AI and Your Documents: Privacy Basics

The plain-English version

When you upload a file to an AI chat:

• It goes over the internet to the company running the AI (OpenAI, Anthropic, Google, Microsoft).
• It gets stored on their servers for some period — usually at least 30 days.
• On free tiers, your conversation may be used to train future versions of the model. On paid business tiers, usually not.
• In rare cases, a human at the company might review a conversation — usually for safety or abuse checks, occasionally for quality improvement.
• After enough time, logs get deleted — though "enough time" varies.

None of this is a scandal. It’s how hosted cloud services have always worked. What’s new is the sheer range of documents people are handing over — things they’d never email to a stranger — without really thinking about it. The rest of this page is about drawing a sensible line.

A useful analogy: two kinds of photocopiers

Imagine the AI as a photocopier at a rental office. You put your document on the glass, press the button, and out comes a summary. Great.

Now imagine there are two models of photocopier in the building. One quietly keeps a scanned copy of every document that passes through it, and the copies get used to improve future photocopier software. The other has a big sticker that says "your originals are not saved or used for anything else."

Both copiers do the same job. The first one is free. The second costs a little per use. Which one do you use for a love letter? For your taxes? For a random takeout menu?

That’s the choice you’re making, every time you upload a file. Match the tool to the document.

Free tier vs paid tier: what actually changes

The key word above is "usually." Policies shift. If you’re uploading anything that could cause real harm if it leaked, open the current privacy page for the tool you’re using and read it yourself that day. Don’t rely on a table on a webpage, including this one.

The traffic-light rule

A simple way to decide what’s okay to upload. We give every file a color in our heads before we drag it into a chat.

Red-zone documents and what to do about them

Red-zone files are the ones that cause real trouble. Let’s walk through the common ones.

HR and performance documents

An employee review mentions real names, real performance issues, real compensation numbers. Uploading that to a free AI chat puts that information on a third-party server, possibly in training data, possibly reviewable by humans at the AI company, and almost certainly outside your employer’s data retention rules. If you want AI help thinking about a review, rewrite the content without identifying details, or use a Team/Enterprise plan with a data processing agreement.

Legal contracts

Contracts are tempting to feed to AI — they’re long, dull, and full of details a summary would catch faster. The catch is they usually contain someone else’s name. Your landlord, your counterparty, your employer. They may also include a confidentiality clause that you’re about to violate by pasting the thing into ChatGPT. If privilege or confidentiality matters, this is Enterprise-plan territory or a local tool.

Medical records

Almost always regulated. In the US, HIPAA covers doctors and hospitals but not you as the patient — meaning you can technically paste your own chart into ChatGPT without breaking the law. But consumer AI chats are not HIPAA-compliant, so doing this frequently, especially with a family member’s records, is worth thinking twice about. The privacy-first conversion post covers this at more length.

Financial records with account numbers

The simple rule: if the document contains a bank account number, a Social Security number, or a credit card number, redact those before upload. Black them out in a PDF viewer, or retype the useful parts by hand and paste them in. An AI can still help you understand the document without seeing the sensitive identifiers.

Small habits that make a big difference

• Turn off chat history in free ChatGPT (Settings → Data Controls) if you don’t want conversations stored for training. You lose chat history, but you gain privacy.
• Redact before upload. Most PDF viewers let you black out text. Ten seconds of redaction beats a nervous hour later.
• Start a new chat per topic. That way a sensitive file from one conversation doesn’t linger in context for an unrelated question.
• Delete after you’re done. Deleting the conversation removes it from your history; whether the company’s servers also delete depends on their retention policy, but deleting is still better than not.
• Pay for a no-training plan when it matters. A Team or Enterprise subscription is the fastest way to get a contract saying your data won’t be used for training. For a one-person operation, Claude Pro and ChatGPT Plus both have sensible opt-out defaults.

A note on our service

Since you’re on our site, here’s how our document conversion handles files, in the same plain terms. We process files to extract text and turn them into markdown. Files are kept only as long as needed to return a result to you, typically under an hour, and then deleted. We don’t train any AI model on uploaded content, and we don’t share files with third parties except the cloud infrastructure that runs the conversion. For the deeper details, see our privacy policy, privacy-first markdown conversion, and the offline vs online post for when neither option feels right.

If you get only one thing from this page

It’s this: match the sensitivity of the document to the tier of the tool. A takeout menu goes anywhere. A tax return goes on a paid plan with no-training guarantees, or not at all. Nine out of ten privacy problems are people skipping that one judgment call.

One more lesson to go — where to go next — pointing you at the right deeper guide based on what you want to actually do with AI.