Home / Blog / Chat & assistants

Data security when using AI tools: what you should not paste into a chatbot

Classify data by sensitivity, learn what to check in a provider's policy, and follow practical rules so individuals and small businesses can use AI more safely.

Edited by PickYourToolUpdated: 2026-10-265 min read

Pasting a contract into a chatbot to get a summary is very convenient, and it can easily cause trouble if you do not know where that data goes. This article is not meant to scare you away from AI, but to help you draw clear boundaries: what can go in, what cannot, and what to check before trusting a tool.

Where your data goes when you press Send

When you type content into an online AI service, that content is sent to the provider's servers for processing. Depending on the service and the plan, the provider may:

• store the conversation history in your account;
• let human reviewers see some conversations to improve quality or investigate abuse;
• use conversations to train models, unless you turn this option off or use a plan that commits not to train;
• pass data to infrastructure subcontractors (cloud, storage).

The specific terms differ between providers, between free, personal and business plans, and between the app and the API. So the exact answer is always in the current privacy policy and terms of service for the exact plan you use, not in an article like this one.

Sort data into three levels

Level 1: public. Content already published on a website, press releases, products already on sale. You can use it freely.

Level 2: internal. Marketing plans, quotes not yet sent, workflows, internal emails that contain no personal data. Use it after reading the policy, turning off the option to use your data for training if there is one, and removing proper names when they are not needed.

Level 3: sensitive. This covers personal data of customers and employees (full names with phone numbers, addresses, ID numbers, bank accounts), health information, contracts with confidentiality clauses, proprietary source code, API keys and passwords. Do not put it into public AI tools. If you truly need to, use only a business plan or a privately deployed service that a qualified person has assessed.

Five things to check in a policy

1. Training: is your data used to train models, and can you turn that off?
2. Retention: how long is it kept, how is it deleted, is it really removed from the systems?
3. Readers: does any human read the conversation content, and in what cases?
4. Location and transfers: where is the data processed, is it transferred abroad?
5. Certifications and contractual commitments: is there a data processing agreement (DPA) and independent security certification for the business plan?

If you cannot find a clear answer to all five points, treat the tool as not yet safe for Level 2 data or above.

Risks beyond storage

Security when using AI is not only about what the provider stores. OWASP, the organization specializing in application security, lists risks specific to applications that use language models, including:

• Prompt injection: content hidden in a web page or document makes the model follow an attacker's instructions. This is especially worrying when the assistant has permission to act, such as sending email or reading files.
• Leakage of sensitive information through answers.
• Excessive permissions: giving an AI agent access to more than it needs.
• Over-trusting the results, for example running code written by AI without reading it.

For tools with "agent" features (browsing the web, sending mail, operating on files by themselves), the safety principle is to grant minimum permissions, keep test accounts separate from real accounts, and require confirmation before actions that cannot be undone.

Practical rules for individuals and small businesses

• Anonymize before pasting. Replace real names with "Customer A", and specific amounts with ranges.
• Separate accounts. Use a dedicated work account and turn on two-factor authentication.
• Turn off history or training if the service allows it and you do not need the history.
• Never put passwords, API keys or OTP codes into any chatbot, even "just to get help fixing an error". If you already did, change the key immediately.
• Write a one-page rule for your team: what is allowed, what is banned, which tools are approved.
• Check extensions and third-party apps. Many browser "AI" extensions ask for permission to read every page you open.
• Reread the policy when you get a notice of change. Terms can change over time.

When the data involves other people

If you process personal data of customers or employees, you may have legal obligations, for example needing a lawful basis, notifying the people concerned, or limiting transfers abroad. The specific rules depend on the country and the industry. In Vietnam, the new personal data protection law has been in effect since 2026; in Europe there is the GDPR. This article is not legal advice; if your business handles large amounts of personal data, ask a lawyer or a compliance specialist.

If you sent sensitive data by mistake

1. Delete the conversation and, if available, remove it from the tool's history or memory.
2. If it was a password, API key or token, revoke it and create a new one immediately.
3. Write down what was sent and when.
4. If it was customer data, tell the person in charge and check any notification obligations.
5. Update your internal rules to avoid a repeat.

Conclusion

Using AI safely does not require you to be a security expert. Classify your data, read the policy for the exact plan, anonymize what you can, and do not put secrets into a chatbot. For important work, choose a business plan with written commitments. Finally, remember that every policy can change, so your habit of caution matters more than any promise.

This article is general reference only and does not replace legal or professional advice. See also the editorial, affiliate and review policy.