Odds are, someone on your team has already pasted something into ChatGPT they shouldn’t have.
PagerDuty’s 2026 survey asked 1,250 office workers at large firms what they had shared with public AI tools. 34% said customer data. 31% said financial data or confidential documents. In the UK, 51% said emails and 23% said customer data.
(Those firms all turn over $500m or more. Don’t relax if yours is smaller. Within the survey, staff at firms under 1,500 people put customer data in more often: 40% against 27%.)
Nobody set out to leak anything. They had a job to finish and a chat box open.
So here’s the list. Print it. Pin it up. Seven things stay out of ChatGPT.
1. Customer data
Names, emails, addresses, order history, complaints. All of it counts as personal data.
The Dutch data regulator received a breach report from a telecoms firm after one employee fed a file of customer addresses into a chatbot. IBM’s 2025 breach study found personal data exposed in 65% of shadow AI incidents. (Shadow AI means tools your business never approved.)
Do this instead: swap real details for labels. Client A. Email 1. Address 1.
2. Staff and candidate records
CVs, pay, sick notes, disciplinary letters, grievances. Your people never signed up for this.
The same regulator logged a worse case. An employee at a medical practice entered patient records into a chatbot, against the rules. Its verdict: “a major violation of the privacy of the people involved.” Health data gets extra protection in law, and your staff sick notes count.
The question alone gives you away too. The UK’s National Cyber Security Centre offers its own example: a CEO asking “how best to lay off an employee”.
Do this instead: ask about the process with no names. “How do I run a fair capability meeting?” gets you what you need.
3. Passwords, API keys and logins
Group-IB found 101,134 infected devices with saved ChatGPT logins, traded on dark web markets over the year to May 2023. ChatGPT saved chat history by default. So one stolen login opened every old chat.
A key pasted in March still sits there in November.
Do this instead: type PLACEHOLDER where the key goes. Use a password manager. If a real one slipped in, change it today.
4. Contracts, NDAs and legal advice
In February 2026, a US judge ruled a fraud defendant’s chats with a public AI tool had no legal privilege. The FBI had seized about 31 documents of those chats. On law firm Orrick’s reading, feeding your lawyer’s advice into a public tool also waives privilege over the advice itself.
The tool in that case was Claude. The logic covers any public chatbot.
That’s US law, not UK law, so ask your solicitor. Either way, don’t volunteer to be the first UK test case. And check your NDAs. An NDA limits who sees the document, and a chatbot firm sits outside that list.
Do this instead: ask what a type of clause means in general. Send the real contract to your solicitor.
5. Unpublished financials and strategy
Margins, forecasts, price rises, the sale you haven’t announced. 26% of UK respondents admit to entering financial information or confidential documents.
The NCSC warned in 2023 about three things. The provider sees your queries. Stored queries get hacked, leaked or made public by accident. And a new owner might bring different privacy rules.
It also flagged what adds up across one login. One chat says little. Twenty chats tell the whole story.
Do this instead: build the model with dummy numbers. Keep the real ones in your spreadsheet.
6. Source code and trade secrets
Samsung learned this in public. Three weeks after lifting a ban, it reportedly found three incidents. One employee pasted source code from a faulty semiconductor database. Another wanted a fix for defective equipment. Samsung banned staff use of ChatGPT soon after.
IBM found intellectual property exposed in 40% of shadow AI incidents.
Do this instead: share the smallest snippet you need, with every name changed. Or give developers an approved tool on business terms.
7. Raw meeting notes and transcripts
This one catches everyone. 40% of those surveyed, and 38% in the UK, have shared meeting notes or summaries. It looks harmless. Samsung’s third incident was an employee asking for minutes of a confidential meeting.
One transcript holds the whole list. A client’s name. Someone’s pay. The price rise. Your throwaway line about a supplier.
Do this instead: use a notetaker your business approved. Failing that, clean the notes first.
“But we pay for ChatGPT”
Fair challenge. The account type matters, so read this list as the rule for every free or personal account.
On personal accounts, OpenAI’s default lets it use chats to improve its models unless you switch off “Improve the model for everyone”. Paying for Plus buys you more features, with the same default. OpenAI doesn’t train on ChatGPT Business or Enterprise data by default, and it supports a Data Processing Addendum.
So a business plan with the paperwork signed moves most of this list from “never” to “on your rules”. Two items stay put. Passwords stay out on every plan. Legal advice stays with your lawyer.
And a business plan fixes training. It doesn’t fix the law. The ICO says there’s no “AI exemption” to data protection law. If a paste counts as a breach and puts people at risk, you have 72 hours to tell the ICO. The Dutch regulator has already warned these pastes can be data breaches.
(I’m not a lawyer. Ask yours.)
Don’t ban it
Here’s the thing nobody says out loud. A ban doesn’t stop this. A ban hides it.
The University of Melbourne and KPMG surveyed over 48,000 people in 47 countries. Staff uploaded company information to public AI tools most often at firms with a ban: 67%, as reported here. That doesn’t prove bans cause it. Firms with more to protect might ban first. It does show a ban won’t save you.
What to do this week
Pick one approved tool, on business terms, with the data agreement signed.
Write one page. Put these seven items on it, plus the NCSC’s test: if this query went public, would we have a problem? If yes, it stays out.
Name one person to tell when something slips. No blame. The ICO clock starts when you find out, so you want to find out fast.
Switch off “Improve the model for everyone” on every personal account today, as a stopgap.
AI makes your people faster. Keep them using it.
Give them the tool, the rule and one person to tell. Then get back to work.