How to Stop Sensitive Data Leaking into ChatGPT
Every day, people paste emails, reports, contracts, and customer records into ChatGPT to get a summary, a rewrite, or a quick answer. Most of them don't notice they've also pasted someone's bank details, National Insurance number, or home address along with it.
This isn't a hypothetical risk. It happens in offices, at home, and in every sector from healthcare to finance to logistics. The text feels routine — a client email, an HR document, a support ticket — but buried inside is personal data that was never meant to leave your system.
Why This Matters More Than Most People Realise
When you paste text into ChatGPT, it is transmitted to OpenAI's servers. Under default settings for free and Plus accounts, that data can be used to improve OpenAI's models. Even if you trust OpenAI, you have no control over what happens to that data once it's sent.
From a compliance perspective, sending personal data to a third-party AI service without a data processing agreement in place is likely a GDPR breach. For businesses handling client or employee data, the consequences can be significant — fines, reputational damage, and notification obligations.
The problem isn't malice. It's habit. People paste text into AI tools the same way they copy-paste into a search bar — without thinking about what's in it. The fix has to be as frictionless as the mistake.
What Types of Data Get Leaked Most Often
Based on common document types, the data most frequently exposed includes:
- Email addresses — in forwarded threads, contact lists, support tickets
- Phone numbers — in customer records, invoices, HR documents
- National Insurance numbers — in payroll documents, employment records
- Bank account and sort code details — in invoices, expense claims
- NHS numbers — in healthcare-adjacent documents
- Passport and driving licence numbers — in onboarding documents, compliance files
- Named individuals — in meeting notes, performance reviews, complaints
- Postcodes and home addresses — in delivery records, customer data exports
None of these are unusual. They appear in the kinds of documents people want AI help with most.
How to Check Before You Send
The safest approach is a scan-before-send habit. Before pasting any document or email into an AI tool, run it through a local PII scanner first. The key requirement is that the scanner should not send your data anywhere — the whole point is to keep sensitive content off third-party servers.
Step 1: Paste your text into a local scanner
A local scanner runs in your browser and checks the text without transmitting it anywhere. Mutant Data Safety Layer does this — it scans for over 12 categories of sensitive data entirely client-side. Nothing leaves your machine.
Step 2: Review what it finds
The scanner highlights everything it detects and classifies the document (PUBLIC / INTERNAL / CONFIDENTIAL). You can see exactly what would have been sent.
Step 3: Redact and copy the clean version
Sensitive items are replaced with typed placeholders like [EMAIL_REDACTED] or [NI_NUMBER_REDACTED] — or with semantic descriptions like [EMAIL_REDACTED: sender contact address] in semantic mode. You then copy the redacted version into ChatGPT instead.
What About ChatGPT's Own Privacy Settings?
OpenAI does offer an option to disable training on your conversations (Settings → Data Controls → Improve the model for everyone). Turning this off reduces — but does not eliminate — the risk. Your data is still transmitted to OpenAI's servers and subject to their privacy policy. Enterprise and Team plans offer stronger contractual protections.
For genuinely sensitive data, turning off training is a necessary step but not a sufficient one. The safest approach is to not send the raw data at all.
Building It Into Your Workflow
The habit that works best is treating the scanner like a spell checker — something you run before sending, not after. For teams, this means making a local PII scanner available and training staff on which document types require a check before AI use.
For individual use, the friction needs to be low. A browser-based tool that requires no login, no account, and no upload to an external server removes every excuse not to use it.
Frequently Asked Questions
Does ChatGPT store the data I paste into it?
By default, OpenAI may use conversations to improve their models unless you opt out in settings. Any sensitive data you paste is transmitted to OpenAI's servers and potentially stored. Enterprise plans offer stronger data protections, but for free and Plus users, treating ChatGPT as a public endpoint is the safest approach.
What types of sensitive data are most commonly leaked into AI tools?
The most common accidental exposures include email addresses, phone numbers, National Insurance numbers, bank account and sort code details, NHS numbers, passport numbers, postcodes, and named individuals. These often appear in documents, reports, and emails that people paste for summarisation or editing.
Is there a free tool to scan text for sensitive data before sending to AI?
Mutant Data Safety Layer runs entirely in your browser with no data sent to any server. It scans pasted text and redacts sensitive content before you copy it into ChatGPT or any other AI tool. Three free scans are available without sign-up.
Is this a GDPR issue?
Sending personal data to a third-party AI service without a data processing agreement is likely a GDPR breach for businesses operating in the UK or EU. ChatGPT's enterprise tier includes a DPA; free and Plus tiers do not. When in doubt, redact before you send.