AI works better when you give it context.
That is one of the first things people learn when they begin using tools such as ChatGPT, Gemini, Copilot, or other AI assistants. Give the system more information about what you are trying to accomplish, and it can often give you a more useful answer.
But that creates another question:
How much information should you give it?
More importantly:
What information should you never give it at all?
This is where good prompting stops being merely a question of getting a better answer. It becomes a question of privacy, security, authorization, and judgment.
The principle I would use is simple:
Do not give an AI information merely because the information would make its answer better. Give it only information you are authorized to disclose, in an environment appropriate for that information, and only when the information is actually necessary.
Context matters.
But so do boundaries.
1. Never give AI your passwords or authentication secrets
This should be the easiest category.
Do not paste your:
- passwords;
- PINs;
- private cryptographic keys;
- API keys;
- access tokens;
- session tokens;
- recovery codes;
- authentication cookies;
- security-question answers;
- or other credentials
into a general-purpose AI chatbot.
You usually do not need to give the actual secret to get help.
Suppose you are troubleshooting an API authentication error.
The AI may need to know that you are using an API key.
It may need to see the format of the request.
It may need the error message.
It does not need your real API key.
Replace it with something such as:
YOUR_API_KEY_HERE
The difference seems obvious when we are talking about a bank password.
It becomes less obvious when someone is debugging code, pasting a configuration file, or asking AI to troubleshoot a system.
That is where people can accidentally disclose credentials because the secret happens to be embedded inside something else.
Redact before you prompt.
2. Do not paste confidential work information merely because AI could help with it
Imagine that you have a confidential internal document.
You could paste the entire thing into an AI system and ask for a summary.
Technically, the AI might be very good at the task.
But capability is not permission.
Before submitting confidential business information, ask:
Am I authorized to put this information into this particular AI system?
Your employer may have an approved organizational AI environment with specific contractual, security, privacy, and retention protections.
That is very different from deciding on your own to paste confidential material into whichever consumer chatbot happens to be open in your browser.
This can apply to:
- internal strategy documents;
- unreleased financial information;
- customer lists;
- contracts;
- source code;
- system architecture;
- internal investigations;
- proprietary processes;
- acquisition discussions;
- personnel information;
- and trade secrets.
The mistake is assuming:
“I am allowed to see this information, therefore I am allowed to give it to AI.”
Those are not the same permission.
Access to information does not automatically include authority to disclose it to another service.
3. Be especially careful with somebody else's personal information
Your privacy is not the only privacy that matters.
Suppose you want AI to help you understand a situation involving another person.
It can be tempting to provide everything:
their name, address, medical condition, financial situation, workplace problem, private messages, family circumstances, or other identifying details.
Most of the time, the AI does not need all of that.
You can often describe the relevant situation without identifying the person.
Instead of:
“John Smith, who lives at [address], has this medical condition and sent me this private message…”
you may be able to ask:
“A family member with this situation told me the following. Help me think through the general issues.”
The exact amount of redaction depends on the task, but the governing question remains:
Does the AI need to know who this person is?
If the answer is no, do not tell it.
4. Do not casually upload medical, financial, legal, or other highly sensitive records
AI can be extremely useful for explaining complicated information.
That usefulness can encourage people to upload entire records when only a small portion is relevant.
For example, someone may want help understanding one line on a medical report but upload a document containing:
- full name;
- birth date;
- home address;
- medical-record number;
- insurance information;
- diagnoses;
- medication history;
- provider information;
- and other unrelated details.
Or someone may ask about one bank transaction by uploading a full statement containing account numbers, balances, addresses, and transaction history.
A better approach is usually:
- identify the actual question;
- isolate the information needed to answer it;
- remove unrelated identifying information;
- use an appropriate AI environment;
- then ask the question.
That is not only a privacy technique.
It is good prompting.
Relevant context is more useful than indiscriminate context.
5. Do not give AI information that you are legally, professionally, or ethically required to protect
Different people carry different responsibilities.
A lawyer, doctor, government employee, system administrator, pastor, counselor, teacher, accountant, contractor, manager, or business owner may possess information that carries obligations beyond ordinary personal privacy.
The exact rules differ by profession, employer, jurisdiction, contract, and type of information.
AI does not erase those obligations.
If you are not sure whether particular information is permitted in an AI system, the answer is not:
“The chatbot probably keeps it private.”
The answer is:
Find out what your actual authorization and policy allow.
AI should fit inside the rules governing the information.
The rules should not be ignored because AI is convenient.
6. Do not paste an entire document when a small excerpt will do
There is a habit worth developing:
minimum necessary context.
AI prompting often teaches us to add context.
Good AI judgment teaches us to add the right context.
If you need help rewriting one paragraph, consider providing one paragraph instead of the entire confidential report.
If you need help with a spreadsheet formula, you may be able to reproduce the structure using fictional values.
If you need help debugging code, replace production credentials, hostnames, customer identifiers, internal addresses, and other unnecessary details.
If you need help drafting a response to a sensitive email, summarize the situation instead of automatically pasting the entire thread.
More information does not always mean a better prompt.
Sometimes it simply means more unnecessary exposure.
7. Remember that AI products do not all handle data the same way
It is dangerous to think of “AI” as though every product has one universal privacy policy.
They do not.
Consumer products may provide settings controlling history, retention, model improvement, personalization, or temporary conversations.
Organizational and enterprise products may operate under different contractual and technical protections.
Those differences matter.
That means you should understand the environment you are actually using.
Ask questions such as:
- Is this my personal account or an employer-managed account?
- Is this service approved for company information?
- Are prompts retained?
- Can conversations be reviewed?
- Can my content be used to improve models?
- What contractual protections apply?
- What does my organization's AI-use policy allow?
- Am I sending information to connected tools or third-party services?
- Is the data necessary for this task at all?
Do not rely on assumptions carried over from another AI product.
And do not assume that changing one privacy setting magically turns an inappropriate disclosure into an appropriate one.
Settings matter.
Authorization still matters more.
8. Temporary or private modes reduce some risks, but they do not create permission
Some AI products provide temporary-chat modes or controls that change how conversations are stored or used.
Those can be valuable privacy tools.
Use them when appropriate.
But there is an important distinction:
A privacy feature can change what the service does with information. It does not change whether you had the right to disclose the information in the first place.
If a document belongs to your employer and policy says it cannot be uploaded to an external consumer AI service, a temporary-chat button does not override that rule.
If another person's confidential information was entrusted to you, deleting the conversation later does not retroactively create permission to disclose it.
Technology controls can support good judgment.
They cannot replace it.
A simple five-question check before you share information with AI
Before submitting anything sensitive, ask:
1. Do I actually need AI for this task?
Sometimes the safest prompt is no prompt.
2. Does the AI need this specific information?
Remove anything unrelated to the task.
3. Am I authorized to disclose it?
Do not confuse your ability to access information with permission to send it somewhere else.
4. Is this AI environment approved for this type of information?
A protected organizational system and a personal consumer account may have very different rules and protections.
5. Can I redact, anonymize, summarize, or substitute the information?
If fictional or reduced data will accomplish the task, use it.
Those five questions will prevent a surprising number of bad disclosures.
Good prompting is also deciding what not to say
People sometimes think prompting is about giving AI as much detail as possible.
It is not.
Good prompting is about giving the AI the information required to perform a clearly defined task.
That means knowing what to include.
It also means knowing what to leave out.
A useful AI assistant does not need unlimited access to your information.
It needs appropriate access to the information necessary for the work you have actually authorized it to help perform.
That is a much healthier relationship with the technology.
You remain responsible for the boundary.
You decide what the AI receives.
You review what it produces.
You verify what matters.
And you approve what happens next.
AI can assist.
The decision to disclose remains yours.
Practical Takeaway
Before you paste text, upload a document, attach an image, or connect another source of information to an AI system, ask two questions:
Am I authorized to give the AI this information?
Does it actually need this information to do the job?
If either answer is no, leave it out.
When possible, redact, anonymize, summarize, or substitute.
Give AI enough context to help.
Do not give it information simply because you can.
Edward Fong is the author of The Necessary Evil series, including The Necessary Evil: AI Prompting for Everyday People, which focuses on practical AI use while preserving human review, verification, judgment, and approval.
Continue with Book 2
Explore The Necessary Evil: AI Prompting for Everyday People for a practical framework that keeps review, verification, judgment, and approval with the human user.