You will be able to explain where a prompt goes, what gets stored, and which settings control it.
Plenty of adults now hand everyday jobs to an AI assistant. They type questions into it as they would into a search bar, keep notes in it, paste in emails for a rewrite, or drop in a spreadsheet and ask for it to be tidied. Because the screen shows only you and a text box, the exchange feels private. It is only partly private, and the gap matters more at work than at home.
Start with the path a message takes. You press send, and your text leaves your device for the provider, the company that runs the assistant. The provider's servers run the model, which writes a reply, and the reply travels back to your screen. Your message and the reply are usually both stored, at least for a while, as chat history. That stored record lets you scroll back through earlier conversations. It also leaves a copy of whatever you typed on a computer you do not own.
Three things decide what becomes of that copy: which provider you use, what kind of account you are on, and how your settings are configured. Three questions cover most of the ground. Is the conversation kept, and if so for how long? Can staff at the company read it, for instance when investigating abuse or checking quality? Can it be used for training, meaning as material to build the next version of the model?
Providers such as ChatGPT, Claude, Gemini and Copilot each answer these questions differently, and their answers shift over time. Rely on two sources only: the provider's current privacy policy and the settings page inside your own account. Some broad patterns do hold. On free and personal accounts, providers generally give themselves more room to use your conversations, and there is often a setting to switch that off. Business and enterprise plans usually include a contract in which the provider agrees not to train on your data, along with controls your employer manages. The same assistant can therefore be acceptable for a work task on the company account and a problem on your personal one.
Most of the worry centres on training. A trained model does not hold your chats as stored files. Even so, researchers have found that models sometimes reproduce fragments of the text they learned from. Training aside, any data a provider stores could leak in a breach, be looked at by a human reviewer, or be demanded in a legal case. You do not have to judge which of these is most likely. Once text has left your device, your control over it is no longer complete.
Optional features send your data further. With memory switched on, the assistant holds on to facts about you from one conversation to the next so it can tailor what it says. Connectors give it access to your email, calendar or files. Plug-ins and custom assistants made by other companies may forward your text to a further server. All of these have real uses, and every one adds to the places where your data ends up.
You can keep using AI. The habit to build is a quick sort before you type: decide which of three groups the information falls into, then act accordingly. Group 1 is safe to paste almost anywhere and covers public information, general questions, and your own writing when it contains nothing sensitive. Group 2 belongs only on an account your employer has approved, and internal documents and plans sit here. Group 3 stays out of every assistant: passwords, one-time codes, NRIC numbers, bank details, health or financial information about other people, and anything a client shared with you in confidence.
At work, Group 3 runs into the law. Under Singapore's Personal Data Protection Act (PDPA), a customer's name, phone number and NRIC number all count as personal data. If you paste personal data into a tool your company has not approved, that can count as a disclosure your company must answer for, however good your intentions were. Module 2 sets out what the PDPA requires of you.
Your task: open the settings of the assistant you use most and find three things. Are your chats used for training? How long is history kept? Is memory switched on? Write the answers down, then change anything you aren't comfortable with.
Open the settings of the assistant you use most. Write down whether your chats train the model, how long history is kept, and whether memory is on.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).