Do not paste anything into a public AI chatbot that you would not be comfortable seeing stored on someone else’s server: customer personal data, ID and bank numbers, passwords and keys, unreleased financial figures, contracts under negotiation, or a colleague’s private information. What you type may be logged, reviewed for safety, kept for a period, and in some products used to improve future models unless you switch that off. The safe habit is to remove or replace sensitive details before you paste, and to use tools your organisation has approved for confidential work.
This is not a reason to avoid AI. Most useful office tasks work just as well with names swapped for “Customer A” and figures rounded. It is a reason to understand where your text goes after you press enter, which settings matter, and why a free personal account and a company-managed account can treat the same message very differently.
Where does your text go after you press enter?
Your message travels to the provider’s servers, where the model runs. To answer, the system needs to see the full text, so the provider can technically read it. What happens next depends on the product and your settings, and it can change when terms are updated, so the honest answer is: check the current policy of the tool you use.
In many consumer products, conversations are stored in your history so you can return to them. Some providers keep messages for a period for abuse monitoring even after you delete them from your view. Some use conversations from personal accounts to train future models by default, with an option to opt out. Business and enterprise plans often promise not to train on your data and offer stronger controls, but you should read the actual terms rather than assume.
There are also people in the loop. Safety teams may review a sample of conversations flagged by automated systems. That is reasonable for preventing abuse, but it means “only the machine saw it” is not a safe assumption.
Risks on your own side, and AI inside everyday apps
Then there is your own side. Chats synced across devices can be read by anyone who picks up your unlocked phone. Shared links to conversations can be forwarded. Browser extensions that promise to “enhance” AI chats may read everything on the page. Leaks often happen at these ordinary points, not through a dramatic hack.
Finally, remember that AI features now sit inside email, office suites and messaging apps. When a tool offers to summarise your inbox, it is reading your inbox. Turning such features on for a company account is a decision for whoever manages data, not just for the individual user.
What should never go into a public chatbot?
Personal identifiers of customers or staff. Full names combined with ID numbers, phone numbers, home addresses, dates of birth or health details. Vietnam has rules on personal data protection; ask your legal adviser how they apply to your use of AI.
Passwords, API keys and access codes. People paste error logs and configuration files that contain secrets without noticing. Treat any key that has been pasted as exposed and change it.
Bank and card details. Account numbers, card numbers, transaction screenshots with balances. There is no task that requires them.
Unreleased financial information. Draft results, pricing plans, bids, salary tables. Leaks here can harm the company and, for listed companies, may raise legal issues.
Contracts under negotiation. Terms, counterparties and red lines are exactly what a competitor would want. Remove names and figures, or use an approved tool.
Source code and internal system details. Proprietary code and network diagrams can help attackers. Many companies restrict this specifically; follow your policy.
Other people’s private messages. Pasting a colleague’s or customer’s private chat to “ask AI how to reply” shares their words without consent. Summarise the situation in your own words instead.
Anything covered by a confidentiality agreement. If you signed an NDA with a partner, uploading their documents to an outside service may breach it, whatever the AI provider’s policy says.
How do you anonymise a document before using AI?
A five-minute routine keeps most of the value and removes most of the risk.
- Decide what the AI actually needs. To improve the tone of a complaint reply, it needs the complaint and your draft, not the customer’s phone number or order history.
- Replace names with roles. Use “Customer A”, “Supplier B”, “our director”. Keep a private key on your side if you need to put names back later.
- Round or scale numbers. If the task is about structure or wording, replace real figures with rounded or fictional ones of the same size.
- Strip hidden data. Files can carry author names, comments, tracked changes and metadata. Copy only the needed text into a clean document.
- Remove screenshots with incidental details. A screenshot of one email often shows other subjects, names and notifications. Crop tightly or retype.
- Re-insert real details offline. Once you have the improved text, put the real names and numbers back yourself in your own document.
If the task genuinely needs the confidential details, that is a sign to use a company-approved tool with a proper data agreement, or not to use AI for that task.
Settings and habits worth five minutes today
Check the training setting. Look in the privacy or data controls section of your AI tool for an option about using your conversations to improve models, and decide deliberately.
Review chat history. Delete old conversations that contain anything you would rather not keep, knowing that the provider may still retain copies for a time.
Use a separate work account. Keep personal and work use apart, ideally with an account managed by your company where one exists.
Turn on two-factor authentication. Your AI history can contain a lot about you. Protect the account like you would protect email.
Be careful with shared links. A shared conversation link may be viewable by anyone who has it. Share a copied excerpt instead when possible.
Audit browser extensions. Remove extensions you do not need, especially ones that read page content on AI sites.
Ask IT before connecting company data. Plugins and connectors that link AI to your email, drive or CRM should be approved by whoever manages those systems.
If you pasted something by mistake, act early. Delete the conversation, change any exposed password or key, and tell your manager or IT so they can assess the risk.
Frequently asked questions
Does deleting a chat remove it from the provider’s servers?
It removes it from your view, but providers may keep copies for a period for safety or legal reasons. Check the retention section of the current privacy policy for the exact rules.
Are paid AI plans automatically private?
Not automatically. Business and enterprise plans often offer stronger commitments, but personal paid plans may follow the same data rules as free ones. Read the terms of the exact plan.
Is it safe to upload a contract for AI to review?
Only if the tool is approved for confidential documents and the contract’s confidentiality terms allow it. Otherwise remove names and figures, and have a lawyer review anything you will sign.
Can AI tools leak my data to other users?
Direct leaks between users are rare but bugs have happened in the industry, and data used for training could in principle influence later outputs. The safest data is the data you never pasted.
Does my company need an AI usage policy?
If staff already use AI, yes, even a one-page version: approved tools, forbidden data types, who to ask, what to do after a mistake. It works better than a silent ban that people ignore.
What if I pasted a customer’s personal data already?
Delete the conversation, tell your manager or data protection contact, and follow your company’s incident process. Depending on the data, there may be notification duties, so take legal advice.
