Is It Safe to Upload Documents to ChatGPT? Privacy Settings and What Not to Share
Short answer: Uploading ordinary documents to ChatGPT is reasonably safe once you adjust settings, but personal and client data needs care. On personal plans, chats can be used to improve models unless you turn that off. Switch off training, use temporary chats for one-off tasks, and never upload Aadhaar, PAN, bank details or client records unredacted.
The real question is not 'is ChatGPT safe' in the abstract, but what happens to your file, who could see it, and how long it stays. The answers depend on your plan and your settings, and most people never open the settings page.
This guide covers individual settings in ChatGPT, Gemini and Claude as of November 2026, a red/amber/green list for Indian data types, and a redaction routine. Company AI policies are a separate topic; if your employer has one, it overrides anything here.
What happens to a file after you upload it to ChatGPT?
The file is sent to OpenAI's servers, processed to answer your question, and stored with that conversation until you delete the chat. It is not published or visible to other users.
Three things affect privacy from there. First, whether your conversations can be used to train future models (a setting, on by default for personal accounts). Second, retention: deleted chats are generally removed from OpenAI's systems after a period, OpenAI has said around 30 days, though legal obligations can extend that. Third, human review: a small sample of conversations may be reviewed for safety and quality.
Business and enterprise workspaces and the developer API work differently: OpenAI states that their data is not used for training by default. If your company provides one of those, use it for work files instead of a personal account.
Does ChatGPT train on my data?
On Free, Go, Plus and Pro personal plans, it can, unless you switch it off. The setting is under Settings, then Data controls, labelled along the lines of 'Improve the model for everyone'. Turning it off applies to new conversations going forward.
Switching off training does not stop ChatGPT from saving your chat history, and it does not stop memory. Those are separate controls, which is the most common confusion we see when people set this up for the first time.
A simple way to remember it: training decides whether your words can shape future models; history decides whether you can see the chat later; memory decides whether ChatGPT carries facts about you into new chats. Set each one deliberately. A practical default for most working professionals is training off, history on (so you can find past work), and memory on but reviewed every month or so.
Privacy settings walkthrough: ChatGPT, Gemini and Claude
Labels below are as of November 2026 and do get renamed. If you cannot find one, search the tool's help centre for the phrase in the first column.
| Control | ChatGPT | Gemini | Claude |
|---|---|---|---|
| Stop training on your chats | Settings > Data controls > improve-the-model toggle off | Turn off Gemini Apps Activity (shown in some regions as Keep Activity) | Settings > Privacy > toggle that lets chats help improve Claude off |
| Chat without saving history | Temporary Chat | Temporary chat option where available | Incognito chat |
| What it remembers about you | Settings > Personalization > Memory: view, delete or turn off | Saved info or personal context settings | Memory settings where available |
| Delete past chats | Delete individually or all from settings | Delete activity from Gemini Apps Activity page | Delete from the chat list or settings |
| Human review note | Samples may be reviewed for safety | Google says reviewers may read some chats; do not enter what you would not want reviewed | Flagged chats may be reviewed for safety |
Two practical points. Gemini's help pages have said that even with activity off, chats are kept briefly to run the service. And with Claude, Anthropic asked consumer users in 2025 to choose whether their chats could be used for training, with retention linked to that choice, so check what you selected. For what memory stores and how to manage it, see our guide to ChatGPT custom instructions and memory; for Claude's interface generally, our Claude AI how-to walks through it.
What should you never share with AI? A red, amber, green list
Use this list for personal accounts on any AI chatbot, with training switched off. With training on, move everything amber into red.
| Rating | Data type | What to do |
|---|---|---|
| Red: never | Full Aadhaar number, PAN with name and date of birth, passport scans | Never upload. If a document contains them, mask or crop first |
| Red: never | Bank account numbers, card numbers, UPI PINs, OTPs, net-banking passwords | Never share in any form |
| Red: never | Client files: GST returns with GSTIN, payroll sheets, medical records, legal case papers | Use only on an employer-approved business plan, or redact fully |
| Amber: redact first | Your own salary slip, rent agreement, insurance policy, CV with phone and address | Remove numbers and addresses, keep the parts you need help with |
| Amber: redact first | Internal business data: sales figures, pricing sheets, unreleased plans | Replace names with labels; round figures if exact values are not needed |
| Green: fine | Public information, published articles, your own drafts, generic templates, anonymised data | Upload freely |
How to redact a document before uploading: a worked example
Say you work at a small CA firm in Janakpuri and want ChatGPT to explain why a client's GST input tax credit looks lower this quarter. Here is the routine.
- Copy only the rows you need into a fresh sheet. You rarely need the whole file.
- Replace identifiers: the client's name becomes 'Client A', the GSTIN becomes 'GSTIN-1', supplier names become 'Supplier 1, 2, 3'.
- Delete PAN, addresses, phone numbers, bank details and invoice numbers.
- Keep what the question depends on: month, invoice value, tax amount, ITC claimed, ITC reflected.
- Turn off training or use a temporary chat, then upload and ask: 'Here is quarterly ITC data for an anonymised client. Explain likely reasons claimed ITC is higher than ITC reflected, and list what to check with suppliers.'
- Delete the chat when done if you used a saved chat.
Before: 'Sharma Traders, GSTIN 07ABCDE1234F1Z5, Supplier: Gupta Steel, Invoice 4471, ₹1,18,000.' After: 'Client A, GSTIN-1, Supplier 2, Invoice row 7, ₹1,18,000.' The analysis is just as good, and nothing identifies anyone. The answer still needs checking against the GST rules, as with any AI output; our guide to verifying AI answers explains how.
Are some file types riskier than others?
The risk is in what a file contains, not its format, but some formats hide more than you expect.
- Spreadsheets often carry hidden sheets, filtered rows and old columns with phone numbers or PAN. Copy only the visible range you need into a new file.
- PDFs from banks and government portals usually print your name, address and account numbers in the header of every page, not just the first.
- Screenshots capture notification banners, contact names and open tabs. Crop before uploading.
- Photos can include location data in their metadata, and faces of people who did not agree to be uploaded.
- Word documents may keep tracked changes and comments with colleagues' names. Accept changes and delete comments first.
Where does India's DPDP Act fit in?
The Digital Personal Data Protection Act, 2023 governs how organisations in India handle personal data, and its Rules were notified in late 2025 with obligations phasing in over the following period. Read the Act and Rules on the MeitY website rather than relying on summaries.
For individuals using AI tools, the practical point is this: if you handle other people's personal data at work (clients, patients, students, employees), pasting it into a personal AI account may conflict with your organisation's obligations. This is not legal advice. When in doubt, anonymise, or ask whoever handles compliance in your organisation.
A five-minute privacy routine for every AI tool you use
- Open the data or privacy settings and decide on training.
- Review and prune what memory has saved about you.
- Use temporary or incognito chats for anything personal.
- Redact before uploading, using the list above.
- Delete old chats with sensitive content every few months.
The same principles carry over to every assistant, which is why the prompting fundamentals in our pillar guide include deciding what context to leave out. Safe, practical use of AI at work is built into every module of our Applied AI Professional course.
Frequently asked questions
Can OpenAI employees read my ChatGPT chats?
OpenAI says a limited number of authorised staff and contractors may review some conversations, for example for safety or abuse investigations. Treat any chat as something that could be read, and keep sensitive identifiers out.
Does deleting a ChatGPT chat delete it permanently?
It removes it from your history immediately, and OpenAI has said deleted chats are generally removed from its systems after a period of around 30 days, unless legal obligations require keeping them longer.
Is ChatGPT Plus more private than the free plan?
Not by default. Personal paid plans have the same training setting as free accounts. Business and enterprise workspaces are the plans that exclude your data from training by default.
Is it safe to upload my bank statement to ChatGPT?
Not unredacted. Remove the account number, name, address and UPI IDs, keep only the transaction columns you need analysed, and use a temporary chat.
Does the advice change for Gemini and Claude?
The principles are the same: switch off training, use temporary or incognito chats, and redact. The setting names differ, as shown in the walkthrough table.
Keep reading in AI tools
- How to Write AI Prompts That Work First Time Start here
- How to Fact-Check AI Answers Before You Use Them
- How to Make AI Videos From Text and Images
- ChatGPT Custom Instructions and Memory: Set Them Up Once
- How to Use Deep Research in ChatGPT and Gemini: Brief It, Then Audit It
Related: Applied AI Professional · How to Fact-Check AI Answers Before You Use Them · All guides in AI tools