What not to paste into an AI.
Every prompt you send is data leaving your computer. Most of the time that's fine. Some of the time it gets people fired. This lesson is why it matters, what's safe, what's never safe, and a 30-second workflow that lets you use AI on sensitive work anyway.
We'll go deeper on the stakes first, then run a 10-question gym where you call "paste it, or not?" About 9 minutes.
01 · Why this is a bigger deal than it feels
A chat window feels private — just you and the bot. It isn't. Your words travel to a company's servers, and depending on the tool they can be stored, learned from, or exposed. And you can't un-send a prompt: once it leaves your machine, it's out of your control. That's the whole risk.
It's not paranoia — it's that three different things can happen to what you type:
It trains the model
On some tools and tiers, your prompt improves the model — fragments can resurface for other users.
It sits in logs
Prompts are retained on vendor servers. A breach or a legal request can expose them.
It becomes a record
Many employers log AI use. Prompt history can be screenshot, audited, or subpoenaed.
None of this means "don't use AI." It means know the line — which is simpler than it sounds.
02 · The one-line test
Treat every prompt like an email you might one day have to read out in court. You won't — but that mindset filters the worst mistakes. Would you put this in an email to a stranger, or a forum post you can't delete? If no, don't paste it into an AI.
Safe to share with most AIs
- Public info — anything you'd post on a blog or LinkedIn
- Anonymized tasks — "draft a follow-up" with no names or company details
- Generic content — outlines, brainstorms, formatting, polish
- Your own writing for editing (if it's clean of the items below)
- Code with no proprietary logic or credentials
Never paste into a consumer AI
- Customer data — names, emails, account numbers
- Health info (PHI) and financial details — SSNs, card/bank numbers
- NDA-covered content — internal strategy, unannounced products, M&A
- Proprietary source code and credentials — keys, tokens, passwords
- Privileged legal docs and other people's private messages
03 · Redact the data, keep the task
Most "sensitive" work isn't sensitive because of the task — it's the data inside it. Swap the data for placeholders, run the task, then substitute the real values back in your own document. Watch:
Same email, zero sensitive data. Find-and-replace the placeholders back afterward — about 30 seconds. This is exactly how lawyers, doctors, and accountants use AI without breaking their obligations.
Enterprise
ChatGPT/Claude Enterprise, M365 Copilot, Gemini Workspace. Data stays in your tenant, never trains the model. Safest.
Pro / Team
ChatGPT Plus, Claude Pro, Perplexity Pro. Generally not trained on — but still avoid the never-paste list.
Free
Free tiers often train on your data by default and have weaker account security. Avoid for any work-sensitive content.
04 · Paste it, or not?
Ten things land on your screen. For each, make the call: safe to paste into a consumer AI, or not?
Gym complete
Three rules you can apply tomorrow:
- When in doubt, don't. The uncertainty itself is your answer — find another path or redact.
- Check your company's AI policy. Use the approved tools for work content.
- Treat prompt history like email history. Both can be screenshot, logged, or subpoenaed.
Tell safe from never-paste on sight, redact sensitive data in 30 seconds, and pick the right privacy tier for the work.
Lesson 6 · Staying safe with AI — security & privacy essentials