Wondering what not to put into ChatGPT or another AI tool? Start with one rule: if the AI doesn’t need a piece of information to do the job, don’t give it that information.
That is the safest useful rule.
You do not have to stop using ChatGPT, Claude, Gemini, Copilot, or whatever new AI assistant appeared while somebody in marketing was renaming the previous one. You just need to stop treating the chat box like a private desk drawer.
Before you paste something in, remove the details that are not necessary for the task. Real names can become Client A. An account number can become [ACCOUNT NUMBER]. A sensitive employee issue can become a generic description of the situation.
Most of the time, the AI will still have everything it needs to help.
Keep the task. Remove the unnecessary identity.
A full email thread. A contract. A spreadsheet. Meeting notes. A customer complaint. An employee review. A medical document.
The useful part might be three paragraphs. The rest may contain names, phone numbers, financial details, internal comments, signatures, addresses, account numbers, or information about somebody who never agreed to become supporting cast in your AI experiment.
The practical habit is simple:
Keep the task. Remove the unnecessary identity.
If you want AI to summarize a customer complaint, it probably needs the complaint. It probably does not need the customer’s full name, phone number, account number, home address, and order history going back to the Bush administration.
What not to paste into a general-purpose AI tool
Unless you are using an approved system specifically configured for the information involved, keep these out:
- Passwords, login credentials, recovery codes, API keys, or security tokens
- Credit card numbers, bank account numbers, routing numbers, or other financial identifiers
- Social Security numbers, passport numbers, driver’s-license numbers, or other government IDs
- Private medical or health information tied to an identifiable person
- Confidential business information such as nonpublic financials, strategy, unreleased product plans, proprietary code, or material covered by an NDA
- Attorney-client privileged information or other legally protected material
- Private information about another person when they have not agreed to its use
- Personal information about minors
- Sensitive employee, customer, student, patient, or client records that your workplace policy does not allow you to put into that tool
This is not because every AI company is waiting behind a curtain reading your spreadsheet. It is because you should know what protection applies before you hand over sensitive information, not after.
The part people miss: ordinary files carry extra baggage
The risky information is often not the thing you are asking about.
Suppose you want AI to rewrite an email to a customer. You paste the whole thread. Buried below the message are:
- the customer’s phone number
- an internal note from a coworker
- a quoted billing reference
- somebody else’s email address
- a signature block with direct contact information
None of that was necessary to rewrite the email. It just came along because copying everything was easier.
The same thing happens with spreadsheets, PDFs, meeting notes, resumes, legal documents, and screenshots.
Before uploading or pasting a file, ask:
What is inside this besides the part I actually need help with?
That ten-second check prevents a surprising amount of nonsense.
Training, storage, and human review are not the same thing
This is where privacy discussions usually become a swamp of settings, policies, and words that sound interchangeable but are not.
An AI provider may have separate rules for:
- whether chats are used to improve or train models
- whether chats appear in your history
- how long data is retained
- whether content can be reviewed for safety, abuse prevention, support, or feedback
- what happens when you delete a chat
- what protections apply to business or enterprise accounts
As of September 2026, current consumer versions of major AI products offer different controls around model improvement and activity history. OpenAI lets personal ChatGPT users turn off use of new conversations for model improvement, while Temporary Chat has separate retention rules. Anthropic currently uses an opt-in model for consumer Claude improvement, with specified safety-review exceptions. Google Gemini personal accounts use a Keep Activity setting that changes how future chats are handled. Microsoft distinguishes consumer Copilot from work and school experiences with enterprise data protection.
The useful conclusion is not to memorize four privacy policies. Those will change approximately five minutes after somebody screenshots them.
The useful conclusion is this:
Turning off training does not automatically mean zero storage, zero retention, or zero review. Check the specific product and account you are using.
A work account can be different from a personal account
This matters a lot if you use AI at work.
Business and enterprise AI products often have stronger data-handling commitments than consumer accounts. Several major providers currently state that commercial customer prompts and outputs are not used to train their foundation models by default.
That still does not mean you can paste anything you want into them.
Your employer may have rules about:
- which AI tools are approved
- what customer or employee information may be entered
- confidential documents
- regulated data
- record retention
- intellectual property
- legal review
If your company has an AI-use policy, use it. If it has an approved enterprise AI account, use that instead of your personal account for work material.
Convenience is not a compliance program. Humanity has tried similar shortcuts before. The paperwork always wins eventually.
The safer replacement habit: use placeholders and excerpts
You can usually keep the usefulness while stripping out the risky details.
Hypothetical example. Instead of this:
“Jordan Example at Example Dental owes invoice 84731 for $6,842.17 and has not paid since August 12. Write a collection email.”
Use this:
“Client A has an overdue invoice for a mid-four-figure amount and is about three weeks late. Write a firm but professional collection email that preserves the relationship.”
The AI still knows the job.
Other easy substitutions:
- real name → Client A or Employee B
- exact account number → [ACCOUNT NUMBER]
- precise salary → [SALARY] or a range, if the exact figure is unnecessary
- company-secret project name → Project X
- full document → only the paragraphs relevant to the question
- real case → a short neutral summary of the facts the AI actually needs
This works especially well when you give the AI clearer instructions about the task, audience, constraints, and desired output.
That is the same basic idea covered in Better Instructions Beat Better Prompts: better context does not mean dumping every available fact into the chat.
The 20-second test before you paste anything
Ask four questions:
- Does the AI actually need this detail?
If no, remove it. - Would I be uncomfortable if this information were stored outside my direct control?
If yes, stop and check the tool and account first. - Does this information belong to somebody else?
If yes, be much more cautious. - Am I under a legal, professional, contractual, or workplace obligation to protect it?
If yes, use only an approved system and follow the applicable policy.
That is a better test than “Is AI safe?” because there is no single answer to that question.
The tool, account, settings, information, and context all matter.
What if you already pasted something sensitive?
Do not turn a mistake into a hobby.
Take the practical steps that fit what you exposed:
- Delete the chat or activity if the provider gives you that option.
- If you exposed a password, API key, access token, or other credential, rotate or revoke it immediately. Deleting the chat is not a substitute for changing the credential.
- Check the provider’s current privacy and deletion documentation so you understand what deletion does and does not mean.
- If the information belongs to your employer, customer, client, patient, student, or another person, follow the relevant internal reporting or privacy process if required.
- Change your workflow so the same thing does not happen again. Placeholders and excerpts are considerably cheaper than an incident review meeting.
The practical rule to remember
You do not need to be frightened of AI tools.
You do need to stop giving them information they do not need.
Use the tool for the job. Strip out the identity, credentials, confidential details, and unrelated baggage first. Check the product’s current privacy controls when the information is sensitive. For work, use the approved account and follow the policy that applies to you.
AI is very good at working with structure, language, and context. It rarely needs the real account number, real patient name, real employee name, or full confidential document to help you think through a task.
That is the point:
Use AI without making the data exposure part of the workflow.
If you want more practical ways to use AI without turning it into a second career, get the newsletter.
Frequently Asked Questions
Do AI companies read my individual conversations?
Policies vary. Major providers use automated systems, and some allow limited human review for purposes such as safety, abuse prevention, support, feedback, or product improvement. The exact rules depend on the provider, product, account type, and settings.
The safer habit is to remove sensitive details before sharing them rather than relying on an assumption that nobody could ever review the content.
Is it safe to use AI for work tasks?
Often, yes, but the account and information matter.
Use your employer’s approved AI tool and follow its policy. A business or enterprise account may have stronger data protections than a personal consumer account, but that does not override confidentiality, privacy, or regulatory obligations.
Does turning off AI model training make a chat completely private?
No.
A training opt-out controls one use of the data. Retention, chat history, safety review, deletion timing, and legal obligations may be handled separately.
Check the current policy for the specific product you use.
Does removing names and numbers make AI less useful?
Usually not when those details are irrelevant to the task.
AI can draft, summarize, compare, restructure, and reason through a scenario using placeholders such as Client A, Employee B, or [ACCOUNT NUMBER].
Keep the context that affects the answer and remove the identity that does not.
What should I do if I pasted a password or API key into an AI chat?
Rotate or revoke it immediately. Treat the credential as exposed.
You can also delete the chat where possible, but deleting a conversation does not make an exposed credential safe again.
