What Should You Never Put Into an AI Tool?
Last Updated: September 2026
What not to put into AI is a short list, not a long document. Six things belong on it: personal details about a real person, client work you promised to keep private, passwords and keys, work you have not made public, anything tied to a legal case, and recordings of people who did not agree to be taped. Most other prompts are a judgment call. These six are not.
AI Smart Ventures has guided growing businesses through AI adoption since long before prompting was part of normal work. Advisory work keeps turning up the same story. Almost nobody leaks data on purpose. Someone pastes a normal file into a normal-looking box, in a hurry, and learns later what that box really was.
The gap is rarely knowledge. Most people know a password does not belong in a chat window. What they lack is a line they can see in the two seconds before they hit paste, which is when the choice gets made. Without it, the answer defaults to whatever saves time.
Key Takeaways
- Use the stranger test: if you would not email it, unlocked, to someone you have never met, it does not go in a prompt.
- Keep six things out: personal details, private client work, passwords and keys, unpublished work, legal matters, and recordings made without asking.
- Treat every chat as a record: a New York judge held in February 2026 that one man’s chats with a public AI tool were neither private nor privileged.
- Watch the account, not the tool: 45% of staff now use AI often on work devices, up from 15%, and 67% of them sign in with a personal account.
Read together, those four points move the problem off the software and onto the person typing. No setting saves a paste that should never have happened, because the harm lands the moment your text leaves the machine. That is the good news too. A rule you can hold in your head travels to every tool, including the one your team picks up next month.
What Should You Never Put Into an AI Tool?
Six kinds of input, and they hold in any tool. Keep out personal details that point to a real person, client work you agreed to keep private, passwords, keys, and tokens, work you have not made public such as source code or deal terms, anything tied to a legal case, and recordings of people who never agreed to be taped. Each one shares a trait. It is not yours to give away.
The list stays short on purpose. A never-list with thirty entries is a page people skim once and forget, while six items survive a busy Tuesday. None of them names a product, either. Tools change their terms and their defaults, and the six kinds do not change at all. Write them down, add a real example from your own week beside each, and a new starter can use the page on day one.
- Personal details: names, addresses, health notes, anything that points to a real customer or colleague.
- Private client work: contracts, briefs, files, and anything you signed a non-disclosure agreement to read.
- Passwords and keys: logins, tokens and connection strings, including ones buried in code you are fixing.
- Unpublished work: source code, deal terms, plans and results your market has not seen.
- Legal matters: advice from your lawyer, draft deals, notes from an inquiry.
- Recordings without consent: call audio, meeting notes and video where the others never agreed.
Why Do These Inputs Cause Real Harm?
Because a prompt is a handover, not a search. Your text leaves the device, lands on someone else’s system, and sits there under their rules rather than yours. Four things can follow. It can train a future model, rest in a log that outlives your account, be pulled into a lawsuit, or be read by staff at the company during a check. All four are routine, and none needs a hacker.
The harm you get depends on what you sent. Personal details can put you on the wrong side of a privacy law, and those duties differ a lot by market, so read the body that covers you rather than a general rule. Private client work can break a contract you signed. Passwords hand over live access. Unpublished work can lose the legal shield that made it worth guarding, which is the hardest harm here to undo.
What Actually Happens After a Bad Paste?
Three real outcomes, all from 2026. A New York federal judge ruled in February that a man’s chats with a public AI tool were not covered by legal privilege, because the tool’s own terms let the company keep and use what he typed. A California court threw out a trade secret claim on similar grounds. And judges have begun ordering AI firms to hand over user chat logs by the million.
Start with the trade secret ruling, because it catches people out. Secrecy is a status you hold by acting as though it matters. Foley Hoag’s May 2026 note on trade secret cases describes a January ruling where a plaintiff lost because she had built the material through ChatGPT, sharing it with a party that owed her nothing. Nothing was stolen. The cover simply stopped applying the moment she typed it in.
How Do You Decide Before You Hit Paste?
Ask one question. Would I email this, unlocked, to a stranger who has signed nothing? If the answer is no, it stays out of the prompt. The test works because it matches what really happens. Your text goes to a company you have no deal with, is kept on terms you did not write, and may be read by people you will never meet. That is a stranger with an inbox, not a private notebook.
Two follow-ups catch what the first one misses. Ownership comes second: client, staff, and patient files belong to someone who has not agreed to any of this. Staying power comes third, since a prompt can turn up in a court file, a company log, or a search result months later. Run all three in five seconds, then write the answers down, because a test that lives only in your head is not a rule.
- Would I email it to a stranger? If not, it stays out of the prompt, whatever the deadline looks like.
- Is it mine to share: client, staff, and patient files belong to people who never agreed to this.
- Would I mind seeing it quoted back: in a log, a lawsuit, or a search result, months from now.
What Can You Use Instead of the Real Thing?
Send the shape of the problem, not what is in it. Most work you want help with is about form: a tone, a format, an argument, a pattern in a bug. Describe the case in general terms, swap real names and numbers for made-up ones, and ask for a template. You keep the help and skip the handover. Where the real thing must go in, use an account your business owns with model training switched off.
That last step is where teams stop too early. Turning training off narrows one risk. It does not clear the company’s logs, beat a court order, or win back secrecy you gave away. Pair it with a habit: nobody uploads a file they have not opened and read. Generic AI consultants sell the account and stop there, which is why capability building beats a settings screen. AI Smart Ventures observes that habits change fastest when teams practise on real prompts.
| Instead of pasting | Send this |
|---|---|
| A signed client contract | The clause type and your question, no names |
| A customer list | Five made-up rows in the same format |
| A call recording | Your own notes, once you have asked the others |
Agreeing on your never-list takes one short session, and it then holds for years. AI Advisory helps founder-led businesses draw that line and build the AI literacy that keeps it in use, drawing on Applied AI work with 20,000+ professionals trained.
Frequently Asked Questions
What are five things I should never tell ChatGPT?
Passwords and keys, personal details about a real person, client work you agreed to keep private, unpublished work such as source code or deal terms, and anything tied to a legal case. A sixth belongs beside them: recordings of people who never agreed to be taped. The thread through all of them is ownership. Each one belongs to someone who has not agreed to the handover.
What information should I not share with ChatGPT?
Anything you would not email, unlocked, to a stranger. In practice, that covers customer records, staff files, health notes, signed contracts, passwords, unpublished results, and anything a lawyer has touched. Source code is the quiet one, because it often carries live keys as well as real value. General questions about method, wording and approach are fine, since none hands over what belongs to someone else.
What must you never use AI for?
Final calls about people, with no named human owning the outcome. That covers hiring, firing, credit, discipline, and any safety or medical call. Never use it as your only check on a fact you are about to publish or file, either. Several markets now police automated decisions about people more tightly, and those duties vary by country, so check your position with a qualified adviser.
Are my AI chats private after I delete them?
Not reliably. Deleting a chat removes it from your view, while company logs and backups can live on, and a legal hold freezes them outright. In January 2026, a New York court backed an order making OpenAI hand over 20 million stripped ChatGPT logs. Shared links leak too: in July 2025, shared chats turned up in Google results before the feature was pulled.
Does a paid account make it safe to paste anything?
No. A business account with training switched off narrows one risk, which is your text feeding a future model. It does not stop a court order, clear the logs, win back secrecy you already gave away, or fix the fact that a colleague on the call never agreed to be taped. Paid tiers change the terms, not the never-list. Read the terms you are on.
Can I paste source code into an AI tool to fix a bug?
Only from an account your business owns, and only after taking out keys, tokens, connection strings, and any customer records sitting in the file. A leaked key is a live door rather than a slip, so it does damage at once. Where the code carries real value, treat it as unpublished work: build a small standalone example of the bug, and send that instead.
Is it alright to let an AI notetaker join a client call?
Ask first, every time, and note the answer somewhere you can find it later. Taping rules differ by market and, in the United States, by state. A Reed Smith analysis counts thirteen states that need consent from everyone present, and warns that voice matching can trigger separate biometric rules. Where one person on the call sits in such a state, plan for the strictest rule in the room, not the loosest.
Does removing names make a client file safe to paste?
Less risky, but not safe. Records can often be matched back to people once you add other records, and a contract barring you from sharing client work usually says nothing about names. The better move is to describe the case and make up the details, so the tool sees a real shape with no real person in it. Check what your client agreement allows first.
What should I do if I already pasted something I should not have?
Say so the same day, because the useful window is short. Delete the chat and any shared link, change every password or key that appeared, and check whether the account was personal or owned by the business. Then tell whoever looks after data protection, and your client, if their work was involved. Contracts and privacy rules run on clocks that start when somebody knows.
How do we agree a never-list across the whole team?
One working session is usually enough. Bring five real prompts from last week, run each through the stranger test out loud, and write the six kinds on a single page with your own examples beside them. Then practise, because AI literacy sticks through repeat use rather than reading. Book a consultation to run that session and leave with a page your team will follow.
Executive Summary
What not to put into AI comes down to six kinds of input: personal details, private client work, passwords and keys, unpublished work, legal matters, and recordings made without asking. The reason is practical, not moral. A prompt is a handover to a company that owes you nothing, and 2026 court rulings have treated those handovers as normal sharing, stripping legal privilege in one case and killing a trade secret claim in another. Use one test at the keyboard: would you email this to a stranger?
What Should You Do Next?
Write the six kinds on one page this week, with a real example from your own work beside each. Run last week’s five most awkward prompts through the stranger test, and check whether each went through a company account or a personal one. Settle the account question first, since it decides where everything else lands.
AI Smart Ventures offers AI Advisory for growing businesses that want practical AI adoption without a security team behind them. Schedule a consultation to agree on your never-list and the AI training that keeps it working.
People Also Read
- AI Content Strategy: What It Means for Service Businesses
- What a Correct AI Marketing Budget Actually Looks Like
About the Author
Nicole A. Donnelly is the Founder of AI Smart Ventures and an AI Adoption Specialist with 20 years of experience as a founder and CEO and over a decade leading AI adoption initiatives. She helps businesses integrate artificial intelligence with clarity and confidence, driving innovation and sustainable growth. Nicole has trained over 20,217 professionals in Applied AI, delivered 624 workshops, and worked with close to 1,000 organizations across diverse industries.
Expertise: AI Transformation, AI Strategy, AI Implementation, AI Adoption, Applied AI, Marketing, Business Operations
Disclaimer: This content is for informational purposes only and does not constitute professional business or technology advice. Results vary based on industry, existing systems, and implementation commitment. Contact AI Smart Ventures for a consultation regarding your specific situation.
