How to De-Identify Data or Build Synthetic Practice Material
Identify direct and indirect identifiers, preserve only the task structure you need, and create synthetic examples that do not represent real people.
When you cannot use real data, you may still be able to test an AI workflow. First decide whether you are testing classification, summarization, formatting, or process behavior. Then preserve only the structure needed for that test through removal, generalization, pseudonymous codes, or fully synthetic material. Replacing a name with “Person A” is usually not enough.
A customer record can remain identifiable after the name is removed. The company, job title, unusual product, exact date, and event description may still allow someone familiar with the situation to recognize the person.
NIST describes de-identification as removing the association between identifying data and a data subject to reduce disclosure risk while retaining useful analysis. It also warns that tools that merely mask personal information may not provide sufficient de-identification. NIST: SP 800-188
NIST SP 800-188 is technical and governance guidance for government datasets. The four choices below—removal, generalization, coding, and creating material from scratch—are my beginner method for low-risk practice. They are not an anonymization certification or legal advice.
Start by naming what the exercise must test
If you want to test whether AI can classify customer messages into three categories, you need different tones and problem types. You do not need real names, order numbers, or company secrets.
This material is used only to test whether AI can ______.
For that purpose, it must preserve ______.
It does not need to preserve ______.The clearer the purpose, the less information you need to move. The safest data is often the data you never collect or copy into the exercise.
Use four approaches for different risks
Remove fields that the test does not need
Delete direct identifiers such as names, phone numbers, email addresses, identity numbers, order numbers, and precise addresses when they are unrelated to the exercise. Do not merely cover text with a visual black box if the underlying content can still be recovered.
Generalize values when exactness is unnecessary
Change “August 27, 2026 at 2:03 p.m.” to “August 2026.” Replace an exact age with a range and a street address with a city or region. The less precise value still needs to support the test.
Use codes when records must remain distinguishable
Replace customers with C001 and C002 when you need to associate several records with the same fictional or protected subject. Coding is not the same as anonymity. If another table maps C001 back to a real person, that table remains sensitive and should not travel with the practice material.
Create synthetic scenarios from scratch
When the goal is to test a process, fully fictional material is often the best beginner option. Write six fictional customer messages covering returns, invoices, and delivery questions. Do not derive the names, orders, or incidents from real records.
Synthetic data is not automatically risk-free. NIST notes a trade-off between privacy and usefulness, and synthetic data generated too closely from source records can still reveal patterns or individual cases. For a beginner exercise, the simplest boundary is not to ask AI to anonymize sensitive source material. Design the fictional situation before the real data enters the system.
Look for clues that identify someone only in combination
Direct identifiers are not the only risk. Quasi-identifiers may be harmless alone but revealing together:
- the only overseas regional manager;
- a rare incident on an exact date;
- a precise title, tenure, and age in a small team;
- an internal project code paired with a customer industry;
- a distinctive sentence that can be searched online.
Ask: “Could someone familiar with this organization narrow the record to a few people?” If yes, generalize further, combine categories, or replace the material entirely.
Practice: create six synthetic customer messages
Do not use real cases. Create:
- two messages that need an ordinary reply;
- two that only need to be recorded;
- two that should be escalated to a manager;
- fictional names, products, and order formats;
- no scenario based on a real customer or recent incident.
Then save:
De-identified or synthetic practice pack
Purpose of the exercise:
Structure that must remain:
Direct identifiers removed:
Fields generalized:
Are codes and lookup tables separated?
How synthetic content avoids representing a real person:
Re-identification check:
Situations where this material is still unsuitable:
Practice material:There are two completion tests: the material can test the intended capability, and another person familiar with the background cannot reasonably infer a real person, customer, or secret.
This is still not a legal anonymization assessment. Medical, financial, employment, child-related, legal, and highly sensitive commercial data should not be uploaded merely because you completed this worksheet.
The next lesson places safe material and durable rules in a work structure where they can be found again instead of being scattered across conversations.