Ethical Data Handling
What this domain tests
Ethical data handling: collecting and using data lawfully, protecting individuals, and communicating limitations. The exam tests privacy principles, de-identification, consent, bias, and responsible sharing.
Core rules
- Minimize: collect only the fields the task needs, and delete them when no longer needed.
- Identify personally identifiable information such as names, emails, precise location, and device IDs, and handle it under an appropriate legal basis such as consent or contract.
- De-identification: pseudonymization replaces identifiers with tokens and is reversible with a key; anonymization aims to be irreversible. Quasi-identifiers such as age, ZIP, and gender can enable re-identification when combined.
- Consent is scope-limited: data collected for support emails is not automatically usable for marketing.
- Check bias in sampling and labels; document provenance, transformations, and known limitations.
- Store and transmit data securely; restrict access on a need-to-know basis.
Worked example
A team shares a customer dataset for a public dashboard. They remove names and emails, replace customer_id with a salted hash, generalize age to 10-year bands, and aggregate to region level before export. They also check that no region has fewer than 20 customers, because a small cell could expose an individual. Reasoning: removing direct identifiers is not enough; generalization and aggregation reduce re-identification risk, and a minimum cell size prevents small-group disclosure.
Traps that cost marks
- Assuming that dropping names makes a dataset anonymous.
- Combining quasi-identifiers that uniquely single out a person.
- Reusing data beyond the consent or purpose for which it was collected.
- Ignoring class imbalance or sampling bias and reporting results as universal.
- Sharing raw data in notebooks or repositories without checking for secrets and PII.