What happens to what you type into an AI chatbot? It is sent to the provider’s servers, processed by the model and usually stored in your history; depending on the service and your settings, it may also be reviewed by staff or used to train future models. Consumer apps and business plans often handle this very differently, so the same prompt can carry very different risks.
Below is a plain explanation of the data path, a list of things that should never go into a chatbot, and the settings and habits that reduce risk for individuals and small companies.
Where does your chat go after you press send?
When you press send, your text, plus any files or images you attached, travels to the provider’s data centre. The model reads it together with the earlier messages in the conversation, produces a reply and sends it back. Unless you are using a model that runs on your own device, none of this happens on your phone; it happens on someone else’s computers.
Most services then keep the conversation so you can come back to it. Many consumer services have also, at least by default, reserved the right to use conversations to improve their models, while offering a setting to opt out. Business and enterprise plans typically promise not to train on customer data by default. The exact rules vary by company and change over time, so the only reliable source is the current privacy policy for the plan you actually use.
Reviewers, retention and third parties: the less obvious paths
Three further paths are less obvious. Providers may have staff or contractors review samples of conversations for safety and quality. Memory features store facts about you across chats. And shared links to conversations can be opened by anyone who has the link; in some cases such links have even ended up in search engine results.
Deleting a chat removes it from your view, but providers often keep data for some time for security or legal reasons before final deletion. If something sensitive went in, deleting reduces exposure but does not guarantee it never existed.
Finally, the chatbot company is not the only party. Browser extensions that promise to “enhance” AI chats, third-party apps built on top of a model, and even the keyboard app on your phone can see what you type. Each adds its own policy and its own risk.
What should you never paste into an AI chatbot?
Passwords, OTPs and recovery codes. No legitimate AI task needs them. If a chatbot-style interface asks for them, you are not dealing with a genuine service.
Citizen ID numbers and ID scans. Your citizen ID number, ID card photos, passport pages and household papers can be misused for account fraud. Blank them out if you need help reading a form.
Bank and card details. Account numbers with names, card numbers and statements belong in your banking app, not in a chat history.
Other people’s personal data. Customer lists, employee records, students’ grades and patients’ notes belong to those people. Pasting them into a consumer tool may break your organisation’s rules and the law.
Health details you would not tell a stranger. A chatbot can explain medical terms, but describe your situation in general terms and make decisions with a doctor.
Confidential business information. Unreleased financial results, pricing strategy, contract drafts under non-disclosure agreements and acquisition plans should stay inside approved systems.
Source code that contains secrets. API keys, database passwords and private certificates often hide in configuration files. Strip them out before asking for coding help.
Unchecked screenshots. A quick screenshot can include notifications, names and numbers you did not notice. Crop before you upload.
Privacy settings worth checking today
Menus differ between apps, but most major chatbots offer similar controls. Ten minutes is enough to go through them.
- Find the training setting. Look for options about improving the model or using your content for training, and decide consciously whether to allow it.
- Check history and retention. See whether history can be switched off, or whether a temporary chat mode exists for sensitive questions.
- Review memory. If the app remembers facts about you, open the memory list, delete what you do not want kept and consider turning the feature off.
- Audit shared links. Find the list of conversations you have shared and revoke the links you no longer need.
- Remove unneeded connections. Disconnect email accounts, cloud drives or calendars you linked once for a test. Every connection widens what the assistant can read.
- Secure the account itself. Use a strong, unique password and two-factor authentication; your chat history may reveal more about you than your inbox.
- Prune extensions and keyboards. Uninstall browser extensions and third-party keyboard apps you do not fully trust; they can read whatever you type into any chatbot.
How can a company use AI without leaking its data?
Most data leaks through AI at work are not hacks. They are well-meaning employees pasting a customer complaint, a contract or a spreadsheet into a free chatbot to save time. Banning AI rarely works, because people simply use it on their phones. A better approach is to give staff an approved tool and clear rules.
For a small company, the rules can fit on one page: which AI tools are approved and on which plan; which kinds of data may never be entered; how to anonymise examples by replacing names with roles and removing figures; and who to ask when unsure.
Anonymising is a skill worth teaching. Instead of pasting a customer’s full name, company and the exact amount owed, write “a long-standing customer has an overdue invoice; draft a polite reminder that keeps the relationship”. The AI does the job just as well, and nothing sensitive leaves the company.
Where data genuinely must be processed by AI, such as summarising large volumes of customer email, use a business plan with a proper data processing agreement or a model hosted inside your own systems, and involve whoever handles personal data compliance. Vietnam has its own personal data protection rules, so confirm your obligations with a lawyer or the relevant authority rather than assuming a foreign provider’s policy covers you.
The everyday test is simple: if this text appeared on a public website tomorrow, would it harm a customer, a colleague or the company? If the answer is yes, it does not go into a chatbot without approval.
Frequently asked questions
Do AI chatbots train on my conversations?
Some consumer services do by default and offer an opt-out, while business plans usually do not. Check the current privacy settings and policy of the exact app and plan you use.
If I delete a chat, is it gone?
It disappears from your history, but providers may keep data for a period for security or legal reasons. Deletion reduces exposure; it is not a guarantee.
Is temporary chat mode fully private?
It usually keeps chats out of your history and out of training, but the provider still processes the text and may retain it briefly. Treat it as reduced exposure, not secrecy.
Can my employer see what I type into AI tools?
On company-managed accounts and devices, administrators often can, depending on the plan and internal policies. Assume work accounts are visible to the employer and keep personal matters on personal tools.
Are AI models running on my phone more private?
Generally yes, because the text does not have to leave the device, but many apps mix on-device and cloud processing. Check which features run locally before assuming.
Is it safe to upload my CV or ID card to a chatbot?
A CV without your ID number, home address and phone number is low risk for writing help. ID card images should not be uploaded; type only the fields you need help with and hide sensitive numbers.
