Remediation cannot erase data already collected
Aliases: data irreversibility · residual data · data persistence
What it is
The irreversibility of data collection means that stopping future collection cannot reliably make already observed, copied, or inferred information unknown. Removing a primary record does not automatically remove backups, logs, downstream copies, model outputs, or decisions already made by recipients. “We can delete it later” is therefore not equivalent to declining to collect it.
Why it happens
Digital copying is nearly costless, while provenance and recall become harder across replicas, organizations, and time. Even byte-level removal of a source record can leave scores, classifications, and human knowledge derived from it. Under some conditions, aggregate models can reveal membership or attributes as well. Remediation can reduce future accessibility and propagation, but cannot recreate the counterfactual in which nobody learned the information.
Studying it
Data-lineage audits can seed a marked test record, trace it through operational stores, caches, warehouses, logs, backups, and processors, then retest after deletion. Relevant measures include replica coverage, propagation delay, unlocatable processors, and clearance of derived fields. Studies should distinguish logical deletion, cryptographic erasure, and physical destruction, and must not use a “deleted” interface state as evidence of backend completion.
Where it stops holding
Irreversibility does not make remediation pointless. Revoking access, rotating keys, notifying recipients, and correcting models can materially reduce future harm; verifiable cryptographic erasure in a controlled system may approach effective deletion. Recall is much weaker once data are public, copied by independent actors, or used in external decisions. Whether a model requires retraining depends on threat, obligation, and impact rather than a universal rule.
Applying it
- Apply a recall test before collection: if this datum leaked or its purpose changed tomorrow, would allowing it into the system still be defensible?
- Maintain lineage for each data class, including processors, derivatives, and deletion-propagation agreements.
- When collection stops, freeze new inference and sharing as well; do not describe a disabled toggle as erasure of history.
- Run deletion drills with uniquely marked records across live stores, restored backups, logs, and vendor attestations, reporting any residue, justification, and expected clearance time.