.png)
How to Use This Page
If you completed the self-assessment in Before You Continue, you evaluated your own readiness as a reader. This page gives you the tool for what follows: a framework to evaluate individual AI outputs every time you use them.
Before You Paste, Trust, or Automate
Seven questions to run through before acting on any AI result.
The Risk Ladder
Use this ladder to decide how much review, verification, and control an AI output needs before it affects the real world.
Every AI output carries a risk level. The ladder maps that level to a rule. When stakes are unclear, go up one level.
- Use case examples
- Secrets, credentials, sensitive personal data, privileged company data, automated execution authority
- Why it matters
- Exposure cannot be undone; access cannot be easily revoked.
- Rule
- Do not paste or connect unless policy, access controls, and monitoring already exist.
- Use case examples
- Payments, trading, hiring/firing, production deployment, safety-impacting decisions
- Why it matters
- Errors cause direct, potentially irreversible harm.
- Rule
- No autonomous action without formal controls, logging, and documented human approval.
- Use case examples
- Legal, medical, financial, HR, security, compliance, or regulated content
- Why it matters
- Errors carry professional or regulatory consequence.
- Rule
- Verify against primary sources or qualified experts.
- Use case examples
- Internal memos, customer emails, summaries, routine analysis
- Why it matters
- Errors reach others or influence decisions.
- Rule
- Review before sending or acting.
- Use case examples
- Brainstorming ideas, rewriting drafts, generating content
- Why it matters
- Errors have low consequence and are easily caught.
- Rule
- Use freely, with normal judgment.
★ If you're unsure which risk level applies to a given AI output, assign it the next level up rather than down. For example, if you can't tell whether something is Medium or High risk, treat it as High. The bias is toward more caution, not less. The cost of unnecessary review is low; the cost of skipping review on something that needed it can be significant.
A note on the examples: Knight Capital and Wells Fargo were conventional automation, not AI. They’re included because the same operating principles apply once a consequential decision is delegated to software.