Privacy and AI: what not to share with a digital assistant

The chat box is the most intimate interface we have ever built for software.

No forms. No fields. No “Are you sure?” Just a blinking cursor that seems to say: tell me everything, I’ll help.

And people do. Contracts pasted in full. Client lists. Salary spreadsheets. Medical test results. Screenshots with someone else’s phone number in the corner. Passwords inside a config file, “just to fix the error”.

Not because they are careless. Because the interface feels like a conversation, and conversations feel private.

A conversation with an AI is a data transfer

When you type into an assistant, your text leaves your device and is processed on someone else’s infrastructure. What happens next depends on the provider, the product, your plan and your settings:

• how long the conversation is kept;
• whether it can be used to improve future models;
• who inside the provider can access it, and when;
• whether your company has a contract that changes any of the above.

I won’t tell you what each tool does today, because these policies change often. You should check the privacy settings and data terms of the specific assistant you use — and check them again every few months.

The principle is simpler than the policies: don’t share what you wouldn’t be comfortable seeing outside your control.

Keep these out of the chat

1. Credentials and secrets. Passwords, API keys, access tokens, recovery codes, private keys. Never, not even inside a log or a config file. Remove them before pasting.
2. Identity and financial numbers. Tax codes, ID and passport numbers, card numbers, IBANs.
3. Other people’s personal data. Names with contact details, health information, HR evaluations, customer records. It isn’t yours to hand over — and in Europe, GDPR has something to say about it.
4. Confidential business material. Unreleased financials, tender documents under NDA, pricing strategies, network diagrams, security configurations, source code your employer hasn’t cleared for external tools.
5. Your own sensitive details, by default. Health, legal situations, family problems. You may decide to share some of them. Make it a decision, not a reflex.

Use AI anyway: three habits that work

Privacy doesn’t mean giving up the tool. It means changing what you put into it.

Anonymise before you paste. Replace names with roles (“Client A”, “Supplier B”), real numbers with realistic placeholders, addresses with “[city]”. The assistant usually needs the structure of the problem, not the identity of the people in it.

Describe, don’t upload. Instead of pasting a 40-page contract, ask: “What should I check in a termination clause for a multi-year maintenance contract?” You get most of the value with none of the exposure.

Use the right channel for the right data. Many organisations now provide approved AI tools with contractual data protections. If yours does, company data goes there — not into a personal account on your phone. If it doesn’t, ask why. That question is worth more than any workaround.

A quick test before you press Enter

Read what you are about to send and ask:

• Would I email this to an external supplier I’ve never met?
• Does it contain anything that identifies a real person?
• Would my company’s security team approve this paste?

If any answer makes you hesitate, edit before sending.

The fear is reasonable. The paralysis isn’t.

In regulated sectors we learned long ago that security isn’t a feeling, it’s a set of controls: what goes where, who can see it, for how long. AI doesn’t change that logic. It just makes the most sensitive door look like a friendly chat window.

Treat it like a door.

Your first step

Open the AI assistant you use most. Find its privacy or data settings and read them for five minutes. Then delete the one past conversation you wish you hadn’t had — and keep in mind that deleting it from your history may not mean it’s gone everywhere; the provider’s terms tell you that too.

Back to Surprise Me — choose another fear to name.