How to know what's safe to put into AI

Sensitive data. Public trust. Organisational rules. All of these factors come into play when determining what information public servants can provide to AI systems.

At Apolitical's recent live session, What's Safe to Put Into AI (and What Isn't), Antonia Mochan, a social scientist at the European Commission's Joint Research Centre, walked public servants through exactly that question.

While there's no universal answer to what's safe to share with AI, a set of habits can help you decide for yourself.

Here are six of the most practical takeaways from the session.

1. The absolute don’ts

Classified material and sensitive personal information don't need a framework; they're simply off limits. By classified, Mochan meant material formally marked as such under a national or organisational framework. Think official-sensitive, restricted or secret, not just information that happens to feel private.

The same goes for anything your organisation hasn't made public, and pre-legislative material sits in a greyer but still cautious zone. Beyond that is where you need to exercise your judgement.

2. Use frameworks to decide whether (and how) to use AI

As well as data sensitivity, it’s important to consider whether AI is the right tool for the task at hand, and how you can carefully use it once you start. Mochan shared two simple frameworks for thinking these through.

• For deciding whether to use AI at all, there's can't, can, should, shouldn't: just because a tool can do something (write a briefing, generate a graphic) doesn't mean it should. Think about the wider implications, like whether it sidelines a colleague's learning or workload.

• For being careful throughout the process, there's slow, fast, slow: think carefully before you prompt, weighing up what you're asking the AI to do and what you're sharing to do it. Then let the AI do its fast thinking. Then slow down again, and check what it gives you back critically rather than taking it at face value.

3. Run it through the café test

One of Mochan's simplest checks is what she calls the café test.

Picture yourself seated in a busy café, papers spread out on the table in front of you. Would you say this out loud in that setting, for the people at the next table to overhear? Or, if you left the papers there by accident, would it be fine?

If the answer is an absolute no, don't share the info with artificial intelligence systems either.

4. Watch what the AI remembers from earlier

Keep an eye on what information an AI has about your work/data.

A budget question on its own might be harmless. But if you've already asked the same tool something else work-related, it holds onto that context and can start inferring things you never told it, joining dots you didn't mean to connect.

Before you type, think about what else the AI already knows about you from contextual memory and whether that changes the picture of what you tell it.

5. Ask the AI to check itself

One of the more practical ideas from the session: build a prompt that gets AI to interrogate your task for sensitivity before you share anything real. Something like:

You are acting as a cautious prompt-engineering assistant, helping me check whether a task I want an AI to do involves sensitive information, before I share anything with you or any other AI tool. Ask me to briefly describe, in general terms only, what I want an AI to do and what kind of material is involved. Based on my description, assess the sensitivity level as: public, internal but low-risk, likely sensitive, or classified/restricted. Ask one clarifying question at a time, then give a plain-language verdict, a one-line reason, and 2-3 concrete suggestions.

Then describe your own budget, briefing, or dataset the same way, in general terms only, and see what it says. You'll get a verdict (one of those four categories), a short reason for it, and a couple of concrete suggestions, such as stripping out names or using an internal tool instead. That gives you a clear answer before you've risked sharing anything real.

6. Never let AI fully off the leash

Mochan was blunt about her own mental model: AI is useful for getting things moving, but not to be trusted without checking. It’s important to double check with every use. A data leak might not happen all at once, it might be small, incremental decisions to share a bit more information each time until suddenly you've offered up too much. Measure twice, cut once, every single time.

None of this replaces your organisation's own AI policy, and Mochan was upfront that she doesn't know yours. But whichever rules you're working within, the habit underneath them is the same: think slow, use the tool, then think slow again before you trust what came back.

Watch the highlights

Catch the highlights video here, or watch the full recording and find Antonia's slides and prompt template on the event page. Full speaker bios are also available there.

This article was written with the help of AI using an anonymised transcript and edited by a human before publishing.