What you never paste into a chat

A chatbot feels like a private conversation, but everything you type ends up on a company’s servers. Sometimes in places you don’t expect, too.

After this lesson you have: a fixed list of what never goes in, a way to still ask for help without sensitive data, and five things you sort out today.

Six things that never go in

What	Examples	Why it’s extra sensitive
Identifying numbers	BSN (Dutch citizen service number), passport or driving licence number, IBAN with name	Name + BSN + address is a complete package for identity fraud.
Photos and scans of your ID	passport, driving licence, selfie with ID	Number and face in a single file.
Medical data	diagnoses, medication, test results, including other people’s	Special category personal data under the GDPR: the most heavily protected category.
Customer and staff data	customer files, CVs, case files, emails with names	You are responsible. Without agreements, entering it can be a notifiable data breach.
Passwords, API keys, tokens, recovery codes	a config or .env file you want “quickly checked”	They end up in your history, in shared links and with anyone who steals your session.
Confidential documents	contracts under NDA, your employer’s source code, strategy	An AI provider is a third party, and an NDA often specifically forbids sharing with third parties.

Six ways your chat travels further

1. Training. On most free accounts training is on by default. What’s already in a trained model can’t be taken out, not even if you turn training off later (Anthropic).
2. Humans read along. Google has some Gemini chats reviewed by humans and keeps those for up to three years.
3. Deleted isn’t gone right away. Claude removes deleted chats from its servers within 30 days; until then they’re still there. Gemini keeps chats for at least 72 hours, even if you turn off Keep Activity.
4. Shared links are public. In August 2025 Forbes found more than 370,000 shared Grok chats in Google, some with names, personal data and at least one password.
5. Clone apps and extensions leak. The big leaks of 2026 weren’t at the makers themselves, but at parties in between. At the app Chat & Ask AI, according to Malwarebytes (February 2026), some 300 million messages were exposed; two Chrome extensions with around 900,000 users combined forwarded chats to attackers.
6. Lawsuits. In The New York Times’ case against OpenAI, OpenAI had to keep deleted and temporary chats too, from May to the end of September 2025. What was kept then stays kept, except chats from the EEA, Switzerland and the UK. The lesson: “deleted” means “according to the policy, unless a judge says otherwise”.

How to still ask for help

• Replace, don’t just delete. “Jan de Vries, Kerkstraat 12” becomes “[CUSTOMER-A], [ADDRESS-A]”. You keep the key list to yourself, never in the chat.
• Remove rare details. Profession + village + illness is often enough to recognise someone, even without a name.
• Ask for the form, not the content. “Write a rejection letter for an applicant with five years of experience” instead of pasting the CV.
• Screenshots: crop, don’t blur. Watch out for tabs, notifications and names in a corner of your screen.
• Paste text, not files. An uploaded document carries its author, comments and sometimes revision history with it.
• Keys out before you paste. Replace them with <YOUR-API-KEY>.

Placeholders don't make it anonymous

With your key list next to it, “[CUSTOMER-A]” can still be traced back to a real person. Placeholders reduce the risk; for real customer files you still need a business account (see below).

Do you work with customer data?

GDPR for freelancers and SMEs

• It really happens. In December 2025 it turned out that employees of the municipality of Eindhoven had put CVs, youth care files and internal reports into public chatbots over 30 days. The municipality blocked those sites and asked OpenAI to delete the data. The Dutch Data Protection Authority (AP) received dozens of similar reports.
• Notifiable. According to the AP, entering personal data into a chatbot against the agreed rules is often a notifiable data breach. Earlier examples: a GP practice employee with patient data and a telecom employee with customer addresses (2024).
• How to arrange it. Customer data only in a business plan (Team, Business or Enterprise) with a data processing agreement and training off. With a free or personal account, don’t count on such an agreement.
• Write it down. One paragraph on “what may and may not go into the chatbot” is enough; that’s also what the AP advises.

Deepfakes and voice clones: the call-back rule

A few seconds of audio from social media is enough to clone someone’s voice (AVROTROS Radar). The Fraudehelpdesk (the Dutch fraud helpdesk) received more than 100,000 reports about all kinds of fraud in 2025, with almost €69 million in reported damage. Voice cloning works because a trusted voice switches off your critical thinking.

• The call-back rule: someone urgently asks for money, codes or data, even with a familiar voice or face? Hang up and call back on a number you already had. Never on the number from the call or message.
• Code word: agree on a word with your family that appears nowhere online.
• At work: payment changes and urgent payments always via a second channel and with two people. “The director is calling on video” is no proof.

Don't give a thumbs up on sensitive chats

Feedback is data too. At Claude, a thumbs up or down stores the entire conversation for up to 5 years, and Anthropic may use that feedback for training. Mistral: feedback with a comment may always be used for training. Unhappy with an answer? Better to start a new chat.

A shared link is a publication

Treat a share link like a blog post: clean it up before you share, delete it when you’re done (via the list of shared links in your settings). What has already been read, you can’t take back: at ChatGPT a viewer can add a shared chat to their own history, and that copy stays.

Security

Pasted something that didn’t belong there after all? Do this in the first 15 minutes:

• Password, key or recovery code? Change or revoke it now. Deleting the chat isn’t enough: it stays on the servers for days.
• Shared? Delete the shared link first, then the chat.
• Customer or patient data? Note what, when and in which assistant, and ask the provider to delete it. If it’s a notifiable data breach, report it to the AP within 72 hours.
• Training off, temporary chat and your account locked down: see privacy settings per assistant.

Checklist

0/5

• Turned off training in the assistant I use
• Went through my shared links and deleted what is no longer needed
• Removed browser extensions I don't know or don't use, especially AI sidebars
• Agreed on a code word with my family and explained the call-back rule
• For work: wrote down what may and may not go into a chatbot, and customer data only in a business account

Your progress is only stored in this browser.

Test yourself

1. 1You want help with an angry email from a customer. What is the best approach?

   Paste the whole email, so the AI understands it betterReplace name, address and details with placeholders like [CUSTOMER-A], then pasteDon't ask, AI is never safe for this kind of thing

2. 2You accidentally pasted an API key into a chat. What is the most important step?

   Delete the chatRevoke the key and create a new oneTurn off training

3. 3Your "son" calls with a familiar voice and urgently needs money. What do you do?

   Transfer it quickly, it really sounds like himCall back on the number he just called fromHang up and call back on the number you already had for him

Next lesson: privacy and account settings per assistant. Training off, temporary chats and your account locked down.