Safety & Judgment
This chapter isn't here to scare you. It's here so you can keep using AI with peace of mind. Getting AI banned at your company because of one careless moment would be a real shame.
The One Rule That Matters Most
If you wouldn't want a coworker, client, or journalist to see it, don't paste it into AI.
This rule covers 90% of situations. The details are below.
Data Tiers: What You Can Paste and What You Can't
Never paste
- ID numbers, passport numbers, bank account numbers, credit card numbers
- Passwords, API keys, access tokens
- Medical records, health checkup results
- Client lists and contact details
- Unreleased financial figures, price quotes, contract terms
- Anyone else's personal information (even a coworker's)
Check your company's policy first
- Internal company documents, presentations, meeting notes
- Product plans, features that haven't launched yet
- Code (especially code that contains business logic)
- HR-related information
Usually fine
- Public information and published content
- Your own study notes and drafts
- Examples with identifying details removed
- General knowledge questions
Anonymizing: What to Do When You Need to Paste Something
Replacing real details with placeholders usually makes it safe to use:
Original: John Smith (555-123-4567) complained on 3/15 that our ABC product at the downtown store...
Revised: Customer A recently complained that our Product X at one of our stores...
Tip: Before pasting, use your text editor's "Find and Replace" to swap company names, people's names, phone numbers, and dollar amounts for placeholders. Then paste.
Why Be Careful: Where Does Your Data Actually Go?
Different plans handle your data very differently:
| Plan type | How your data may be used | Recommendation |
|---|---|---|
| Free personal plan | Some services may use it to improve their models by default | Only use public or personal content |
| Paid personal plan | You can usually turn off training use in settings | Go into settings, check, and turn it off |
| Business / team plan | Contracts usually guarantee it won't be used for training | Follow your company's guidelines |
| Self-hosted or local model | Your data never leaves your machine | The choice for the most sensitive data |
Each service's policies change, and the location of the "turn off training" option moves around often too. Check the settings page of every service you use on a regular basis.
Fact-Checking: What AI Says Isn't Always True
For the full guide, see Hallucinations and Fact-Checking. Here's the short version:
Always check: numbers, dates, laws and regulations, cited sources, claims about people, product specs
The 30-second fact-check:
- What happens if this is wrong? If nothing much, just use it
- Pick out the specific claims (numbers, quotes)
- Ask the AI: "Which parts are you confident about? Which parts are guesses? Please cite your sources"
- Confirm the key points with a search engine or official website, and actually click the links
Copyright and Commercial Use
What you put in
Don't upload complete copyrighted works (entire books, full paywalled articles) for large-scale processing. Summarizing or analyzing short excerpts is usually considered fair use, but keep in mind that laws differ from country to country.
What comes out
- Countries treat copyright for AI-generated content differently. Don't assume you automatically own all the rights
- Check the service's license terms before using output commercially (especially for image generation)
- AI can produce content that's very similar to existing work, so do a reverse search before using it for anything important
Disclosure
More and more platforms, journals, and clients are asking people to disclose AI use. Being honest about which parts you used AI to help with works in your favor over the long run.
Rolling It Out to a Team: A Minimum Viable Usage Policy
If you're championing AI at your company, here's a skeleton you can edit and use right away:
[1. Approved tools]
- Company-purchased version: ___ (use this first)
- Allowed personal versions: ___ (non-company data only)
[2. Data red lines]
The following must never be entered into any AI tool:
- Client personal information and contact details
- Unreleased financials, price quotes, and contracts
- Employee personal information
- System passwords and keys
[3. Responsibility for output]
- All AI output is treated as a draft and must be checked by the person responsible before it goes out
- Numbers, regulations, and citations in external documents must be verified separately
- Anything involving legal, tax, financial, or medical matters must be reviewed by a qualified professional
[4. Disclosure]
- If client deliverables rely heavily on AI-generated content, tell the client upfront
[5. When in doubt]
- If you're not sure whether something can be pasted in, ask ___ (contact person) first
Three Common Myths
"It's just a little snippet, it should be fine"
The risk isn't about length, it's about what kind of content it is. A single line with a client's phone number is a personal data leak.
"The company has no rule against it, so it must be okay"
No rule usually means nobody has thought about it yet, not that it's allowed. Asking once protects you.
"The AI told me it won't remember my data"
AI gets things wrong about its own product policies, too. Go by the official documentation, not by asking the model.
Being safety-minded isn't about making you afraid to use AI. It's about letting you use it longer and more boldly. Once you've drawn your red lines clearly, you're free to go all in on everything inside them.