When the terms change

Agreements are not fixed. In January 2024, staff in the Federal Trade Commission’s Office of Technology wrote that companies offering AI models as a service must honor the privacy commitments they make to users.¹ A month later, they warned that quietly changing terms to use people’s data in new ways, such as for AI training, could be unfair or deceptive.² These were staff statements, not court rulings.

Companies do change their terms in the open, too. In August 2025, Anthropic announced that users of its personal plans would choose whether their chats help train its models. Chats would be kept for five years for those who agreed, and for 30 days for those who did not.³ So read the terms once, and again whenever the company says they have changed.

Working one through

Here is the reasoning in full, on an example written for this book. An HR manager has rough notes from a meeting about one employee’s performance and wants an AI tool to turn them into a tidy summary. Before pasting, she asks five questions.

• Is this my personal account or my organization’s?
• Who signed the agreement, and does it cover this kind of information?
• How sensitive is this, and what would a leak cost someone?
• Is this particular feature covered, or open but outside the agreement?
• Will a tool or connector send this to another company?

She starts with sensitivity. The notes name a person and judge their work, so a leak could do that person serious harm. She rates them at least moderate, and high if they touch on health or pay. That rules out her personal account, which is an agreement between her and the AI company, not between her employer and anyone. Her work account is the right room if the employer’s policy allows HR records in it, so she checks the policy rather than guessing. She notices that the summary doesn’t need the employee’s name, so she leaves it out. Last, she sees that the app is set to save results to a shared drive through a connector. Once the summary lands there, the drive company’s terms apply, so she saves it by hand to the folder HR already uses.

Your turn

Try it. Find out which room you are in. Open the AI tool you use most and answer three questions from what it shows you. Whose name or email is on the account, yours or your employer’s? Is there a setting about using your chats to improve or train models, and how is it set? Are any connections to other services switched on? If you would rather work on paper, list three things you might paste into an AI tool this week. Rate each low, moderate or high by the harm a leak would do, and write beside it which of the four arrangements it would need. Then test one claim from this chapter: does your tool’s own privacy page say who the agreement is between?

Think about it. What is the most sensitive thing you might be tempted to paste at work? Who in your organization would know which room it belongs in, and have you asked them?

So where do your words go? Through an app, to a model, and sometimes on to other companies, each hop under an agreement that someone signed. Which model it is matters less than whose account it is, whose agreement covers it, which features are on and whose machine runs it. Knowing where your words go tells you what to leave out. The next chapter turns to what to put in: how to ask so the model has what it needs.