Almost anything can go into an AI. What cannot fits in three questions.
In an organised company, whoever has the doubt asks someone. You have no such someone. The file is open, the question is can I paste this?, and the answer has to arrive in ten seconds.
This text is that answer. It fits in three questions, and their purpose is not to make you use AI less — quite the opposite.
What changes when you hit enter
Start with the plain description, no drama: what you paste into an AI leaves your computer and enters another company's. Sometimes it is stored. Sometimes a person reads it to improve the service. Sometimes it becomes training for the model, depending on the plan you use and the setting you never opened.
That is not a reason to use it less. It is a reason to know where things go — which is exactly what separates someone who uses AI from someone who obeys it.
And notice the asymmetry, because it holds up everything else: being too careful with a text costs you thirty seconds of rewriting. Being careless once costs you a piece of data that does not come back.
The three questions
1. Is this about another person?
This is the one almost nobody asks, and the most important — because a good share of what we paste is not about us. It is your father's test results, which you want to understand before the appointment. It is the screenshot of a conversation with a friend, so you know how to reply. It is someone's CV, a client's contract, a photo of your child.
You chose to use AI. Those people did not. It is the only item on the list where the decision was never yours to make.
2. Does this identify someone?
Full name, national ID, address, phone number, card number, licence plate, account, a test result with the name on top. The catch is that identifying someone rarely needs a document: the combination is enough. "My neighbour in flat 43, the one who works at the city council" identifies a person just as well as an ID card.
3. If this showed up on a screen that is not mine, what changes?
Not "will it leak?", which nobody can answer. The question is about the size of the damage, and it is an honest one: for most of what you paste, the answer is nothing. A generic work text, a coding question, a draft email, tonight's recipe — nothing changes.
That is precisely why the yardstick allows more than it forbids. Once you know what is serious, you stop treating everything as serious.
How to use it anyway
The mistake most people make is not pasting what they should not — it is giving up on what they needed, because it felt risky. Three habits solve almost every case:
Drop the name. "A 62-year-old with these test results" works just as well as the test with the name on top. The AI does not need to know whose it is to explain what it means.
Change the number. To understand a contract, an approximate value will do. To check a calculation, a round number will do. You want the reasoning, not the record.
Describe instead of attaching. Much of what we dump whole into an AI could be three lines of summary — and the answer improves, because the essential stands out instead of drowning.
In all three cases you use the AI for exactly what it is good at, and hand over none of what it never needed.
It is worth opening that setting once
Most tools let you choose whether your conversations can be used to train the model. It is usually on by default, and takes a minute to turn off — once, for life, per tool.
It is not the most important part of this text, but it has the best return: a minute now is worth more than all the vigilance you will try to keep later.
And when it is work, the yardstick is not only yours
Here the conversation changes owner. If what you want to paste belongs to the company you work for — a contract, a spreadsheet, client data, the project that has not been announced — the decision stopped being individual, even if nobody told you anything.
Most companies have not written that yardstick yet. No rule is not permission, it is merely the absence of a rule — and data that leaves through a personal account leaves all the same, whether the company wrote anything or not.
If you are the one who answers for the company in this story, the problem stops being pocket-sized and becomes one of method: it is what can go in, before deciding who uses it.
The test
Open the last long conversation you had with an AI and read the first message.
Is anyone's name in there? Does that person know?
If the answer is no, you do not need to feel bad — you need ten extra seconds next time. That is all the yardstick is.
This is not legal advice. What counts as personal data comes from the law of your country, and what a specific tool does with what you type is in its terms — which change.