Privacy and safety: what you can and can't put into AI

In the last lesson — the one about your first prompt — I dropped a one-line warning: don't paste confidential data into free AI chats. It's an important line, so let's take it apart. This is the most underrated beginner mistake, and even large corporations make it.

Where's the risk, exactly?

Remember the lesson on how a network learns: a model grows on the data it's shown. At many services the default rule on free plans is that your conversations may be used to train future versions of the model. You paste text into the chat box and it goes off to the company's servers, where it may be looked at during quality checks and folded into the training data. Confidential data is anything you wouldn't want outsiders to see: passwords, customer lists, internal documents, commercial terms. Send that to a free chat and you've effectively carried it out of the office.

Samsung, spring 2023: what this looks like in real life

The story has become a textbook case. In spring 2023, engineers at one Samsung division were allowed to use ChatGPT for work. According to media reports, several incidents happened within weeks: employees pasted internal source code and the contents of a closed internal meeting into the chat — so the AI could help find a bug and write up the notes. The data went to someone else's servers, and there's no getting it back.

The company's reaction was firm: Samsung restricted the use of generative AI on work devices. That same year, according to the business press, major banks and corporations brought in similar restrictions — JPMorgan and Apple were among those named — worried about exactly the same thing.

The moral is the same as in the hallucinations lesson: the person is answerable, not the tool. There, it was whoever didn't check the result; here, it's whoever didn't think about where their text was going.

The red zone: what must never go in

• 🔑 Passwords, keys, access codes — never, anywhere.
• 👥 Personal data. Personal data is information that can identify a specific person: other people's names, phone numbers, addresses, passports, medical records. Protecting it is a matter of law (GDPR in Europe, and most countries have their own personal-data laws), and whoever caused a leak answers for it.
• 📄 Trade secrets. A trade secret is information that gives a business an advantage precisely because competitors don't have it: private contracts, pricing, finances, strategy, source code.
• 🏥 Sensitive things about people — health, salaries, private correspondence.

The green zone: what's safe to put in

• Text that's already public or is being prepared for publication anyway.
• Your own drafts, with nobody else's secrets in them.
• Practice problems and abstract questions.
• Data with every identifying detail stripped out — which is what's next.

The technique that solves half the problem: de-identification

De-identification (de-personalisation) means swapping real details for placeholders before anything goes to the AI. Need to work through a real contract? Replace the parties with "Company A" and "Company B", take out the amounts and the surnames, keep the substance. To help with the structure of a letter or to find a hole in an argument, the model doesn't need the real names and numbers. This is the most practical skill in the lesson: the overwhelming majority of work tasks can go to AI de-identified — same benefit, no risk.

An important caveat. What you're doing is pseudonymisation (swapping names for placeholders), not full anonymisation. It lowers the risk but doesn't remove your responsibility: fragments of data can sometimes still be pieced back together to point at a specific person (that's called re-identification), and personal-data law (GDPR in Europe) still applies. So de-identification is the required minimum, but for genuinely sensitive material add a business tier on top and check your company's own policy.

Three rules worth knowing

• Check your privacy settings. ChatGPT, Gemini and the rest have a switch in settings that turns off the use of your chats for training (often bundled with saving your history). Find it — and turn it off for work tasks.
• For work data, use business tiers. On corporate and team plans (Business/Enterprise), companies generally promise not to use your data for training. If you handle anything sensitive, that isn't a luxury, it's the basics.
• If your company has a policy, it wins. Plenty of employers have already written down what may and may not go into AI. Not sure? Ask before you paste in a work document.

A simple everyday rule

Before you hit send, ask yourself one question: "Would I be comfortable if a stranger saw this text?" If yes, paste it. If you feel a flicker of doubt, de-identify it first or switch to a tool on a business tier. That three-second filter protects you better than any technical measure — because the decision happens before you send, not after a leak.

Do this now

Take any real work text — an email, an application, a chunk of a report — and practise de-identification, right there on paper or in your notes. Go through it and find everything that counts as personal or commercial: names, phone numbers, addresses, company names, contract amounts. Replace each with a placeholder — "Client A", "Company B", "amount N". Compare the two versions: the de-identified one explains the substance just as clearly, but gives nothing away. Do this once deliberately and from then on your hand will clean the text automatically before you send it anywhere.