Home / Writing / You can’t delete what you can’t find
You can’t delete what you can’t find
Why a GDPR erasure request is, underneath, an entity resolution problem, and why the same gap follows you into AI training.
A patient asks a hospital to erase her data. GDPR gives them one month.
Her records sit in four systems. Different spellings, a typo in the surname, two digits swapped in a birth date, an empty postcode. Three systems store tables. The fourth stores FHIR JSON. There is no shared ID.
nora keller 1984-03-12 69003KELLER, Nora J. 12/03/1984 69003nora kellar 1984-03-21 (empty){"family": "Keller", "given": ["Nora"], "birthDate": "1984-03-12"}Finding all of them is not a legal problem. It is an entity resolution problem, and I have met it in every data role: company registries, scraped property listings, patient records.
Every product inherits it
- Search: the same item appears five times and pushes better results off the first page.
- Support tickets: one customer, three accounts, and no agent sees the full history.
- CRM and billing: the same client contacted twice, invoiced twice, or missed.
- Dashboards: “customers” counted per record, not per person, so every KPI is off.
It follows you into AI training
- The same person counted four times gives the model a skewed picture.
- Duplicates leak between training and test data, so scores look better than they are.
- A record you missed gets trained on, and removing it from a model is far harder than deleting a row.
What works in practice
- Resolve identities first: one per person, linked to every source record.
- Keep a human in the loop for borderline matches. A wrong merge is a privacy incident too.
- Protect the identity map itself: restrict access, encrypt it, log every lookup.
Clean identities are not a data-team chore. They are the foundation every product, and every model, stands on.
One catch: scale
Comparing every record with every other grows with the square of your data. Ten times more records, a hundred times more comparisons. The brute force tab of the matching demo shows where that bottleneck appears.
And the erasure tab shows what an exact-match delete leaves behind.
First published on LinkedIn.