Zarlent

How personal information is removed before AI training

De-identification removes or replaces names, contact details, account numbers and other identifiers in company records so the people in them cannot be identified. It runs before the records reach an AI lab; it lowers privacy risk but does not remove it, so good licenses add an independent check and a ban on re-identification.

No buyer pays to be mentioned

What could your company's records earn? Get a range in 30 seconds. Free, no obligation.

What does de-identification remove?

US health-privacy rules give the most detailed public list. The HIPAA "Safe Harbor" method requires removing 18 kinds of identifiers, including names, street addresses and ZIP codes (with narrow exceptions), all date elements except the year, phone and fax numbers, email addresses, Social Security numbers, account numbers, IP addresses, web addresses, device serial numbers, biometric identifiers and full-face photos (HHS de-identification guidance).

That list was written for patient records, but buyers of company records remove similar identifiers from email, chat and documents: names become placeholders such as "Employee 14", and addresses, phone numbers and account numbers are deleted or masked.

What are the two accepted methods in US health law?

MethodHow it worksSource
Safe HarborRemove the 18 listed identifiers, and have no actual knowledge that the rest could identify someone45 CFR 164.514(b)(2)
Expert DeterminationA qualified expert finds "the risk is very small" that the data could identify someone, and documents the method45 CFR 164.514(b)(1)

Source: HHS guidance on de-identification under the HIPAA Privacy Rule. Once health data meets either method, HHS says it is no longer protected health information.

How does EU law treat de-identified data?

Under the GDPR, truly anonymous information falls outside the regulation, but "pseudonymised" data that could be linked back to a person "should be considered to be information on an identifiable natural person" (GDPR Recital 26). Swapping names for codes is not enough on its own if a key or other data could reverse it. EU and UK companies should expect stricter scoping.

What are the limits of de-identification?

Workplace records are hard to anonymize fully. A message that says "the night-shift supervisor at our Tulsa plant" can point to one person with no name in it. Privacy experts warn that each copy or breach of a dataset creates new chances to link data points back to people (The Record, 18 Nov 2026).

That is why the strongest public terms go further than removing names. In the court-approved Spirit Airlines sale, the buyer agreed to:

  • Exclude health and medical data, drug and alcohol test results, safety and disciplinary records, and identity documents
  • A "rigorous de-identification process"
  • Review of the de-identified data by an independent assessor chosen with the union
  • Not attempt to re-identify the data

Source: The Record, 18 Nov 2026.

What should your license say?

  1. Who de-identifies, and before which point (before any AI lab sees the records).
  2. Which identifiers are removed, as a list.
  3. Which categories are excluded entirely.
  4. An independent check, or your right to review a sample.
  5. A ban on re-identification that passes to anyone who later receives the data.
  6. What happens if personal details are found later.

See the 10 contract clauses to check.

Questions

Who sees the raw records before de-identification?

Usually the buyer's own processing team or a third-party processor, under contract. Ask the buyer to name who touches raw records and where.

Is de-identified data still personal data?

In the US it depends on the law; HIPAA-compliant de-identified health data is no longer protected health information. Under the GDPR, pseudonymised data that can be linked back is still personal data.

Can AI models leak names from training data?

Models can repeat rare text they saw in training. Removing identifiers before training is the main protection, which is why it must happen before delivery.

Sources