What you'll learn
Prompt engineering hands you real power to shape what an AI produces — and real responsibility for the consequences. This module is about using it well: recognizing bias, protecting privacy, being honest with users, and keeping humans in charge of decisions that matter.
By the end of this module you'll be able to:
- Recognize where bias enters model outputs and reduce it
- Handle personal and sensitive data responsibly
- Be transparent with users about AI's role and limits
- Design a human check into consequential decisions
Bias in, bias out
Models learn from vast human-written text, so they absorb its biases — around gender, race, age, culture, and more. Those can surface in outputs: skewed assumptions, stereotyped language, uneven quality across groups. Your prompt can amplify bias ("describe a typical nurse") or help counter it. It's your job to notice and mitigate.
- Ask neutrally. Avoid assumptions baked into the prompt; specify diversity where relevant.
- Test across groups. Run your evals (Module 26) on varied names, dialects, and demographics.
- Constrain outputs away from stereotyping, and review for it explicitly.
Watch out
Privacy & what you send
Every prompt you send leaves your control. Treat the model like any third-party service: don't send personal, confidential, or regulated data unless you know how the provider handles it and you have the right to share it.
| Practice | Why it matters |
|---|---|
| Minimize data sent | Include only what the task needs — less to leak or misuse |
| Redact / anonymize | Strip names, IDs, and secrets before they reach the model |
| Know the provider's policy | Whether inputs are logged, retained, or used for training |
| Never put secrets in prompts | API keys and passwords in a prompt are a breach waiting to happen |
Being transparent with users
People deserve to know when they're interacting with an AI and how far to trust it. Disclose that it's an AI, be honest about its limits (it can be confidently wrong — Module 18), and don't design prompts that deceive or manipulate. A model that cites its sources and admits uncertainty earns appropriate trust rather than blind trust.
Keeping a human in the loop
For decisions with real stakes — medical, legal, financial, safety, anything hard to reverse — the model should assist, not decide. Route consequential outputs to a human for review, and design the prompt to support that: ask for reasoning and sources the reviewer can check, and have it flag low-confidence or out-of-scope cases instead of pushing ahead.
Key idea
A responsible-prompting checklist
Before shipping a prompt or an AI feature, run through these:
- ✓Could this output reflect or amplify bias? Test across groups.
- ✓Am I sending personal or sensitive data I shouldn't?
- ✓Does the user know they're talking to an AI?
- ✓Is a human reviewing consequential decisions?
- ✓Have I told the model what NOT to do (harmful, medical, legal)?
- ✓Can I stand behind how this system is used?
Recap & quick check
Key takeaways
- Models absorb bias from training data; your prompt can amplify or reduce it, so ask neutrally and test across groups.
- Every prompt leaves your control — minimize and redact sensitive data, know the provider's policy, and never send secrets.
- Be transparent: tell users it's an AI, be honest about its limits, and don't design deceptive prompts.
- Keep a human in the loop for consequential decisions, and prompt for reasoning and sources they can check.
- Match oversight to stakes — the higher the impact on people, the tighter the human review.
Quick check
1. Where does bias in model outputs mainly come from?
2. What's the right stance on sending sensitive data in prompts?
3. What does transparency with users require?
4. When is a human-in-the-loop most important?
You can now build prompts and systems that are effective, reliable, secure, and responsible. The final phase is about putting it all to work — reusable patterns and playbooks for real tasks. Next up: Phase 8, Module 30 — Prompt Patterns & Recipes.