Field notes·AI·2026

What You Should Never Type Into an AI Chatbot

The famous AI risk is believing a wrong answer. The faster one is what you type in with your own two hands, on a night you need help badly enough to forget who's listening.

The design of these apps quietly lies to you. A chat window feels like a diary — private-looking, talks only to you, remembers yesterday. Every cue whispers this is between us. Here's the truer picture: a chat window is an email to a company. A brilliant, genuinely useful company — but a company, with servers, employees, security reviews, lawyers, and the occasional breach, like every company you've ever heard of. Chat logs have already been demanded and produced in real court cases, because a conversation with a machine is a business record, and business records can be summoned.

You wouldn't email your Social Security number to a helpful stranger at a corporation. Same rule here. Five categories should never make the trip.

The Never List

• The keys. Passwords. Full account numbers. Card numbers. Your Social Security number. These aren't information about you — they're the keys to you, and a chatbot never needs a key to help. It can explain a bank statement without your account number and decode a tax notice without your Social.
• Your name married to a sensitive fact. Here's the nuance most advice misses: the sensitive fact is usually fine to discuss — that's often exactly the help you need. It's the marriage of fact to identity that creates risk. "A 61-year-old warehouse worker with a denied injury claim" gets identical help to the version with a full name attached. One creates a permanent record about a person. The other is about no one.
• Other people's secrets. Your sister's diagnosis, your friend's marriage, your coworker's mistake. You have every right to seek advice about these situations. You don't have a right to file someone else's private life in a company's servers with their name on it. Change every name — the advice comes out identical.
• Your employer's confidential material. People have been fired for pasting work secrets into public chatbots — not because the machine did anything sinister, but because sending confidential data to an outside company is precisely what the confidentiality rules were written about.
• Anything that fails the headline test. Before hitting enter: if this exact message appeared in a news story with my name on it, would my stomach drop? Not paranoia — the same judgment you already apply to email, finally applied to the box that feels like a diary.

The workaround that costs nothing

Now the good news: none of this limits the help you can get, because the machine helps by recognizing your situation's pattern — and patterns don't have names. The move is redact and generalize. Names become roles: "my landlord," "a 58-year-old teacher," not the actual humans. Precise numbers become ranges when precision isn't the point: "about $40,000," "early sixties." Documents get their headers torn off before uploading — the name, address, and ID block go, the body text stays, and the body was always where the help lived.

Sixty seconds of repackaging, identical quality of answer, and the permanent record is about nobody. Notice what's not on the Never List: your problems, your questions, your fears at nine at night. Those are exactly what the box is for. You just don't have to sign them.

The machine runs on patterns, and patterns don't have names. Identity was never an ingredient — it was just a habit.

Two settings worth changing tonight while you're at it: find the control that decides whether your conversations train future models, and switch it off. Then give the account a password you use nowhere else, with two-factor on — because the likeliest way a stranger ever reads your chats isn't a corporate conspiracy; it's a password you also used on a pizza site that got breached years ago. And since who has write-access to your attention is the same question in another costume: audit those permissions too.