The four things a tool can do with your text

When you paste text into an AI tool, up to four things can happen to it. It is sent to a server somewhere to be processed. It may be stored, for a period, in your chat history and on that server. It may be reviewed by a human, either for quality or safety. And it may be used to train future versions of the model, which is the one people worry about most, and the one that is most often switched off.

Which of the four apply depends entirely on the tool, the account, and the settings. A free consumer chatbot with default settings may do all four. A business licence from the same company typically does the first, may do the second briefly, contractually rules out the fourth, limits human access to rare cases such as abuse investigations, and lets your employer's admins decide how long chats are kept. Your organisation's own deployment may keep everything inside its own systems.

So the question is never "is AI safe." It is "which of the four things does this tool, on this account, do with this text." Once you can answer that, every other decision gets easier.

In practice. An administrator pasted a list of forty names and email addresses into a free chatbot to have it formatted into a table. The task was harmless. The list was personal data. On that account, chat history was on, training was on, and she had no idea where the server was.

Nothing bad happened, as far as anyone knows. But she could not have told her manager where those forty addresses had gone, and that, not the task, was the problem.

At your desk. Open the AI tool you use most at work. Find its data settings: look for history, training, privacy or data controls. Write down, for your account, which of the four things it does. If you cannot find out, that is your answer for now: assume all four.

Write it down.

• The tool I use most:
• Which of the four it does on my account:
• The setting I changed, if any: