What You Should Never Share With AI Tools: Privacy Guide for 2026

A freelance designer is running late on a client report. To save time, she pastes the full signed contract into a chatbot — client’s home address, bank details and all — and asks for a quick summary.

The summary is excellent. She has also just handed a private legal document to an outside company, and there is no way to pull it back.

This happens every day. A chat window feels personal, almost like a private notebook. It is not. On most consumer plans what you type can be stored on a company’s servers, retained for a set period, reviewed by people in some cases, and — unless you have changed a setting — used to help train future models.

This guide covers what to keep out of any AI tool, why each category carries real risk, and how to get the same quality of help without exposing anything you cannot take back. It is not about fear. It is about using powerful tools deliberately.

The short version

• Never paste passwords, API keys, full financial or card numbers, government IDs, confidential client or company data, private health or legal details, or proprietary source code into a public AI chat.
• Assume anything you type may be stored, seen, or used for training. On consumer plans this is usually the default unless you opt out.
• Redact and use placeholders. Replace real secrets with [API_KEY], [CLIENT_NAME], or [AMOUNT]. The AI can still help; the secret stays with you.
• Match the tier to the sensitivity. Business, Enterprise and API tiers generally do not train on your data by default. Free consumer chat is the least private surface.
• The structure of a prompt is what makes it work — not the sensitive data inside it.

What to keep out, at a glance

The pattern is consistent. The chatbot almost never needs the actual secret to help you. It needs the shape of the problem.

1. What actually happens to what you paste

When you paste text into a consumer AI tool, it does not vanish when you close the tab. Three things are worth understanding.

Deletion is a promise, and legal processes can override it. In May 2025 a US magistrate judge ordered OpenAI to preserve output log data that would otherwise have been deleted, as part of the New York Times copyright case. That broad order has since ended — OpenAI has confirmed its preservation obligations ceased on 26 September 2025, and deleted chats are once again removed from your account immediately and purged within about 30 days. But a limited set of data from April to September 2025 remains under legal hold, and the episode makes the general point: at any provider, deletion policies can be suspended by a court. Do not rely on the delete button to protect something you should never have pasted.

Your chats may train the model by default. Anthropic changed its consumer terms starting 28 August 2025, giving Free, Pro and Max users — including Claude Code — a choice about whether their conversations are used for training, with a deadline of 8 October 2025 to decide. Allowing it extends retention to five years; declining keeps the standard 30 days. Claude for Work, Government, Education and API access are excluded entirely. Other providers run comparable schemes with their own settings and defaults.

People leak sensitive data far more often than they think. Cyberhaven’s 2026 AI Adoption and Risk Report, published in February 2026, found that 39.7% of all data movements into AI tools involve sensitive data, and that the average employee enters sensitive data into an AI tool roughly once every three days. Nearly a third of ChatGPT usage it measured came through personal accounts, which sit outside any company’s visibility. The best-known single case is still Samsung, where engineers pasted proprietary semiconductor code and internal meeting content into ChatGPT across a handful of incidents in 2023, prompting the company to restrict public AI tools.

None of this means AI tools are unsafe. It means they are powerful external services rather than private vaults, and a few habits keep you on the right side of that line.

Caption: Pasted text can travel further than the chat window suggests

Where the major tools stand

Defaults change often, so treat this as a starting point and confirm in each tool’s own settings.

The through-line: on consumer plans, assume your chat could be stored and used for training unless you have actively changed the setting.

2. Passwords and login credentials

Never paste a password, passphrase, PIN, or security question answer into an AI tool — not even to check whether it is strong, and not buried inside a longer block of text you are asking the model to tidy up.

A password typed into a chat now sits outside your control and may remain in logs for weeks. If that account also protects email, banking, or work systems, you have handed over the master key. Reused passwords make it worse: one exposed credential can unlock many accounts.

Instead: use a password manager to generate and test strength. If you want help writing a password policy — length, rotation, MFA rules — ask about the rules in general terms, with no real credentials in the prompt.

3. API keys and access tokens

API keys, access tokens, secret keys, OAuth credentials and connection strings are among the most dangerous things to paste into a chatbot. Anyone who obtains one can call services, run up charges, or reach data as if they were you.

Developers do this without meaning to. You paste a config file or an error log to debug it and do not notice the live key sitting in line twelve. Automated scanners actively hunt for leaked keys, and exposure can turn into real cost within minutes.

Instead: strip secrets first and replace them with obvious placeholders like [API_KEY] or sk-XXXXXXXX. The model can still explain the code, find the bug, or refactor the logic — it does not need the real value. If a key has ever been exposed, revoke and reissue it immediately.

4. Banking and payment information

Keep full card numbers, CVV codes, bank account and routing numbers, IBANs and online banking logins out of prompts entirely. This is precisely why several major banks restricted staff use of public chatbots.

Even an innocuous-sounding request — “help me format this bank statement” — can expose account numbers, balances, transaction history and spending patterns that are valuable to fraudsters.

Instead: mask all but the last four digits, or describe the task without the data. For budgeting help, use rounded, non-identifying figures. And remember that an AI tool is not a licensed financial adviser; treat its output as a starting point rather than advice.

5. Personal identity documents

Do not upload or type the contents of passports, national ID cards, driving licences, Social Security or national insurance numbers, birth certificates, or visa documents. These are the raw material of identity theft. A single ID number alongside a name and date of birth can be enough to open an account in someone else’s name.

This applies to other people’s details too. Pasting a colleague’s, client’s, or family member’s personal data into an AI tool can breach privacy law such as GDPR or CCPA, and their trust, even where your intentions are good.

Instead: if you need help with a form or template, ask about the structure and required fields generically, then fill in real details yourself, offline. Never use real ID numbers as examples.

Caption: Remove sensitive details and replace them with placeholders before submitting information to an AI chatbot

6. Confidential client and company data

Client lists, contracts, pricing agreements, unreleased product plans, internal financials, strategy documents, employee records and anything under an NDA should stay out of public AI tools. Sharing them can breach confidentiality agreements, expose trade secrets, and create legal liability for you and your employer.

This is the category that catches professionals out, because it feels productive: summarising a confidential contract, analysing a client’s private data, drafting strategy from internal numbers. But the moment that material enters a consumer chatbot, you have shared protected information with an outside company.

Instead: anonymise aggressively. Swap real names for [CLIENT], [COMPANY] or [PROJECT], and replace specific figures with ranges. If your organisation genuinely needs AI for sensitive work, use a Business or Enterprise tier with a no-training guarantee and proper data controls, not a personal account.

7. Private health or legal information

Medical records, diagnoses, mental-health details, prescriptions and test results are deeply sensitive and, in professional contexts, legally protected — under HIPAA and equivalent regimes. The same caution applies to legal matters: case details, privileged communications, settlement terms, or anything tied to active litigation.

Pasting a patient’s or client’s information into a consumer AI tool can breach professional confidentiality and regulatory obligations. For individuals, health and legal details are exactly what you do not want sitting on a third-party server.

Instead: ask general educational questions — what usually happens in this type of case, what a term means — without identifying anyone. For your own sensitive matters, keep names and specifics out and rely on a qualified professional for actual advice. AI can explain concepts; it cannot be your doctor or your lawyer.

8. Source code and internal business documents

Proprietary source code, internal documentation, infrastructure diagrams, database schemas and security configurations are valuable intellectual property. Pasting them into a public AI tool can reveal how your systems work, including their weaknesses.

Developers face this trade-off daily, because AI genuinely helps with debugging and review. The problem is not using AI for code. It is using a public consumer tool with real proprietary code and live secrets inside it.

Instead: share the smallest generic snippet that reproduces the problem, with secrets, internal URLs and identifying details removed. For ongoing work, prefer enterprise coding tools or business tiers that contractually exclude your code from training. Sanitise logs and config files first — placeholders like [DB_PASSWORD] and [INTERNAL_URL] work fine.

Caption: Confidential files, source code, and internal business information may leave your control once they are shared with an external AI service.

The same task, done safely

Go back to the freelancer and her contract. Here is the risky version and the safe one.

Risky. Paste the entire signed contract — real names, home address, bank details — and ask for a summary.

Safe:

I have a service contract between a design freelancer and a small business client. Summarise the payment terms, deadlines, and cancellation policy in plain language. Use CLIENT and FREELANCER instead of real names. Here are only the relevant clauses: [paste the clauses with names and numbers removed].

Same summary. None of the exposure. She keeps the client’s trust and still saves the time.

Notice what did the work in that prompt: the task, the audience, the output format, the constraint about names. None of it required the client’s bank details. That is the general case, not the exception.

Caption: Placeholders keep the request useful while removing the private values.

Four habits that keep you safe

You do not have to give up AI help to protect private data. These four habits let you keep both.

Use placeholders. Replace real values with labels the model can still reason about. [CARD_NUMBER] instead of a real card number, [CLIENT_A] instead of a client name. The structure of your request survives; the sensitive value never leaves your device.

Share the minimum necessary. Ask what the model actually needs. A chatbot can help you rewrite a difficult message without seeing the sender’s full name, address and account history. Give it the situation and the goal, not the entire file.

Redact before you paste. If you must paste a document, strip names, numbers and account details first. A quick pass usually removes the risky parts while keeping the substance you need help with.

Match the account to the sensitivity. For routine work with no private data, a personal account with training turned off is often fine. For client or company data, a business or team tier — or a temporary chat mode excluded from training — gives you stronger protection.

A five-second pause before you send

• Would I be comfortable if this appeared publicly? If not, do not paste it as-is.
• Is there a secret, ID, or number in here? Redact it or swap in a placeholder.
• Is this someone else’s private data? If so, you likely need their consent, or you should not share it at all.
• Does the model actually need the real data to help? Usually not.
• Am I on the right plan and settings? Training off, temporary chat for one-offs, business tier for sensitive work.
• Have I checked the output? Review before you publish, send, or rely on it.

Caption: Pause before sending and check whether your prompt contains secrets, personal information, financial data, client records, or proprietary content.

Where Promptlywise fits

Here is the reframing that makes this easier: most of the time, the sensitive data is not what makes a prompt work. The structure is. A good prompt needs clear context, audience, goal, tone, constraints and output format. It rarely needs your actual passwords, client names, or account numbers.

Promptlywise helps you turn a rough idea into a clear, structured prompt — choosing a niche, adding context, defining the goal and format — so you get a better result without pasting the risky parts. You can build the instruction around placeholders like [CLIENT] or [FIGURE], then fill in real details privately, only where they are genuinely needed.

You still decide what to share, and reviewing both the prompt and the output before you rely on them remains your job.

Frequently asked questions

Is it safe to share personal information with AI tools? Treat consumer AI chats as external services, not private vaults. General, non-sensitive information is fine. Keep passwords, financial and identity data, health or legal details, and confidential work data out of prompts. When in doubt, redact.

Do ChatGPT, Claude and Gemini use my conversations to train their models? On consumer plans, generally yes unless you turn it off in settings. Business, Enterprise and API tiers typically exclude your data from training by default. Because these defaults change, check the setting in each tool rather than assuming.

If I delete a chat, is it gone? Usually within about 30 days, but not always permanently. A US court order in 2025 required OpenAI to preserve output logs that would otherwise have been deleted; those obligations ended on 26 September 2025, though a limited set of data from that window remains under legal hold. Deletion is a policy, and policies can be overridden.

Can I use AI for coding without leaking secrets? Yes. Remove API keys, passwords, internal URLs and identifying details, and share the smallest generic snippet that reproduces your issue. For ongoing proprietary work, use tiers that contractually exclude your code from training.

How do I make AI tools more private? Turn off model training in privacy settings, use temporary chat for sensitive one-offs, choose business or enterprise plans for confidential work, and apply the minimum-necessary rule.

What should I never type into an AI tool? Passwords and logins, API keys and tokens, full financial or card numbers, government ID numbers, confidential client or company data, private health or legal information, and proprietary source code.

Final thoughts

AI tools are among the most useful assistants ever built, and usefulness and privacy are not opposites. The safest users are not the ones who avoid AI. They are the ones who share deliberately.

Keep secrets, identifiers and confidential data out of your prompts. Redact what you do not need to reveal. Check your settings. Review every output. Do that, and you get the powerful help without handing over information you cannot take back.

Build a safer prompt with Promptlywise and keep the sensitive parts where they belong.

Related reading

• What Makes a Good AI Prompt?
• How to Turn a Vague Idea Into a Structured AI Prompt
• The 7 Most Common AI Prompt Mistakes
• Free vs Paid AI Tools: When Is It Worth Upgrading?
• How to Fact-Check Anything an AI Tool Tells You