Your team wants to use ChatGPT to summarize a client contract, draft a response to a legal memo, or get help with a code review. The question your compliance officer asks: is it safe? The honest answer: no, not without guardrails. Here's why, and how to fix it.
What actually happens when you paste into ChatGPT
When you type anything into the ChatGPT prompt box and hit send, that text is transmitted to OpenAI's servers. It's stored in logs. It may be used in training data (unless you've opted out in settings). It's no longer under your control. If that text includes a client name and SSN, an account number, a confidential strategy, or proprietary code, it has now left your organization and landed in a system you don't own. OpenAI's terms promise not to use it in training if you have an enterprise agreement, but the text is still in their logs. That's the exposure. For regulated industries — legal, accounting, healthcare, finance — this single paste can be a reportable disclosure. The client might have grounds to claim their data was mishandled. Your insurance might not cover it. Auditors will ask how it happened.
Why 'be careful' doesn't work
The instinct is to tell your team 'just don't paste sensitive data.' But that instruction fails in practice. Your team is under deadline pressure. They paste something, they catch themselves, but it's too late. Or they paste something they thought was fine, but it included an account number buried in a paragraph. Or they work on a personal device with a personal ChatGPT account, and nobody controls it. One careful employee handles it. But in a team of 50, statistically, someone will slip. The control needs to be technical, not behavioral.
Enterprise agreements don't eliminate the exposure
OpenAI offers enterprise agreements that promise your data won't be used in training. That's helpful. But the data is still in OpenAI's systems. If you're HIPAA-covered, your PHI shouldn't be there at all — it's a breach. If you're subject to GDPR, you're sending EU data to the US, which has legal implications. If you're in legal services, attorney-client privilege might be waived by disclosure. Enterprise agreements reduce risk, but they don't eliminate it. The real fix is preventing the data from being sent in the first place.
How to use ChatGPT safely: screen before send
The reliable solution is a layer that screens prompts before they're submitted. Emil's browser extension runs on ChatGPT and catches sensitive data as you type — SSNs (with or without dashes), account numbers, API keys, confidential content — and redacts it or warns you before the message is sent. The safe parts of your prompt still go through. 'Summarize this client contract' gets sent. The client's SSN doesn't. 'Review this code snippet' gets sent. Your API key doesn't. Your team keeps the speed of ChatGPT; you get the control and the evidence that sensitive data stayed in.
Prove it to auditors and counsel
When an auditor asks 'how do you control what leaves through AI tools,' you need an answer that isn't 'we tell people to be careful.' Emil logs every screening decision (metadata only, never the content), so you can run a report: 'In the last quarter, the extension screened 8,000 prompts, caught 247 instances of sensitive data, and redacted or blocked them before they reached the AI provider.' That's the evidence that transforms a handwavy policy into a measurable control.