Safe means three answers you can read, not a feeling
A document feels different from a typed question because it is the whole thing at once: the contract, the payslip, the board pack. You're right that more is at stake. But instinct isn't how you decide. Each vendor answers three questions in its policy, and a document is safe to upload where the three answers are acceptable to you.
Does the service train on it? Consumer tiers of ChatGPT and Claude use conversations for training by default with an opt out; the AI chatbot privacy comparison tracks each provider's default and the dates they changed. Whizi's policy is one sentence: "Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models. Whizi does not sell this content as training data."
How long is it kept? Whizi configures uploaded attachments to expire after up to 30 days, and you can delete a file or the whole conversation sooner. OpenAI's API does not train on what you send unless you opt in, and keeps abuse monitoring logs for up to 30 days by default. A consumer chat product keeps an uploaded file as long as the conversation exists unless you delete it.
Who else receives it? Every chat product in front of a model sends the document to the model's maker. In Whizi a file attached to a Claude message reaches Anthropic and one attached to a GPT message reaches OpenAI, through OpenRouter, and nothing is sent to a model you did not select. Whizi routes API traffic with OpenRouter's data collection flag set so the request is excluded from provider training, and providers may still hold temporary caches under their own policies. Whizi doesn't claim a universal zero retention setting, and neither should anyone else.
How much of the document the model actually sees
One fact changes the risk picture more than any policy line: the model never reads a large document whole. Whizi gives every catalogue model a flat budget of 40,000 input tokens per turn, roughly 80 pages of ordinary text. An upload that runs past it is trimmed at the cut, with a toast and an in-message marker saying so. That's a quality limit, and it is also a privacy one, because the part of a 300 page report that leaves your machine is the part that fits.
If the sensitive part of a document is on page 200, uploading the whole file may never send it, and uploading pages 195 to 205 certainly will. And if the document is a scan, Whizi runs it once through a server side vision pass to produce a per page transcript. A model reads the pages as images before the text exists; the chat with PDF guide explains when that fires. Splitting a document by purpose and uploading the sections you need, which the upload documents guide recommends for accuracy, is also the smaller privacy footprint.
The documents that fail the test at every vendor
Some files should not go into any AI product, however good its policy, because the risk is not the vendor's handling but your permission to share at all.
| Document | Why it fails everywhere | What to do instead |
|---|---|---|
| Anything under someone else's NDA or a client data processing agreement | Your contract, not the vendor's policy, governs it, and most forbid sending to a third party processor | Redact, or ask the client, or use a tool inside the contract |
| Files holding other people's personal data: patient lists, payroll, customer records | You need a lawful basis to send it anywhere, and "the policy says no training" is not one | Strip names and identifiers first; the analysis rarely needs them |
| Credentials: API keys, passwords, tokens, recovery codes | A file with a key in it is a leaked key the moment it leaves your machine | Never; rotate any key that has already been uploaded |
| Unreleased source code or product plans under an employer policy | The exposure is to your employer's rules, not the vendor's | Check the approved provider list; a multi model workspace lets you route to an approved one |
| Signed contracts, tax returns and bank statements that are yours | These can be safe where the three answers hold; the risk is the recipient list | Upload the pages you need, delete the file after |
The full list of what never belongs in a chat window, with the reasoning, is what never to paste into an AI chat. The rule that covers the table in one line: if you could not email the document to a subcontractor, you can't upload it either.
The vendor answers, side by side
| Vendor and tier | Trains on uploads by default | Stated retention for uploads | Who receives the file |
|---|---|---|---|
| ChatGPT, consumer tiers | Yes unless you turn off the training setting | Kept with the conversation until deleted; deleted chats removed within 30 days | OpenAI |
| Claude, consumer tiers | Yes since the policy change announced 28 August 2025, with an opt out | Up to 5 years for accounts that allow training | Anthropic |
| OpenAI API | No, unless you opt in | Abuse monitoring logs kept up to 30 days by default | OpenAI |
| Whizi | No, on any plan | Configured to expire after up to 30 days; deletable sooner | OpenRouter and the maker of the model you selected, excluded from provider training by routing flag |
The two consumer rows are the ones to read twice, because they are where most people upload most documents. Both vendors let you switch training off, and the privacy comparison shows where the toggle is on each. Neither toggle shortens retention on its own: at OpenAI the file lives as long as the chat does, and at Anthropic the training choice decides whether the five year window applies to you.
The Whizi row is our own and we have kept it to what the privacy policy states. Deleting a file does not remove what was already discussed about it in the conversation, so a document sensitive enough to delete is a conversation worth deleting too. Two separate 30 day figures sit in the policy: uploads are configured to expire after up to 30 days, and deleted database records can remain in Cloudflare D1's point in time recovery history for up to 30 days before ageing out. Neither is a promise of immediate erasure, and the policy states no exact expiry value. Where Whizi stores data has the full list of processors.
A four step habit for uploading anything that matters
- Check the three answers for the product you are using: training, retention, recipient. If any of them is not written down, treat it as the worst case.
- Strip what the task does not need. Names, account numbers, addresses and identifiers rarely change the analysis, and removing them takes a minute.
- Upload the pages, not the file. The sections you need are a sharper answer and a smaller exposure, and past 40,000 tokens the rest was never going to be read anyway.
- Delete the file and the conversation when the task is done, and rotate any credential that slipped through.
The decision rule: a document is safe to upload where the vendor answers all three questions in writing and your own contracts allow it, and it is not safe anywhere those two things are not both true.
- Read the vendor's training, retention and recipient answers before the first upload
- Never upload files under someone else's NDA, other people's personal data, or credentials
- Strip identifiers the analysis does not need
- Upload the relevant pages rather than the whole document
- Delete the file and the conversation when you are finished, and rotate any exposed key
Frequently asked questions
Is it safe to upload confidential documents to ChatGPT?
Only after two checks. Consumer ChatGPT uses conversations for training by default unless you turn the setting off, and keeps an uploaded file with the conversation until you delete it. And your own contracts govern whether you may send the document to a third party at all; a client NDA usually says no regardless of OpenAI's policy.
Does AI keep my uploaded files?
It depends on the product. Consumer chat products keep a file as long as the conversation exists. The OpenAI API keeps abuse monitoring logs for up to 30 days by default. Whizi configures uploaded attachments to expire after up to 30 days and lets you delete a file or conversation sooner, and deleted database records can remain in Cloudflare D1 recovery history for up to 30 days.
Does Whizi train on documents I upload?
No. Whizi's privacy policy states that it does not use prompts, files, conversations, voice transcripts or generated content to train Whizi-owned AI models and does not sell that content as training data. The file is sent, through OpenRouter, only to the provider of the model you selected, with the routing flag that excludes the request from provider training.
Does the AI read the whole document I upload?
Not if it is long. In Whizi every catalogue model gets a flat budget of 40,000 input tokens per turn, roughly 80 pages of ordinary text, and an upload beyond it is trimmed with a toast and an in-message marker at the cut. A 60 page deck, roughly 30,000 tokens, fits; a 300 page report sends less than a third of itself. Upload the sections you need for both a better answer and a smaller exposure.
What should I never upload to an AI tool?
Anything under someone else's NDA or data processing agreement, files containing other people's personal data such as patient or payroll records, and credentials of any kind. Those fail at every vendor because the problem is your permission to share, not the vendor's handling. Rotate any key that has already been uploaded.
