Most of what people paste into ChatGPT is harmless. Some of it isn't.
This post is a list of the things that should never leave your office. The examples are real. The stats are from public research. None of this is legal advice, and where the answer involves a lawyer or a doctor, talk to one.
A 2026 Cyberhaven study tracked 7 million workers and found that 39.7% of all data movements into AI tools involve sensitive data, including prompts or copy-paste actions. About 32.3% of ChatGPT usage occurs through personal accounts, which means most of your sensitive data going into AI tools is going through accounts your IT team can't see.
The volume of corporate data pasted into AI tools grew 485% between March 2023 and March 2024. That's not a typo. Five times more, in twelve months.
So this isn't a hypothetical risk. It's already happening, probably in your business too.
The famous case: in April 2023, Samsung engineers pasted confidential source code into ChatGPT. Within less than 20 days of allowing ChatGPT access, Samsung experienced three separate incidents where engineers exposed confidential company information. One engineer pasted source code looking for a fix. Another pasted code to optimize a defective-chip test sequence. A third employee converted a smartphone recording of a company meeting to a document file using an AI tool, and entered it into ChatGPT to get meeting minutes.
Samsung banned ChatGPT after the third incident.
Why it matters: once code goes into a public AI tool, you don't get it back. Even on paid tiers, the data has been on someone else's server. If you have any contract that promises confidentiality on your code, you may have just broken it.
What to do instead: redact before pasting. Replace API keys, customer names, and identifying values with placeholders. For real coding work, use enterprise tools with confidentiality contracts in place (Claude Code, GitHub Copilot Business, etc.) and check your client agreements first.
This one changed in February 2026. A federal judge ruled in United States v. Heppner that ChatGPT conversations are not privileged. The defendant used Claude to research legal questions related to the government's investigation after receiving a grand jury subpoena and engaging counsel, but before his arrest. He fed information he had learned from his defense counsel into the AI tool, generated 31 documents of prompts and responses, and then transmitted those documents to his lawyers.
The FBI seized the documents. Defense counsel claimed privilege. The judge said no.
The ruling went further: disclosing information received from counsel to a third-party AI platform could waive privilege over the underlying attorney communications. Meaning, if you take what your lawyer told you and paste it into ChatGPT, you may have damaged the privilege over the lawyer's original advice too.
It gets worse. In May 2025, a federal magistrate judge in The New York Times v. OpenAI ordered OpenAI to preserve every single ChatGPT conversation indefinitely, including ones users had already deleted. So even your "deleted" chats may still be sitting on OpenAI's servers, subject to subpoena.
What to do instead: don't paste anything related to active legal matters into a consumer AI tool. If you have a lawyer, talk to them. If your lawyer wants to use AI on the matter, they need to be using an enterprise-tier tool with proper confidentiality controls, and the use needs to be at their direction.
If you're a healthcare provider or work for one, pasting patient data into a consumer AI tool is a HIPAA violation. The fines start at $100 per violation and scale up sharply for repeated or willful breaches. The Business Associate Agreement protections you need don't exist on free or basic-paid tiers.
As one healthcare law analysis put it: a well-meaning physician pasting patient notes into ChatGPT to save time creates the same HIPAA violation as intentional data theft. Compliance doesn't care about intent.
What to do instead: use HIPAA-compliant AI through providers who will sign a BAA with you, or run AI on hardware you own where data never leaves the office. This is one of the few cases where local AI on your own hardware is genuinely the right answer rather than a paranoid one.
A recurring example from Cyberhaven's research: an executive pastes bullet points from the company's strategy document into ChatGPT and asks it to rewrite it in the format of a PowerPoint slide deck. Their warning: in the future, if a third party asks "what are [company name]'s strategic priorities this year," ChatGPT could answer based on the information the executive provided.
In the 2025 Cyberhaven report, the most common types of sensitive data employees put into AI are source code (18.7% of sensitive data), R&D materials (17.1%), and Sales and Marketing data (10.7%).
What to do instead: use Microsoft Copilot for Business, Claude Team, or ChatGPT Enterprise. These have stricter data-handling than the consumer tiers and contractually exclude your data from training. Or do the work locally.
If you're a bookkeeper, tax preparer, or financial professional, pasting customer financial details into a consumer AI tool exposes you to a contract breach with your customer at minimum, and to professional discipline depending on your license and your state. Most engagement letters with financial clients include confidentiality terms that consumer AI tools straightforwardly violate.
What to do instead: most accounting software has reporting built in. QuickBooks, Xero, FreshBooks all have summary reports that don't require you to paste raw customer data anywhere. If you need AI help, ask the generic question without the data attached.
If you've ever pasted a config file or a script into ChatGPT to debug it, you may have included credentials without thinking about it. Once those are in the chat history, treat them as compromised. Rotate them. Don't argue with yourself about whether they were "really" exposed. They were.
What to do instead: redact before pasting. API_KEY=sk-real-thing-here becomes API_KEY=PLACEHOLDER. Make this a habit so you don't forget on the day it matters.
A 2024 Cyberhaven study found that in March 2024, 27.4% of corporate data employees put into AI tools was sensitive, up from 10.7% a year ago. A lot of that sensitive data is customer information.
Even on paid business tiers, you've now created a record of customer information on a third-party server. That can become a problem in privacy lawsuits, in M&A diligence, in regulatory investigations, or just in customer trust if the customer ever finds out.
What to do instead: rewrite without names. "Help me respond to a customer complaint about a delayed delivery" works without putting the actual customer's name and details into the tool.
This is the rule of thumb that catches almost everything else.
Before you paste anything into a consumer AI tool, ask: would I email this exact text to a stranger I don't know?
If the answer is no, don't paste it. The AI tool is functionally a stranger you don't know, with terms of service you didn't read, sitting on a server you don't control.
Most things. Generic questions. Hypothetical scenarios. Drafts you wrote yourself with no customer or employee data. Public information. Marketing copy. Recipe ideas. Most code that doesn't contain credentials or proprietary algorithms.
The line is "data that belongs to someone other than you, or that you've promised to protect, or that creates a legal exposure if it leaks." Everything else, paste away.
The blunter version: if you wouldn't put it on a post-it note stuck to a bus stop wall, you probably shouldn't drop it in a chat with AI.
Most businesses don't have a written AI policy. The ones that do are rarely complicated. One paragraph in the employee handbook is enough for most small businesses:
We use AI tools at work. Here's what doesn't go into them: customer data, employee personal data, financial records, source code, anything covered by HIPAA or attorney-client privilege, and credentials of any kind. If in doubt, ask.
That paragraph costs nothing and takes ten minutes to write. It would have prevented every incident in this post.
If you want help writing a real AI policy for your business, that's part of what an AI Readiness Audit covers. Reach out and we'll talk.
We'll email you when each post is up. Two emails a month, max. Easy unsubscribe.
