How to safely use ChatGPT with confidential documents
In this post
AI assistants have become genuinely good at document work: summarizing a contract, drafting a client letter, finding the risks in a lease. The catch is what you have to give them. For professionals who handle confidential material, pasting a document into ChatGPT or any other AI tool raises a real question: where does that information go, and who controls it now?
This guide covers the risk plainly, and a workflow that removes it.
What actually happens when you paste a document
When you put text into a cloud AI tool, it travels to servers you don't control. Depending on the service and your settings, it may be retained for some period, reviewed for safety purposes, or used in ways governed by terms you accepted quickly. Policies differ between tools and change over time, and consumer tiers usually offer fewer guarantees than enterprise contracts.
None of that makes AI tools bad. It makes them the wrong place for raw confidential data. A client's name, a patient's record number, an unannounced deal term: once pasted, that information has left your custody, and no deletion button on your side changes what already traveled.
For regulated professions the stakes are sharper. Confidentiality duties, privacy laws, and client agreements generally don't have an exception for "but the chatbot was helpful."
The wrong fixes
Abstinence. Banning AI tools outright protects the data and forfeits the productivity. In practice it also fails quietly: people use the tools anyway and stop telling anyone.
Trimming by hand. Manually deleting names before pasting sounds workable and isn't. Identifiers hide in dozens of formats, documents repeat them in variations, and one missed instance defeats the whole exercise. It also destroys the document's usefulness if you cut too much.
The right fix: give the AI structure, not identities
What an AI tool actually needs to summarize or analyze a document is its structure and logic, not the real identities inside it. That's the idea behind the Safe AI Workflow:
01Anonymize first. Run the document through Conseal and choose Anonymize. Every sensitive value becomes a consistent label: [PERSON_1] was admitted to [ORG_1] on [DATE_1]. The document stays fully readable, and the mapping between labels and real values stays on your machine.
02Let the AI work on the safe copy. Summaries, rewrites, analysis: the tool sees everything it needs and nothing it shouldn't.
03Stitch the real values back. Bring the AI's output into Conseal, and it restores the original values wherever the labels survived, entirely on your device.
The AI never sees a single real name, number, or identifier, and you still get its full output with your real data back in place. Full walkthrough: The Safe AI Workflow.
Habits that keep it safe
- Anonymize the whole document rather than fragments, so context stays intact
- Tell the AI to keep bracketed labels unchanged when rewriting
- Read the stitched result before it goes anywhere, and scan for leftover brackets
- Treat this as policy, not preference: if a document is confidential, the AI gets the anonymized copy, every time
Used this way, the choice between productivity and confidentiality stops being a choice.
- Introducing Conseal
- How to safely use ChatGPT with confidential documents
- That black box isn't redaction: 4 mistakes that expose sensitive data
- Locked or open? Which save to pick for your redacted file