AI chatbots have become useful enough that people increasingly treat them like another person sitting beside them.
- First: AI Chatbots Aren’t Private Diaries
- 1. Never Share Your Passwords
- 2. Never Paste API Keys or Private Access Tokens
- 3. Don’t Share Your Credit or Debit Card Details
- 4. Bank Account Information Deserves Similar Treatment
- 5. Government Identification Documents Should Be Treated as Highly Sensitive
- 6. Be Careful With Medical Records
- 7. Confidential Business Information Can Be Even More Complicated
- 8. Don’t Upload an Entire Confidential Document When Three Paragraphs Will Do
- 9. Other People’s Personal Information Isn’t Yours to Give Away
- 10. Be Especially Careful With Children’s Information
- 11. Exact Location Information Is Often Unnecessary
- 12. Small Pieces of Information Can Become Sensitive When Combined
- 13. Cryptocurrency Seed Phrases Are an Absolute No
- What Can You Safely Share Instead?
- Privacy Settings Help, but They Don’t Replace Good Judgment
- What If You’ve Already Shared Something Sensitive?
- A Simple Rule Before You Press Enter
You can paste an email and ask for a better reply, upload a contract for an explanation, troubleshoot a computer problem, analyze a spreadsheet, discuss financial decisions, or ask questions you might feel uncomfortable asking another person.
That convenience creates a new privacy problem: it is extremely easy to provide an AI system with far more information than it actually needs.
In 2026, major AI providers offer increasingly sophisticated privacy controls, and business or enterprise AI products can have significantly different data protections from ordinary consumer accounts. However, the safest principle remains surprisingly simple: if an AI doesn’t need a piece of sensitive information to complete the task, don’t give it that information.
Here is what that means in practice.
First: AI Chatbots Aren’t Private Diaries
It would be misleading to claim that everything you type into an AI chatbot automatically becomes public or that every provider uses every conversation to train AI models. Neither statement is true.
Different AI services have different policies.
For example, ChatGPT allows users to control whether conversations contribute to model improvement, while Temporary Chat conversations don’t appear in history, don’t create memories, and aren’t used to train models. Google similarly provides activity and temporary-chat controls for Gemini, while Microsoft’s consumer Copilot provides controls over whether conversations can be used for model training.
Enterprise products can operate under still different rules. Microsoft states that prompts and responses in Microsoft 365 Copilot aren’t used to train foundation models, for example.
The important lesson is therefore not “AI companies read everything.”
It is that you should understand the privacy model of the particular AI service and account you’re using before entering sensitive information.
1. Never Share Your Passwords
Your password is one of the clearest examples of information an AI assistant almost never needs.
Don’t paste your email password into a chatbot and ask whether it’s strong. Don’t provide your banking password while troubleshooting a login problem, and don’t include passwords when pasting configuration files, internal documentation, screenshots, or support conversations.
If you want an AI to evaluate your password strategy, describe its characteristics instead. You could say that your password contains 18 random characters and is unique to one account without revealing the password itself.
The same rule applies to:
- Password-manager master passwords
- Recovery codes
- One-time authentication codes
- Backup codes
- Security-question answers
- Passcodes and PINs
An AI can explain how authentication works without knowing the credentials that authenticate you.
2. Never Paste API Keys or Private Access Tokens
This mistake is particularly relevant to programmers.
Suppose your code isn’t working, so you paste the entire configuration into an AI assistant:
API_KEY=...
DATABASE_PASSWORD=...
PRIVATE_TOKEN=...
The chatbot doesn’t need the real secret to debug most problems.
Replace credentials with obvious placeholders before submitting code:
API_KEY=YOUR_API_KEY
DATABASE_PASSWORD=REDACTED
Developers should apply the same rule to private SSH keys, cloud credentials, database connection strings containing passwords, OAuth tokens, session cookies, webhook secrets, cryptocurrency seed phrases, and other machine credentials.
Treat these values exactly as you would treat passwords.
3. Don’t Share Your Credit or Debit Card Details
An AI chatbot does not need your full card number to explain why a transaction might have failed.
It doesn’t need the CVV printed on your card, either.
Avoid entering combinations of financial information such as your full card number, CVV/CVC, PIN, online-banking credentials, or authentication codes.
You can usually obtain the same assistance by describing the problem:
Bad: “My Visa 4123 XXXX XXXX 5678 with CVV 123 isn’t working.”
Better: “My debit card keeps being declined for online international purchases. What are the common causes?”
The second version gives the AI everything it needs to answer the question without exposing financial credentials.
4. Bank Account Information Deserves Similar Treatment
Sometimes you may legitimately want AI help analyzing financial documents.
Perhaps you want to understand a bank statement, calculate expenses, categorize transactions, or prepare a household budget.
That doesn’t mean the AI needs every identifier printed on the document.
Before uploading financial records, consider removing or obscuring account numbers, IBANs where unnecessary, card information, customer identifiers, barcodes, QR codes, addresses, and other information irrelevant to the analysis.
A chatbot can calculate that you spent €460 on restaurants without knowing the number of the bank account that paid for dinner.
5. Government Identification Documents Should Be Treated as Highly Sensitive
Be extremely cautious about uploading photographs or scans of:
- Passports
- National identity cards
- Driving licences
- Residence permits
- Social-security documents
- Tax-identification documents
- Birth certificates
These documents frequently combine several valuable pieces of identity information in one place: your legal name, photograph, date of birth, document number, nationality, signature, and other identifiers.
If you want an AI to explain something written on an identity document, provide only the relevant text where possible.
If an AI service specifically requires identity verification as part of an official process, that’s different from casually pasting the document into a normal conversation. Verify that you’re using the legitimate provider and understand why the information is being requested.
6. Be Careful With Medical Records
Health questions are one of the areas where AI can be genuinely useful, but they also tempt users to provide enormous amounts of sensitive personal information.
An AI may be able to explain what a medical term means without knowing your full name, home address, national identification number, insurance number, telephone number, hospital patient ID, and every other identifier printed across a medical report.
If you upload a medical document for general explanation, consider redacting identifying information first.
You can also often describe the relevant information directly. For example, providing the name and value of a laboratory measurement may be sufficient for a general explanation without uploading the entire document.
Microsoft explicitly advises consumer Copilot users not to provide confidential or sensitive personal information they wouldn’t want processed under the service’s stated privacy practices.
AI can help you understand medical terminology, but minimizing unnecessary personal information remains sensible.
7. Confidential Business Information Can Be Even More Complicated
This is where individual privacy becomes a corporate security problem.
Imagine an employee receives a confidential 80-page document from a client and uploads the entire thing to a personal AI account because they want a five-paragraph summary.
That document might contain:
- Customer information
- Employee records
- Unreleased financial results
- Pricing agreements
- Internal strategy
- Proprietary algorithms
- Trade secrets
- Acquisition discussions
- Legal advice
- Product roadmaps
- Source code
- Security vulnerabilities
The employee may have saved an hour of reading while potentially violating internal information-security policies, contractual confidentiality requirements, or other obligations.
Before using AI with workplace information, check whether your employer has an approved AI platform and AI-use policy.
The distinction between consumer and enterprise products matters here. Enterprise AI services can offer different retention, training, administrative, contractual, and security protections. For example, Microsoft says Microsoft 365 Copilot prompts and responses are not used to train its foundation models.
Don’t assume the privacy rules for your company’s approved AI environment are identical to those of a free personal chatbot—or vice versa.
8. Don’t Upload an Entire Confidential Document When Three Paragraphs Will Do
Data minimization is one of the best habits you can develop when using AI.
Suppose you have a 70-page employment contract but only want to understand a paragraph about notice periods.
The chatbot probably doesn’t need all 70 pages.
Provide the relevant section and remove names, addresses, signatures, employee numbers, compensation information, and other details if they aren’t needed for the question.
The same principle works for invoices, tax documents, legal agreements, spreadsheets, emails, source code, medical reports, and customer records.
Ask yourself:
“What is the minimum information the AI needs to answer this question properly?”
Then provide that.
9. Other People’s Personal Information Isn’t Yours to Give Away
Privacy isn’t only about protecting yourself.
Suppose a colleague sends you a private email describing a family problem. You paste the entire message into an AI chatbot and ask:
“What do you think this person is really feeling?”
You’ve now submitted somebody else’s private correspondence to an external service without necessarily considering whether they would be comfortable with that.
The same issue applies to customer records, children’s information, employee evaluations, private photographs, medical details, addresses, telephone numbers, relationship conversations, and identifying information about friends or relatives.
Google’s current Gemini privacy information explicitly notes that users should not submit confidential information they wouldn’t want reviewers to see or Google to use under the applicable settings.
Before sharing information with AI, ask a useful second question:
“Does this information belong to me?”
10. Be Especially Careful With Children’s Information
Children’s information deserves additional caution because a child cannot always meaningfully decide how their data should be shared.
Avoid unnecessarily submitting a child’s full name together with identifying details such as their home address, school, exact daily schedule, identification documents, medical records, private photographs, or detailed location information.
You can still ask useful questions without identifying the child.
Instead of:
“My 11-year-old daughter Jane Smith attends Green Valley School at 123 Main Street and walks home alone at 15:20. Is this route safe?”
you could ask:
“What factors should a parent consider when deciding whether an 11-year-old can safely walk home from school?”
You retain the useful part of the question while removing information the AI doesn’t need.
11. Exact Location Information Is Often Unnecessary
Your city can be useful context.
Your exact home address often isn’t.
If you’re asking for restaurant recommendations, saying you’re in central Rome may be sufficient. If you’re troubleshooting a delivery problem, the AI probably doesn’t need your apartment number.
Be particularly cautious about combining precise location with routines:
“I live alone at [exact address], leave every weekday at 7:30 a.m., and return around 7 p.m.”
No ordinary AI question needs that complete picture.
The problem isn’t necessarily any individual fact. It’s the combination.
12. Small Pieces of Information Can Become Sensitive When Combined
This is an important privacy concept.
Perhaps you’ve never uploaded your passport or credit card, so you assume you’ve shared nothing particularly sensitive.
But across multiple conversations, you might reveal:
Your full name, birthday, employer, city, partner’s name, children’s names, travel dates, financial situation, health concerns, hobbies, daily routine, and upcoming vacation.
Individually, some of those details appear harmless.
Together, they create a surprisingly detailed profile.
This matters even more as AI services introduce memory and personalization features. ChatGPT, for example, provides controls over memory as well as whether conversations can contribute to model improvement.
Periodically review what your AI service stores, remembers, and allows you to delete.
13. Cryptocurrency Seed Phrases Are an Absolute No
If you use cryptocurrency, your wallet’s seed or recovery phrase deserves the same treatment as the keys to a bank vault.
Never paste it into an AI chatbot to ask whether it is valid, how to import it, or why your wallet isn’t working.
Anyone possessing the appropriate recovery phrase may potentially gain control of the associated wallet. Unlike a forgotten social-media password, blockchain transactions may also be extremely difficult or impossible to reverse.
An AI assistant can explain wallet recovery using an example phrase.
It does not need yours.
What Can You Safely Share Instead?
The solution isn’t to avoid AI whenever a topic involves personal information. Instead, learn to abstract, redact, and minimize.
| Instead of Sharing… | Share This |
|---|---|
| Actual password | Length and type of password |
| Credit-card number | Type of card/problem |
| Full bank statement | Relevant transactions with identifiers removed |
| Passport scan | Relevant non-identifying text |
| Complete medical record | Relevant result or symptom information |
| Confidential 50-page contract | Relevant clause, if permitted |
| Customer database | Anonymized sample data |
| Production API key | ___SOP_V5_PROTECTED_23___ placeholder |
| Exact home address | General area or city |
| Private email with names | Redacted message |
| Crypto seed phrase | A fictional example |
| Employee personal records | Anonymized scenario |
This is one of the most useful habits for working with AI because it doesn’t substantially reduce what the system can do.
You still receive assistance while exposing far less information.
Privacy Settings Help, but They Don’t Replace Good Judgment
AI privacy controls are worth understanding.
In ChatGPT, users can disable the use of new conversations for model improvement through Data Controls. Temporary Chat doesn’t appear in history, doesn’t create memories, and isn’t used to train models; OpenAI says Temporary Chats are retained for up to 30 days for safety purposes before deletion.
Google’s Gemini also offers controls around saved activity and temporary chats. Its current privacy documentation explains that when Keep Activity is enabled, activity can be used to provide, develop, and improve services, including generative AI models, with some conversations reviewed by humans; Google specifically advises users not to enter confidential information they wouldn’t want reviewed or used under those settings.
Microsoft’s consumer Copilot similarly allows users to opt out of conversation use for model training, while conversations are stored for 18 months by default according to its current privacy FAQ.
These differences demonstrate why you shouldn’t assume every chatbot has identical rules.
Privacy settings can reduce exposure, but they are not permission to paste every secret you possess into a chatbot.
What If You’ve Already Shared Something Sensitive?
Don’t panic, but don’t ignore it either.
If you’ve accidentally pasted a password, API key, authentication token, recovery code, or another credential into an AI conversation, treat the credential as potentially exposed and rotate or replace it. Changing the secret is more important than merely deleting the conversation.
If you’ve uploaded a document containing sensitive personal information, review the provider’s deletion and retention controls. Be aware that deleting a conversation may not always mean every associated item is handled identically; for example, OpenAI’s current documentation notes that files saved to ChatGPT’s Library are managed separately from chat deletion.
If the information belongs to your employer or another organization, follow its incident-reporting procedure rather than attempting to quietly fix the mistake yourself. Security or privacy teams may need to determine whether additional action is necessary.
A Simple Rule Before You Press Enter
Before sending something to an AI chatbot, imagine that the prompt contains three layers.
The first is what the AI actually needs.
The second is useful context that improves the answer.
The third is information you included simply because it happened to be present in the document, screenshot, email, or story.
Remove as much of that third layer as possible.
You don’t need to become paranoid about using AI. Modern AI services have real security and privacy controls, and enterprise systems can provide substantial organizational protections.
You simply need to stop treating the prompt box as a place where every piece of available information automatically belongs.
Give the AI enough information to solve the problem—not enough information to become the problem.
