10 rules to avoid ending up on AIFail

From real cases and typical scenarios we distilled ten simple rules to use an AI without becoming the next anecdote.

1. Check every cited source

   Case law, studies, books, software packages, API functions: an AI confidently invents plausible references. Open them and check they exist.

2. Never paste secrets

   Passwords, API keys, a full device configuration, confidential code, customer data: whatever you paste leaves your company.

3. Read before you run

   A destructive command (rm -rf, DROP TABLE, iptables -F, terraform destroy, chmod -R 777) must be read word by word, variables included.

4. Test on a disposable environment

   Staging, virtual machine, snapshot: never straight into production, and not on Friday at 5:58 pm.

5. Give an agent minimum permissions

   Read-only by default, no production access, no admin keys, and a human approving irreversible actions.

6. Keep a tested backup

   Snapshot, a backup restored at least once, Git history, "reload in 5" on network devices: plan the way back.

7. Don't delegate the decision

   Health, law, finance, HR: the AI suggests, you are responsible. A confident answer is not a correct answer.

8. Protect your public chatbot

   Scope limits, adversarial testing, no contractual commitments, escalation to a human, conversation logging.

9. Students: cite, verify, understand

   Proofread your work, check every source, be able to explain what you hand in. AI detectors get it wrong, but your teacher will notice.

10. Ask for sources, then verify elsewhere

    When in doubt, ask the AI where the information comes from, then cross-check with an independent source.

Frequently asked questions