Skip to main content
Privacy and data

Erasing a person

When an administrator erases a person, AI Governance:

Written By Christopher Scaminaci

Last updated 4 days ago

At a glance

  • What it does: erases the identity and the stored text of one person, and keeps the evidence that does not identify them.
  • Plan: every plan. Erasure and data exports are never paywalled.
  • Who: an Administrator.
  • Where: AI Governance → People → Staff, in the Erase this person dialog. People → Data requests lists the requests staff make.

What erasing a person does

When an administrator erases a person, AI Governance:

  • removes the identities that tie the person to their usage: their email addresses, their sign-in identities and their accounts in the tools they use;
  • destroys the key that protects the text stored for that person, so that text can never be opened again by anyone, StackJack included;
  • removes the masked value and the fingerprint from every finding about the person, and keeps the rest of each finding (which rule matched, where, when and how serious) as evidence;
  • removes the fingerprint of the person's email addresses from other people's findings too, so a colleague's prompt that quoted the address no longer points back at the person.

The person's pseudonym stays on their usage records, so your totals and history stay correct. The erasure is recorded in your audit trail with the reason you gave; the record holds no name, address or other identity.

What the erasure does not reach

  • Mentions that appear later. The erasure cleans the mentions that existed when it ran. A prompt written afterwards that quotes the person's address is a new finding, and it is not cleaned by that erasure.
  • Findings under a legal hold. A finding about someone else that another legal hold protects is left exactly as it is, and the erasure does not go back to it when that hold ends.
  • Mentions StackJack cannot recognize. A mention is found by the fingerprint of the person's email address. It cannot be found when the matching rule stores no fingerprint (card numbers, national ID numbers, phone numbers and dates never do, and neither does a rule you set not to), when the mention is a name, or when it is something other than an email address, such as an account name in a tool.
  • Mentions fingerprinted under a key that no longer exists. If a key StackJack once used has since been deleted, the mentions fingerprinted under it cannot be compared. The erasure still goes ahead, and its audit entry says how many keys could not be read.

If the erasure is refused or fails

  • A legal hold on the person, on their organization, on your account or on a source that holds their text or their findings refuses the erasure. The refusal is recorded and nothing changes. Release the hold, then erase.
  • If StackJack cannot reach the key store while it erases, the erasure stops with an error before it changes anything. Run it again after about a minute.
  • Erasing a person who is already erased changes nothing, except that it finishes anything an earlier erasure could not (for example a finding written for that person just after the erasure).

The Erase this person dialog

The hint under the reason. Do not write the person's name here: the reason is kept in the audit trail, which cannot be edited.

What the dialog says afterwards.

  • Erased. Their text is destroyed; facts remain without their identity.
  • This person was already erased.
  • This person is under a legal hold and cannot be erased now.
  • Write the reason for the erasure.

A person's data export and the erasure decision

An Admin prepares an export of everything AI Governance holds about one person, under every name and address they use: their activity facts, the text of their own prompts and answers, findings about them (without masked values), their policy acceptances, training, tool requests, endpoint coverage, their own data requests and messages, reveals about them and who looked at their data. Text that belongs to another person (a name or contact detail inside someone's prompt) is redacted; a part that cannot be checked safely is withheld and counted, never shown; and the export says when other people's names could not be looked for. Text of a person the organization excludes from collection is listed and not read. An export can be downloaded until the date the page shows, and every download is audited.

The export's README says from which day its list of who looked at the person's data starts. If your own storage is your main store and older audit trail entries are only there, the README says that those entries are not in the export and that an administrator can read them one day at a time on the Audit trail page.

The erasure decision is the controller's, and a person decides it, never a rule. The Admin either erases now, or refuses with one of the grounds below and a written reason. The legal-hold ground is never offered as a choice: where a hold stops the erasure, the system records it itself.

We could not erase your data because a legal hold applies to it. It is kept until the hold is released.

We are keeping some of your data because we need it to establish, exercise or defend a legal claim.

We are keeping some of your data because the law requires us to keep it.

We are keeping some of your data because we need it for a task carried out in the public interest.

We are keeping some of your data for archiving, research or statistical purposes.

We are keeping some of your data because it is needed to protect freedom of expression and information.

We have decided not to erase your data at this time. Contact the organization's data protection contact for the reason.