When an AI system says 'I deleted the data', how should an…

Direct answer

First define the deletion scope and every copy path in a coverage list. Apply an appropriate method to each target. Then test absence of the record, denial of prohibited access, prevention of reuse, clearance of caches and memory, restoration from backup and, where relevant, effective destruction of the encryption key. Deleting stored data does not automatically remove what a model learned from it. A machine-unlearning or retraining claim is tested separately with fit-for-purpose evaluations such as memorisation, membership and behavioural tests. If evidence is unavailable, the result is 'not verified'.

In plain language

Moving a computer file to the recycle bin does not mean that it cannot be recovered from the disk. The bin, backup, cloud copy and any other system that learned from the file must be checked separately.

Why this matters

A false deletion statement gives a person false confidence while the data may continue to be used in retrieval, decisions, backups or model outputs. One screenshot cannot prove the status of every copy.

Do not confuse

Logical deletion may remove a record from the ordinary interface without ending physical accessibility.
Media sanitisation aims to make access to target data infeasible at a defined level of effort.
Cryptographic erasure can, under the right conditions, end access by rendering the encryption key unusable.
Clearing a cache or memory store does not automatically remove training-data influence from a model.
Machine unlearning seeks to reduce or remove the influence of specified training data; the result still requires testing.

What should you do?

List source, copies, caches, indexes, vector records, memory, event records, backups, exports, models and third parties in the scope.
Choose the appropriate deletion, destruction, anonymisation or restriction method for each object and medium.
Record the operation identifier, tool, operator, time, result, error and remaining exception.
Test read-back, search, access and reuse against the target.
Test scenarios in which a backup is restored, a cache is rebuilt or a downstream system retrieves the data again.
For a model-influence claim, run fit-for-purpose memorisation, membership, behavioural and retained-utility tests.
State what the evidence does not cover and do not extend an absolute-deletion claim to unsupported surfaces.

How do you audit it?

Does the deletion scope include every copy and derived object?
Is the selected method appropriate for the medium and sensitivity?
Is the deletion claim supported by evidence beyond the system provider's own statement?
Were read-back, access and reuse tests actually performed?
Does restoring a backup or rebuilding a cache bring the old data back?
Was storage deletion presented as machine unlearning?
Are remaining uncertainty and unverifiable third parties explicit?

Limit

Absolute irrecoverability may not be provable for every medium and third party. NIST media sanitisation is not a machine-unlearning standard. In a closed model, assurance that learned influence has been removed completely may remain limited by provider access and available test methods.