Research ·
Can someone be recognised? How data providers can assess identification risk
Before you share record-level data, work out who will receive it and what they could match it against. Then choose your privacy tools.
By Nakato · 7 min read
Insights
What we learn putting regulated data to work with AI.
Articles from the team: what a ruling means for sending data to an AI model, how a method holds up under test, and what we have built and why.
Research ·
Is your de-identified data still any good? Evaluating dataset usefulness
Run the same analyses on the de-identified copy and the original, then compare the answers – and check the privacy risk in the same review.
8 min read
Engineering ·
How can teams test for bypasses around an AI de-identification layer?
A de-identification layer protects only the traffic routed through it. Test every route a prompt can take, not just the detector.
5 min read
Regulation ·
How can teams safely send customer data to an LLM?
Approve one task and one data flow, keep the key that reverses the twins in your own environment, and test the detector on your own text.
7 min read
Regulation ·
Swapping names isn’t enough: getting pseudonymised data approved for an LLM
Pseudonymised data stays personal data while you hold the key. A committee approving it for an LLM needs written evidence on seven points.
6 min read
Research ·
How do masking, tokenisation and pseudonymisation change an LLM prompt?
The right de-identification method depends on what the prompt needs.
6 min read
Research ·
Masking, tokenisation and semantic pseudonymisation for LLMs
Masking deletes values, tokenisation swaps in labels and semantic pseudonymisation gives realistic stand-ins the model can still reason with.
9 min read