designate one person and one only, without any name appearing.

The points of exposure

In an AI system, the question arises in more places than the original database. What a person writes in a prompt, what a system retains from one conversation to the next, the documents indexed so as to be retrieved, the technical logs kept in order to understand an incident and the answers produced all form places where the data circulates. Those logs are in fact the most common blind spot, because they are taken for mere machine traces when they attach dated actions to named people. To this is added a question specific to training: when a model has been built on personal data, knowing whether its parameters still carry any has no general answer and is examined case by case. The practical consequence is that personal data has a life cycle inside the system, and that a useful inventory follows that cycle rather than the list of applications.

The stricter regime

Some categories receive stricter treatment, because they expose more: health, opinions, beliefs, origin, orientation, trade union membership, biometric data used to identify. AI adds a difficulty of its own, that of inference: a system can deduce an attribute of this nature from information that does not fall under it, which brings the processing into the strictest regime without anyone having collected anything of the sort. The second trap is claiming anonymity too quickly, often asserted after a simple deletion of names when cross-referencing remains possible. The third is reasoning by tool: an organisation that classifies its rules by software finds itself helpless as soon as a piece of software changes, whereas a classification by category of data outlives the tools.