Privacy, data, and what NOT to paste.
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 covers what's actually safe to share with AI, what's actually risky (and why), and a simple workflow that lets you use AI on sensitive work without putting your company or your customers in the headlines.
The mental model
Treat AI prompts like emails you might one day have to read out in court.
Not because that's literally going to happen for most prompts. But because that mindset filters out the worst mistakes. Would you put this in an email to a stranger? Would you put it in a forum post you can't delete? If the answer is no, don't paste it into an AI either.
Your AI prompts can become training data, logs, or leak vectors — depending on the tool, the plan, and the configuration. Some AIs explicitly never use your data for training. Others do unless you opt out. Knowing the difference is part of using AI professionally.
What's safe vs. risky
Safe to share with most AIs
- Public information — anything you could publish on a blog or post on LinkedIn
- Anonymized work tasks — "draft a follow-up email" without names or company details
- Generic professional content — outlines, brainstorms, formatting, polish
- Your own writing for editing or feedback (as long as it doesn't contain anything from the lists below)
- Code with no proprietary logic or credentials — open patterns, common algorithms, common bugs
NEVER paste into a consumer AI
- Customer data — names, emails, account numbers, anything linking a real person to their relationship with you
- Health information (PHI) — diagnoses, treatments, anything HIPAA-touched
- Financial account details — credit card numbers, bank accounts, SSNs, routing numbers
- NDA-covered content — internal strategy decks, unannounced product details, M&A discussions
- Source code with proprietary IP — your company's secret-sauce algorithms, security implementations
- Credentials — passwords, API keys, tokens, certificates, internal URLs
- Legal documents under privilege — anything where attorney-client privilege would be broken by disclosure
- Other people's personal communications — emails you received from coworkers, customer DMs, etc.
The three tiers of AI privacy
Enterprise tier
ChatGPT Enterprise, Claude for Enterprise, Microsoft 365 Copilot, Gemini Workspace. Data stays in your tenant, never trains the model. Safest for sensitive work — if your company has it deployed.
Pro / Team tier
ChatGPT Plus, Claude Pro, Perplexity Pro. Generally not used for training (per current policies), but you're still trusting the vendor's storage and access controls. Fine for most professional work; avoid for the never-paste list.
Free tier
Free versions of ChatGPT, Gemini, Claude. Often used for training by default (you may need to opt out). Plus: lower account security, no admin controls. Avoid for any work-sensitive content.
The redaction workflow
The simple workflow for using AI on sensitive work
Most "sensitive" work isn't actually about the AI doing the task — it's about the data IN the task. Redact the data; keep the task. Three steps:
- Replace names with placeholders. Instead of "Sarah Chen at Acme Corp," write "[customer contact at customer company]." Same email, no sensitive data.
- Replace numbers and identifiers. Account numbers, dollar amounts you don't want logged, SSNs — replace with
[number]or$[amount]. - Substitute back after. AI generates the output with placeholders. You copy it into your real document and find/replace the placeholders with the actual data. Total added time: 30 seconds.
This is the workflow lawyers, doctors, and accountants use to leverage AI without violating their obligations. It's not paranoid — it's professional.
Three rules you can apply tomorrow
Rule 1: When in doubt, don't. If you're not sure whether something is OK to paste into an AI, that uncertainty itself is your answer. Find another path.
Rule 2: Check your company's AI policy. Most companies now have one. Many have approved AI tools (where data is contracted not to be used for training) vs. unapproved ones. Use the approved ones for work content.
Rule 3: Treat your prompt history like email history. Both can be subpoenaed. Both can be screenshot. Both create records. Behave accordingly.
Final challenge: audit your last 10 prompts
Open your AI of choice. Look at your last 10 conversations. For each one, ask:
- Did this prompt include any real names, accounts, or identifiers I shouldn't have included?
- Was this prompt sent to an AI tier appropriate for the sensitivity of the content?
- Would I be comfortable if a screenshot of this prompt ended up in a public document?
For any "no" answer — that's a habit to fix. Most people find at least one. The fix is easy once you notice the pattern.
What you can do now
- Know exactly what's safe vs. risky to paste into a consumer AI
- Recognize the three tiers of AI privacy (Enterprise, Pro, Free) and pick the right one
- Apply the redaction workflow for sensitive work in 30 seconds
- Check (and follow) your company's AI policy
- Treat your AI prompt history with the same seriousness as your email history