# What you should never paste into a chatbot > A chat box looks like a notebook and behaves like a form you're submitting to a company. Four kinds of thing should never go into it, and each app has one switch that takes two minutes to find. - Author: Abhay Bagda (https://abhaybagda.com/about/) - Published: 2026-10-19 - Topic: AI 101 - Series: Part three - Reading time: 6 min - Canonical URL: https://abhaybagda.com/writing/never-paste-into-chatbot/ --- [Part two](/writing/rag-open-book-exam) ended with an instruction, which was to give the tool the actual page when the answer matters. This piece is the warning that should have come attached to it. I've watched a whole document go into a chat box at close range, name and numbers and all, because the question was urgent and the box was right there. The chat box looks like a notebook. It behaves like a form you're submitting to a company, and what you type leaves your phone the moment you press send. It's easy to forget that when the thing answers back so politely. In [part one](/writing/inside-chatgpt) I said the ingredients for one of these models are text and electricity. Your chats are text, which is why the companies are so keen to hang on to them. ## Your message goes to more places than the reply comes from When you press send, at least three things can happen to what you typed, and the app only shows you one of them. **It gets stored.** Your conversation history lives on the company's computers, which is how it follows you from your phone to your laptop. Anything a company keeps can later be leaked, demanded by a court, or read by someone whose job it is to read it. **A person might read it.** These companies have staff and contractors who go through a sample of real conversations to check the quality of the answers. Google puts this in writing for Gemini, in a line more candid than most privacy pages manage: "Please don't enter confidential information that you wouldn't want a reviewer to see". The chats that do get picked are [kept for up to three years](https://support.google.com/gemini/answer/13594961). **It might become training material.** This is where part one comes back. A model is made by showing it text and nudging its numbers, trillions of times over, and unless you've said otherwise, your conversations can go into the pile for the next round. Once something has been nudged into those numbers there's no page to tear out, because none of it is written down anywhere you could point at. There's a fourth thing, and it has nothing to do with what the companies want. Sam Altman, who runs OpenAI, raised it himself on [Theo Von's podcast](https://techcrunch.com/2025/07/25/sam-altman-warns-theres-no-legal-confidentiality-when-using-chatgpt-as-a-therapist) last year. "And right now, if you talk to a therapist or a lawyer or a doctor about those problems, there's legal privilege for it," he said. A chatbot has no such protection, and he called that gap "very screwed up". If a court asks for your chats, the company can be made to hand them over. > Type into a chatbot the way you'd write on a postcard. Assume someone other than the person you addressed it to is going to read it. ## Four kinds of thing stay out of the box **Identity numbers** stay out, which means your NRIC, your passport page and your Singpass details. A model never needs these to help you, and they're the first thing a scammer needs in order to become you. **Anything that moves money** stays out too. That covers card numbers and bank logins, and above all the one-time codes your bank texts you. If a chatbot, or anything pretending to be one, ever asks you for a code, close it. **Other people's information** isn't yours to hand over, whether that's a screenshot of a group chat, your mother's test results with her name at the top, or a spreadsheet of customers. Nobody in those agreed to be in there, and you can't take it back on their behalf. **Your employer's material** is the one that gets people into real trouble. Samsung [restricted these tools on company devices in 2023](https://techcrunch.com/2023/05/02/samsung-bans-use-of-generative-ai-tools-like-chatgpt-after-april-internal-data-leak/) after internal data ended up in ChatGPT, and plenty of companies have a rule about it that nobody reads until somebody breaks it. ## You can still ask the question None of this means you can't use these tools for the things that matter to you. You almost never need the identifying part to get the answer. A model doesn't need your name to explain what a cholesterol number means, and it doesn't need your landlord's name to tell you whether a clause in a tenancy agreement is unusual. So strip the document before it goes in. Type the numbers from the test result instead of uploading a photo of the whole page. Replace names with "Person A" and "Person B". Delete the account numbers and the address block, and crop the letterhead off a photo before you attach it. You'll get the same answer, and the copy sitting on somebody's server won't have your name on it. ## Each app has a switch, and it only covers one of the three These are the names as they stand this month. Companies move them around, so if yours looks different, search the settings for the word "training" or "improve". **In ChatGPT**, open Settings, then Data Controls, and turn off "Improve the model for everyone". There's also Temporary Chat, which keeps a conversation out of your history and out of training. **In Gemini**, the setting is called Keep Activity, and it sits inside Gemini Apps Activity. Turning it off costs you your chat history, and Google still holds each conversation for 72 hours. **In Claude**, open Settings, then Privacy, and turn off "Help improve Claude". Anthropic's [own terms](https://www.anthropic.com/news/updates-to-our-consumer-terms) say that with it on, chats can be kept for five years, and with it off, thirty days. **Meta AI** doesn't have a switch. Since last December, what you say to it is also used to choose the ads and posts you see on Facebook and Instagram, in most countries outside the EU, the UK and South Korea, and [you can't opt out](https://www.macrumors.com/2025/10/01/meta-ai-ad-targeting/). Of the four, it's the one I'd tell my parents not to confide in. Go back to the three places your message goes. The switch deals with the third and leaves the other two alone, so your chats are still stored and can still be reviewed or demanded. It also applies only to conversations from now on. That's why I'd still strip a document before pasting it, even with every switch turned off. ## What to do tonight Open whichever of these you use and flip the switch, which takes about two minutes an app. Then, before the next thing you paste, run the postcard test. Would you be fine with this arriving at the company with your name on it, readable by a stranger? If the answer's no, take the names and numbers out and ask anyway. And if someone you know has started treating one of these as a diary, send them this. If you want to read how a company describes all of this in its own words, [Google's privacy page for Gemini](https://support.google.com/gemini/answer/13594961) is the most candid of the lot. Fair warning, it's long.