10 Practical Principles for Responsible AI
A compact operating guide for using AI at work or building it into a product — focused on controls that still make sense as models change.
Ten principles you can actually use
Define the task and the boundary
Be explicit about what the AI is allowed to decide, recommend or execute — and what remains a human responsibility.
Ground important claims
Connect factual work to trusted documents, data or current sources. Verify the evidence, not just the wording of the answer.
Use the least privilege
Give agents only the data and tools required for the current task. Read-only is safer than write access; scoped access is safer than administrator access.
Put approvals before irreversible actions
Payments, deletion, publishing, external messages and permission changes should have an explicit approval boundary where appropriate.
Minimize sensitive data
Do not send or retain more personal, confidential or secret information than the task requires.
Test failures, not just demos
Evaluate edge cases, ambiguous inputs, adversarial content, prompt injection, missing data and tool errors.
Make uncertainty visible
Design outputs so users can distinguish sourced facts, inference, estimates and unknowns.
Keep useful logs
For automated workflows, record what the system received, which tools it used, what it changed and when approvals occurred.
Provide a way to recover
Prefer reversible actions, drafts, versioning and rollback. When something cannot be undone, raise the approval standard.
Re-test after change
New models, prompts, tools and data sources can alter behavior. Important workflows should be evaluated again after material changes.
For personal use
- Do not confuse confidence with truth.
- Check primary sources for consequential facts.
- Verify urgent requests through a second channel.
- Review AI-generated code and calculations before using them.
- Know what data the service is allowed to keep or use.
For teams and products
- Assign an owner for the system and its risks.
- Maintain a small set of representative evals.
- Document allowed tools and permission scopes.
- Have an incident path for harmful or incorrect behavior.
- Track model and prompt changes so regressions can be traced.
Before you automate
Ask three questions: What can it read? What can it change? What happens if it is wrong? Those answers should determine permissions, approvals and monitoring.