How to redact sensitive data before pasting into ChatGPT
To redact sensitive data before ChatGPT sees it, replace every name with a label, delete contact details and account numbers, and cut the email headers and signature. Then read the whole thing once more before you paste. Done properly, the question still goes through and ChatGPT can still help. What it gets is the problem, not the person.
This guide walks through a real email, line by line. Then it covers where doing this by hand breaks down.
What counts as sensitive
Start with the legal line. Under the UK GDPR, personal data is any information about a person who can be identified directly or indirectly. That includes a name, an ID number, location data or an online identifier. The ICO sets this out plainly.
Indirectly matters. A job title plus a small town can identify someone. So can a rare event.
Some data is more serious still. The UK GDPR calls it special category data. It covers racial or ethnic origin, political opinions, religious or philosophical beliefs, trade union membership, genetic data, biometric data used for identification, and data about health, sex life or sexual orientation. See the ICO guidance on special category data.
Then add business data the law does not list. Card numbers. Bank details. Passwords and secret keys. Prices agreed with a client. Anything under a confidentiality clause.
Why redaction still matters on a paid plan
Settings help. They do not replace redaction.
On Free and Plus personal accounts, OpenAI may use conversations to train its models unless the user opts out in Data Controls, according to Strac. On business plans, data is not used for training by default. That is confirmed on OpenAI's own page, Enterprise privacy at OpenAI. Our guide to ChatGPT Team vs Enterprise vs free goes through each plan.
Three catches remain.
- Not training is not the same as not storing. Conversations are still sent, processed and kept for a period for abuse monitoring.
- A thumbs up or thumbs down on an answer lets OpenAI use that chat for training, even with training switched off. NordVPN flags this.
- A Temporary Chat is not used for training and is not in your history. It is still deleted only after 30 days, per Danube Labs.
Rule of thumb: settings decide what happens to the data. Redaction decides whether the data is there at all. Do both.
Redact a real email, step by step
Here is an email from Ashcombe, a fictional firm. Priya wants ChatGPT to help draft a reply.
From: Gareth Lowe <gareth@fenwickjoinery.co.uk>
To: priya@ashcombe.co.uk
Subject: Invoice 4471 overdue
Hi Priya, sorry for the delay on invoice 4471 for £3,240. My wife has been in hospital and things slipped. Can we pay half now from our Barclays account, sort code 20-45-77, account 41558902, and the rest next month? Call me on 07700 900812.
Gareth Lowe, Director, Fenwick Joinery Ltd, 12 Mill Lane, Hexham
Step 1: Cut the headers
Delete the From, To and Subject lines. ChatGPT does not need them. They hold two email addresses and an invoice number.
Step 2: Swap names for labels
Replace each name with a consistent label. Gareth becomes [CLIENT_CONTACT]. Fenwick Joinery becomes [CLIENT_COMPANY]. Priya becomes [ME]. Use the same label every time the same person appears.
Step 3: Take out numbers that identify
Sort code, account number and phone number come out completely. Replace the invoice number with [INVOICE]. Keep the amount only if the answer depends on it. Here a round figure such as "about £3,000" is enough.
Step 4: Strip the signature block
Job title, company, street and town all go. Signatures are the part people most often forget.
Step 5: Check for special category data
"My wife has been in hospital" is health data about a third person. Rewrite it as "a family emergency". The tone of the reply is the same. The detail is gone.
The result
Hi [ME], sorry for the delay on [INVOICE] for about £3,000. There has been a family emergency and things slipped. Can we pay half now and the rest next month? [CLIENT_CONTACT], [CLIENT_COMPANY]
Ask ChatGPT for a kind but firm reply. Then put the real names back yourself, in your email program.
Final check: read the redacted text as a stranger. If you could work out who it is about, keep going.
Where manual redaction breaks down
The method works. The person doing it is the weak point.
- Time pressure. Five steps feel fine once. At the end of a busy day, people skip the check.
- Long threads. Forwarded emails hide older headers and signatures halfway down.
- Files. A spreadsheet may hold hundreds of rows. Nobody reads them all.
- Inconsistent labels. Gareth becomes [CLIENT] in one place and "G" in another. The answer gets muddled, or a real name slips through.
- Indirect clues. A rare job and a small town can identify someone with no name in sight.
A common pattern is that redaction is done carefully by the people who already care, and not at all by everyone else. Sector rules make that gap matter more. See our pieces for accountants using ChatGPT with client data and for solicitors.
Do you need automatic checks?
Manual redaction is probably enough if one or two careful people use AI now and then, with short text.
Automatic checks are worth a look if several staff use AI daily, if files get uploaded, or if clients, insurers or auditors ask how you control it. Free browser extensions can help. We compare them in Helio Guard versus free redaction extensions, and say where each is the better pick.
Helio Guard does the steps above on each Windows PC. It checks pastes, messages and uploads on AI sites. Sensitive items are taken out and swapped for labels, and client names become stand-ins such as "Client A". Checked on the PC, with fixed rules. We never see what your team types. It is £1 a person a month, from 5 people, with a 14-day free trial and no card needed.
Common questions about redacting data for ChatGPT
Is using initials instead of names enough to redact data?
Initials are rarely enough to redact data properly before using ChatGPT. Combined with a company name, a town or an unusual event, initials can still identify a person. Use neutral labels such as [CLIENT_CONTACT] instead, and remove the surrounding details that would let someone work out who the text is about.
Does Temporary Chat in ChatGPT remove the need to redact?
Temporary Chat does not remove the need to redact. A Temporary Chat is not used for training and does not appear in history, but it is still sent to OpenAI and kept for up to 30 days. Sensitive client data should still be taken out before it is pasted in.
Can I paste anonymised data into ChatGPT under UK GDPR?
Truly anonymised data falls outside UK GDPR, because nobody can be identified from it. The test is strict, though. If a name, number, location or combination of details could still point to a person, directly or indirectly, the text is still personal data and the normal rules apply.
Related: What cyber insurers ask about AI use, and how to answer honestly, Does ChatGPT or Copilot train on your business data?, Is pasting a client email into ChatGPT a data breach under UK GDPR?