Identify
Spot the personal information and the part genuinely needed. Detection works on known fields and formats; an identifier spelled out in a paragraph calls for a read-through.
Personal data
Reduce the personal data an AI needs to see.
For many tasks, personal data is not needed at all. Look into protecting it before sending anything to a model.
A simple exampleKeep what is useful.
Keep what is useful.
Limit the personal data.
A made-up illustration of a substitution. The real mechanism depends on the processing and its configuration.
Made-up original document
Camille Martin is requesting a return for order ORD-0042. Their device will not start.
Context with identifiers substituted
[PERSON_1] is requesting a return for order [ORDER_1]. Their device will not start.
Substituting these fields illustrates pseudonymisation. If re-identification remains possible, this is not irreversible anonymisation.
A path to define
Protecting personal data must stay useful to the task.
Protect
Look into detecting and replacing identifiers, within the options available. The method is decided task by task: an internal summary and published content do not call for the same level.
Check
Review what remains and the scope for re-identification. A check on a real sample is worth more than a promise about the configuration: that is where the edge cases show up.
Different notions
Personal data: choose the words that match the processing.
Minimisation
Reducing the data used to what the task strictly requires.
Pseudonymisation
Replacing identifiers while potentially keeping re-identification possible.
Anonymisation
Preventing re-identification effectively and lastingly: a requirement to assess rigorously.
Frequently asked questions
What you want to know about personal data.
Does automatic detection catch everything?
A check is still needed. Indirect information, an unusual context or a poor-quality document can leave scope for identification.
Can this be used on every document?
The formats, the categories of information, the quality of the extraction and the expected result all have to be checked. The demonstration is where the terms that apply to your case are framed.
Let us start simply
Let us talk about your sensitive use case.
We can start from a made-up example and define the protection requirements.
