AINUA AI Academy
F4.3 · Foundations

Beginner · responsible practice · 19 min core lesson · 4 min quick note

Protect personal and confidential data

Share the minimum and use the right environment.

Ethics, safety and citizenshipInformation and dataHuman agency
Nia and Kio introducing Protect personal and confidential data
Cold open

A useful answer does not undo a disclosure. Once a client list, a medical note or a colleague's home address has left your control, no amount of quality in the output buys it back.

By the end, you can

redact a real task down to the minimum information a tool needs, and decide which environment a given piece of information may enter.

Quick note · 4 min

Ask what the tool actually needs to do the job. It is almost always less than the document in front of you.

Do now: Take one real task, redact it to the minimum, and check the output is still useful.

Minimise first, then choose the environment. Never the other way round.
Why it matters

This is the mistake with no recovery path. Bad drafts get rewritten and wrong figures get corrected, but a breach involving other people's information cannot be undone, and in many places it carries legal consequences as well as human ones.

01

Four questions before anything is pasted

Whose information is this? Yours, your employer's, a client's, or a person who has no idea this conversation is happening. The further from yourself, the less right you have to make the decision alone.

What does the tool actually need? A model helping you draft a difficult reply needs the situation and the tone, not the person's full name, account number and address. A model helping analyse a pattern needs the pattern, not the identifiers. Strip first, then ask.

Where does it go, and for how long? Consumer tools, business tiers and self-hosted systems differ enormously in retention, training use and access. Find the actual answer for the tool you are using before it matters, not afterwards.

What would happen if this appeared publicly? Not because it is likely, but because it is the question that reveals what you have been treating casually. If the answer is 'somebody gets hurt', the information does not go in.

Minimisation is the control that works even when every other control fails.
02

Redaction that actually works

Replace identifiers with placeholders that keep the structure: CLIENT A, SUPPLIER B, AMOUNT X. This preserves everything the tool needs to reason about the situation while removing everything it does not need to know.

Be careful with combinations. A name removed but a rare job title, a small town and a specific date left in can identify someone as precisely as the name did. Ask whether the remaining details would let a local reader work out who this is.

Watch the file itself. Documents, spreadsheets and photographs carry metadata — author names, edit history, locations, device identifiers — that you are not looking at when you upload. Screenshots capture whatever else was on the screen, including the notification that arrived at that moment.

03

Match the environment to the sensitivity

Think in three tiers. Public or non-sensitive material can go anywhere reasonable. Internal business material belongs in an environment your organisation has actually approved, with terms someone has read. Personal data about identifiable people, health information, financial records and anything about children belong in an approved environment or nowhere.

Where you work under a data protection framework — Kenya's Data Protection Act and the Office of the Data Protection Commissioner, or an equivalent elsewhere — the obligations apply to AI tools exactly as they apply to any other processing. Convenience is not a lawful basis.

Write your organisation's tiers down and make them short enough that people remember them. A one-page rule that gets followed beats a fifteen-page policy that gets skimmed once during onboarding.

Green, amber, red. If people cannot recall your data rules under time pressure, they will not use them under time pressure.