"The real question is not whether machines think but whether people do." — B. F. Skinner, psychologist

Chapter Overview

Priya Sharma stares at a blank document at 11:47 p.m. on a Tuesday. She has a policy brief due in 36 hours, a statistics problem set due in 24, and she just picked up an extra shift at the campus bookstore because rent does not negotiate. She opens a chatbot and types: "Write me a policy brief about renewable energy incentives."

The chatbot produces something. It is coherent. It sounds authoritative. It even includes what look like citations — except when Priya checks two of them, neither exists. The third is real but says the opposite of what the chatbot claimed. The policy brief is fluent, confident, and substantially wrong.

Priya deletes the whole thing and starts over. This time she tries something different. She types: "I am a second-year political science student writing a policy brief about renewable energy tax incentives in the United States. My thesis is that the Investment Tax Credit has been more effective than production-based incentives for residential solar adoption. Can you help me outline the main arguments and counterarguments? Do not include any citations — I will find my own sources."

The result is dramatically better. Not perfect — she still needs to verify claims, reshape the structure, and write the actual brief herself — but the outline gives her a useful scaffolding. The difference between the two attempts was not the AI. It was the prompt.

This chapter is about the space between those two moments. It is a practical guide to working with AI tools — not blindly trusting them, not reflexively refusing them, but developing the judgment to use them effectively, ethically, and with your eyes open. Whether you are a student, a professional, or someone who just wants to get more out of the AI tools you already use, the skills in this chapter will serve you immediately and durably.

In this chapter you will learn to:

1. Write effective prompts for large language models using established techniques
2. Evaluate AI-generated outputs for accuracy, bias, and completeness
3. Identify appropriate and inappropriate uses of AI tools in academic and professional contexts
4. Develop a personal AI use policy that balances efficiency with integrity
5. Collaborate productively with AI systems while maintaining critical judgment

Learning Paths

Fast Track (50 minutes): Read sections 14.1, 14.3, and 14.6. Complete the Self-Assessment and Project Checkpoint.

Deep Dive (2.5–3 hours): Read all sections, complete the Check Your Understanding prompts and Scenario Walkthroughs, build your Personal AI Policy using the template, read both case studies, and work through the exercises.

Spaced Review — Concepts from Earlier Chapters

Before we dive in, let us reconnect with three ideas from earlier in the book that are essential to this chapter.

🔁 From Chapter 5 (Large Language Models): LLMs generate text by predicting the next token in a sequence. They do not "understand" your question or "know" the answer — they produce the most statistically probable continuation of the text you gave them. This means the way you frame your input has an enormous effect on the quality of the output.

🔁 From Chapter 8 (When AI Gets It Wrong): AI confidence and AI correctness are different things. A chatbot can state a completely fabricated fact with the same fluent certainty it uses for accurate information. This is why verification is not optional — it is the core skill of effective AI use.

🔁 From Chapter 9 (Bias and Fairness): AI systems reflect the biases in their training data. When you use an AI tool to research a topic, draft text, or analyze information, the output may carry systematic blind spots — particularly around marginalized communities, non-English-language contexts, and recent events.

14.1 The Art and Science of Prompting

Let us start with a concept that would have seemed bizarre just a few years ago: the quality of what you get out of an AI system depends enormously on what you put in. This is not a minor technical detail. It is the single most important practical skill for anyone using generative AI tools today.

A prompt is the input you give to an AI system — the text you type into a chatbot, the instructions you provide to an image generator, the query you pose to a code assistant. Prompt engineering is the practice of designing prompts to get better, more useful, more accurate outputs.

Here is why prompting matters so much. Remember from Chapter 5 that large language models work by predicting the most likely next token given all the tokens that came before. Your prompt is not a question that the model "answers" — it is the beginning of a text that the model continues. When you type "Write me an essay about climate change," the model is essentially asking itself: "What kind of text would most likely follow this beginning?" The answer is: generic, surface-level, vaguely authoritative text about climate change — because that is what most resembles the pattern of text that followed similar prompts in its training data.

But when you type something more specific — when you tell the model who you are, what you need, what format you want, what level of detail is appropriate, and what constraints apply — you are narrowing the space of probable continuations to a much more useful region.

Think of it this way. Imagine you walk into a vast library and say to the librarian, "Tell me about history." The librarian stares at you. Which history? What time period? What region? For what purpose? At what level — a kindergarten report or a doctoral dissertation? Now imagine instead you say, "I am an undergraduate writing a 10-page paper about the causes of the 2008 financial crisis. I need an overview of the three most widely cited contributing factors, with the strongest evidence for and against each one." The librarian knows exactly where to take you.

AI prompting works on the same principle, with one crucial difference: the librarian might ask you clarifying questions. The AI model, in most cases, will simply produce something — and if your prompt is vague, that something will be vague too.

💡 Key Insight: A prompt is not a question you ask an AI. It is a set of constraints that shape the space of possible outputs. The more precise your constraints, the more useful the output.

What Makes a Good Prompt?

While there is no single formula, effective prompts tend to share several characteristics:

1. Specificity about the task. Instead of "Write about renewable energy," try "Summarize the three main arguments for and against wind energy subsidies in the United States, in approximately 300 words, for a reader with no background in energy policy."

2. Context about the user. Tell the model who you are and why you need this. "I am a nursing student preparing for a pharmacology exam" produces different (and more useful) output than a generic query about drug interactions.

3. Format instructions. If you want a bulleted list, say so. If you want a comparison table, specify it. If you want the response to be under 200 words, state that constraint. Models are remarkably responsive to format instructions.

4. Constraints and guardrails. Tell the model what not to do. "Do not make up citations." "If you are uncertain, say so." "Do not use jargon without defining it." These negative instructions can be as valuable as positive ones.

5. An explicit audience. "Explain this to a 10-year-old" produces very different output from "Explain this to a physician." Specifying your audience helps the model calibrate its language, detail level, and assumptions.

🔄 Check Your Understanding: Look back at the two prompts Priya used at the beginning of this chapter. Identify which of the five characteristics above are present in her first prompt and which are present in her second. How does this explain the difference in output quality?

14.2 Prompt Engineering Techniques: Zero-Shot, Few-Shot, Chain-of-Thought

Now let us move from general principles to specific, named techniques that researchers and practitioners have found to consistently improve AI outputs. These are not secret hacks — they are well-documented approaches that work because of how language models process information.

Zero-Shot Prompting

Zero-shot prompting means giving the model a task without providing any examples of the desired output. You simply describe what you want.

Example: "Classify the following movie review as positive, negative, or neutral: 'The acting was superb but the plot made absolutely no sense. I left the theater confused but impressed.' "

This works for many straightforward tasks because large language models have seen similar tasks during training. The model recognizes the pattern — "classify this text" — and produces a reasonable response. Zero-shot prompting is your default starting point. Try it first, and if the results are not good enough, move to more structured techniques.

Few-Shot Prompting

Few-shot prompting means providing the model with a few examples of the input-output pattern you want before giving it the actual task. Think of it as showing the model what "good" looks like before asking it to produce something.

Example:

"Classify each movie review as positive, negative, or neutral.

Review: 'A masterpiece of storytelling. I cried three times.' Classification: Positive

Review: 'Terrible. The worst two hours of my life.' Classification: Negative

Review: 'It was fine. Nothing special, nothing terrible.' Classification: Neutral

Review: 'The acting was superb but the plot made absolutely no sense. I left the theater confused but impressed.' Classification: "

By providing three examples, you have shown the model exactly how you want the task performed — the format, the tone, the level of nuance. Few-shot prompting is particularly powerful when you want a very specific output format or when the task is unusual enough that zero-shot attempts produce inconsistent results.

💡 Intuition: Few-shot prompting works because of how language models predict the next token. By showing a pattern of (input, output) pairs, you are essentially telling the model: "The text I am writing follows this structure. Continue the pattern." The model does not "learn" from your examples in the way a student would — it recognizes the pattern and continues it.

Chain-of-Thought Prompting

Chain-of-thought (CoT) prompting asks the model to show its reasoning step by step before arriving at an answer. This technique, documented in research from Google Brain in 2022, can dramatically improve performance on tasks that require multi-step reasoning — math problems, logical puzzles, complex analysis.

Example without chain-of-thought:

"A store sells apples for $2 each and oranges for $3 each. If Maria buys 4 apples and some oranges and spends $20 total, how many oranges did she buy?"

The model might jump straight to an answer, sometimes getting it wrong.

Example with chain-of-thought:

"A store sells apples for $2 each and oranges for $3 each. If Maria buys 4 apples and some oranges and spends $20 total, how many oranges did she buy? Think through this step by step."

Adding "Think through this step by step" (or "Let us work through this reasoning") prompts the model to generate intermediate steps: "4 apples at $2 each = $8. Total spent is $20. $20 - $8 = $12 remaining for oranges. At $3 each, $12 / $3 = 4 oranges." The step-by-step process acts as a kind of scaffolding that reduces errors.

⚠️ Important Caveat: Chain-of-thought prompting makes the model show reasoning steps, but this does not mean the model is actually "reasoning" in the way humans do. The steps are still generated by next-token prediction. Sometimes the model produces steps that look logical but contain subtle errors. Always verify the reasoning, not just the final answer.

Role Prompting

A fourth technique worth knowing is role prompting — asking the model to adopt a specific persona or expertise.

Example: "You are an experienced data journalist who specializes in explaining statistical concepts to a general audience. Explain what a confidence interval is, using a real-world example."

Role prompting works because it activates patterns in the training data associated with how experts in a particular field communicate. It can improve the accuracy, tone, and specificity of responses. However, a role prompt does not give the model actual expertise — it gives it a style of communication that may or may not be accompanied by accurate content.

Combining Techniques

These techniques are not mutually exclusive. In practice, the most effective prompts often combine several approaches:

"You are a constitutional law professor preparing a study guide for undergraduate students [role]. I will give you three examples of how to summarize a Supreme Court case, and then I want you to summarize a fourth case in the same format [few-shot]. For each case, explain the key legal question, the court's reasoning, and the implications, step by step [chain-of-thought]. Keep each summary under 200 words [constraint]."

📊 Scenario Walkthrough: Building a Prompt

Imagine you are researching the environmental impact of cryptocurrency mining for a term paper. Start with this weak prompt: "Tell me about crypto mining and the environment."

Now improve it using what you have learned: 1. Add specificity: What exactly do you want to know? 2. Add context: Who are you and what is this for? 3. Add format: How should the information be structured? 4. Add constraints: What should the model avoid? 5. Choose a technique: Would role prompting, chain-of-thought, or few-shot help?

Try writing your improved prompt before reading on. There is no single correct answer — the goal is to practice the skill.

14.3 Evaluating AI Outputs: A Verification Framework

Here is the uncomfortable truth about using AI tools: the better you get at prompting, the more convincing the outputs become — and the more convincing the outputs, the more tempting it is to stop checking them. This is exactly the wrong instinct. Skilled AI users are not people who get such good outputs that verification becomes unnecessary. They are people who always verify, no matter how good the output looks.

Why? Because the fundamental nature of how language models work has not changed just because your prompt was excellent. The model is still predicting likely text, not retrieving verified facts. It can still hallucinate. It can still present outdated information. It can still reflect biases in its training data. A great prompt reduces the probability of these problems but never eliminates them.

The VIBE Check: A Verification Framework

We propose a simple, memorable framework for evaluating any AI-generated output. We call it the VIBE Check — not because it is casual, but because it gives you a systematic way to assess whether an output should be trusted.

Let us apply the VIBE Check to a real scenario. Suppose you ask a chatbot: "What are the health effects of intermittent fasting?"

The chatbot responds with a detailed, confident answer listing benefits (weight loss, improved insulin sensitivity, reduced inflammation) and citing "a 2019 study in the New England Journal of Medicine" and "research from Johns Hopkins University."

• V — Verifiable: Search for the specific study mentioned. Does a 2019 NEJM paper on intermittent fasting actually exist? (In this case, one does — a review article by Rafael de Cabo and Mark Mattson. But AI models frequently fabricate citations that sound plausible but do not exist.)
• I — Internally consistent: Does the response claim intermittent fasting is "safe for everyone" in one paragraph and then mention "people with diabetes should consult a doctor" in another? Internal contradictions are a red flag.
• B — Balanced: Does the response mention potential risks and limitations, or does it read like an advocacy piece? Research on intermittent fasting is genuinely mixed — a one-sided answer is incomplete.
• E — Evidence-backed: Are the claimed benefits described in general terms ("studies show...") or with specific, verifiable evidence? Vague appeals to "research" should increase your skepticism.

⚠️ Critical Rule: Never cite an AI-generated citation without verifying it yourself. AI models routinely generate citations that look real — correct journal names, plausible author names, reasonable publication years — but that do not actually exist. This is one of the most common and most dangerous forms of hallucination. If you include an unverified AI citation in academic or professional work, you are the one responsible for the error, not the AI.

Common Failure Patterns to Watch For

As you evaluate AI outputs, be alert to these recurring problems:

Confident fabrication. The model states something false with the same tone and fluency it uses for true statements. There is no "uncertainty signal" in the text — no hedging, no qualifiers. This is why you cannot use the model's tone as an indicator of accuracy.

Plausible but wrong. The output sounds like something that could be true — it uses the right vocabulary, the right sentence structures, the right level of specificity — but the actual content is incorrect. This is particularly dangerous because it takes domain knowledge to catch.

Outdated information. Models have a training data cutoff — they do not know about events after a certain date. If you ask about recent developments, the model may present outdated information as current or simply make up recent events to fill the gap.

Systematic blind spots. Due to biases in training data, models may consistently underrepresent certain perspectives, cultures, or communities. If you are researching a topic that involves marginalized groups, non-English-speaking communities, or the Global South, scrutinize the output for gaps in representation.

Hedging as a cover. Some models have been fine-tuned to include hedging language ("It is important to note that..." or "Some researchers argue..."). This can create the appearance of balance without the substance. Check whether the hedges actually lead to different conclusions or are just decorative.

🔄 Check Your Understanding: You ask a chatbot to explain the causes of the 2008 financial crisis. The response mentions "the landmark Glass-Steagall Act of 1933, which was repealed in 2001." What VIBE Check step would catch this error, and what is actually wrong?

(The Glass-Steagall Act was partially repealed by the Gramm-Leach-Bliley Act in 1999, not 2001. This would be caught by the V — Verifiable step. The error is the kind of "close but wrong" factual claim that models frequently produce.)

14.4 Academic Integrity and AI: Where's the Line?

Let us return to Priya. She has gotten better at prompting — much better. She knows how to write specific, contextualized prompts. She applies the VIBE Check to everything. She uses AI as a thinking partner, not a ghostwriter. But she is still wrestling with a fundamental question: Where exactly is the line between using AI as a tool and using it as a crutch?

This is not a question with a single, universal answer. It depends on the assignment, the course, the institution, and — most importantly — the purpose of the work. But we can build a framework for thinking about it clearly.

The Skill-Building Test

Here is a useful heuristic: ask yourself, "Is this assignment designed to help me build a skill, or to demonstrate a skill I have already built?"

If the purpose is skill-building — if the professor assigned a close reading of a poem so you would develop the ability to analyze literary language — then having AI do the close reading defeats the purpose. You might produce an A-quality analysis, but you have not built the skill. You have outsourced it.

If the purpose is demonstrating or applying a skill you have already developed — say, formatting a bibliography or structuring a presentation — then using AI to handle the mechanical aspects may be perfectly appropriate, freeing you to focus on the intellectual work.

The difficulty, of course, is that many assignments blend both purposes. And reasonable people disagree about where the line falls.

A Spectrum, Not a Binary

Rather than drawing a single line, it is more useful to think about a spectrum of AI use, from clearly acceptable to clearly problematic:

⚖️ The Transparency Principle: When in doubt, disclose. If you are unsure whether your use of AI crosses a line, the safest and most intellectually honest approach is to tell your professor or supervisor exactly what you did. "I used ChatGPT to help me brainstorm an outline, then I wrote the paper myself" is a statement that few instructors would find objectionable — and the act of making it shows that you are engaging with the ethical dimensions of your work.

What Professors Are Actually Worried About

It is worth understanding why academic integrity policies exist in the first place. Most professors are not trying to make your life harder. They are worried about three things:

1. Skill atrophy. If you outsource your writing to AI for four years of college, you will graduate without the ability to construct a clear, persuasive argument in your own voice. That ability matters in ways that extend far beyond the classroom.

2. Assessment integrity. If an AI did the work, your grade reflects the AI's capability, not yours. The professor cannot accurately assess what you have learned, which undermines the entire purpose of education.

3. Intellectual development. The struggle of writing — organizing messy thoughts, confronting gaps in your understanding, revising a paragraph for the fifth time — is not a bug. It is the process by which learning happens. Bypassing that process means bypassing the learning itself.

👁️ Perspective-Taking: Imagine you are a professor who spent 20 years developing expertise in your field. You assign a paper because you believe the process of writing it teaches students to think like a practitioner of your discipline. Then you receive 30 papers, and you suspect many were substantially AI-generated. How would you feel? What would you do?

Now imagine you are Priya — overwhelmed, working part-time, genuinely trying to learn but facing time pressures that feel impossible. How does the situation look from her perspective?

Neither perspective is wrong. The challenge is designing policies that honor both.

Priya's Decision Framework

Here is what Priya eventually develops for herself — a set of questions she asks before using AI on any academic task:

1. What skill is this assignment trying to build? If I use AI, will I still build that skill?
2. Could I do this without AI? If the answer is "I have no idea," then I need to learn the material first, not outsource it.
3. Am I using AI to enhance my thinking or replace it? Enhancement: "Help me see this problem from a different angle." Replacement: "Write this for me."
4. Would I be comfortable showing my professor exactly how I used AI? If the answer is no, that is a signal.
5. What does the course policy say? When in doubt, follow the explicit rules and ask the instructor about edge cases.

🔄 Check Your Understanding: Priya uses an AI chatbot to explain a statistical concept (p-values) that she is struggling with in her statistics course. She reads the explanation, thinks she understands, and then works through the problem set on her own. Is this appropriate AI use? Why or why not? Apply Priya's five questions.

14.5 Professional AI Use: Best Practices by Field

The academic integrity conversation is important, but it addresses only one context. Most of you will spend far more time using AI in professional settings than in academic ones. And professional AI use brings its own set of challenges, opportunities, and ethical considerations.

Journalism and Media

Journalists have been among the most thoughtful adopters of AI — partly because accuracy and attribution are foundational to their profession, and partly because they have seen firsthand what happens when AI-generated content goes wrong.

Emerging best practices in journalism include:

• AI for research, not for writing. Use AI to find patterns in data, summarize lengthy documents, or identify leads — but write the story yourself.
• Never trust AI-generated facts. Verify every claim independently. A fabricated quote attributed to a public official is not just an error — it is a potential lawsuit.
• Disclose AI use. Major outlets like the Associated Press and The New York Times have policies requiring disclosure when AI tools are used in reporting.
• Protect sources. Be cautious about entering sensitive or confidential information into AI tools, as the data may be used for model training.

Healthcare

In healthcare settings (which we will explore in depth in Chapter 15), AI use carries particularly high stakes:

• AI as decision support, not decision maker. A diagnostic AI should inform a physician's judgment, not replace it. The physician remains clinically and legally responsible.
• Patient consent and transparency. Patients have a right to know when AI is being used in their care.
• Verification against clinical evidence. AI recommendations should be checked against established clinical guidelines and the physician's own expertise.

Law

Legal professionals are grappling with AI in real time, and several high-profile incidents have highlighted the risks:

• In 2023, a lawyer submitted a brief containing citations fabricated by ChatGPT. The court imposed sanctions. The lawyer said he had assumed the AI was a "super search engine" that would only return real cases.
• The lesson: AI-generated legal research must be verified against actual legal databases (Westlaw, LexisNexis). Hallucinated case law is not a minor inconvenience — it is professional malpractice.

Education

Teachers and instructors are both users and subjects of AI disruption:

• Using AI to develop materials. Many instructors use AI to help generate quiz questions, discussion prompts, or illustrative examples — then review and edit the results.
• Using AI to provide feedback. Some educators use AI to give students initial feedback on drafts, freeing the instructor to focus on higher-order comments.
• Modeling good AI use. Instructors who use AI transparently in their own work model the kind of critical, disclosed AI use they want students to practice.

Business and Management

In corporate settings, AI use raises questions about confidentiality, quality control, and accountability:

• Confidential information. Entering proprietary data, financial information, or customer details into public AI tools may violate data protection obligations. Many companies now provide enterprise AI tools with data privacy protections.
• Quality control. AI-generated reports, analyses, and communications should be reviewed by a human with relevant expertise before being shared externally.
• Accountability. If an AI-generated recommendation leads to a bad business decision, the human who acted on it — not the AI — bears professional responsibility.

📋 Action Checklist: Professional AI Use

Before using AI in any professional context, ask yourself:

• [ ] Am I entering confidential or sensitive information? If so, is this tool approved for that data?
• [ ] Will I verify every factual claim in the output before using it?
• [ ] Am I transparent with relevant parties (colleagues, clients, readers) about my AI use?
• [ ] Do I understand the limitations of this tool for this specific task?
• [ ] Am I prepared to take full responsibility for the final output, regardless of how it was generated?
• [ ] Does my organization have an AI use policy? Am I in compliance?

14.6 Building Your Personal AI Policy

We have covered a lot of ground — prompting techniques, verification frameworks, academic integrity, professional best practices. Now it is time to pull it all together into something personal and practical: your own AI use policy.

A personal AI policy is not a legal document. It is a set of principles and guidelines you set for yourself — a deliberate, considered answer to the question: "How do I want to use AI in my life?"

Why bother writing one down? Because in the heat of a deadline, good intentions evaporate. When you are exhausted, overwhelmed, and a chatbot is offering to do your work for you, having a pre-committed set of principles gives you something to fall back on. Athletes do not decide their training regimen in the middle of a race. You should not decide your AI ethics in the middle of a crisis.

Template: My Personal AI Policy

Here is a template you can adapt. Fill in each section honestly — there are no right answers, only your answers.

📝 My Personal AI Policy

1. My values regarding AI use: What matters most to me when it comes to using AI? (Examples: honesty, learning, efficiency, creativity, fairness)

2. Academic use guidelines: - I will use AI for: ___ - I will not use AI for: ___ - When I am unsure, I will: ______

3. Professional use guidelines: - I will use AI for: ___ - I will not use AI for: ___ - I will disclose AI use when: ______

4. Verification commitment: Before acting on any AI-generated output, I will: ______

5. Skill protection: The skills I want to protect from atrophy by doing them myself are: ______

6. Privacy boundaries: I will not enter the following types of information into AI tools: ______

7. Review schedule: I will revisit and update this policy every: ______

Why Your Policy Needs a Review Schedule

AI tools change rapidly. Your circumstances change. A policy you wrote as a sophomore may not serve you as a senior. A policy you wrote as an individual contributor may not work when you manage a team. Build in a regular review — every semester, every six months, or whenever you change roles — to ensure your policy still reflects your values and your reality.

🪞 Self-Assessment: Your AI Readiness

Rate yourself on a scale of 1 (not confident) to 5 (very confident):

1. I can write a prompt that gets useful output on the first or second try.
2. I can identify when an AI output contains fabricated information.
3. I know the difference between using AI as a tool and using it as a crutch.
4. I can explain my AI use decisions to a professor or employer without embarrassment.
5. I have a clear set of principles guiding when and how I use AI.

If any score is below 3, revisit the relevant section of this chapter. If all scores are 4 or above, you are well positioned for the collaborative future we are building.

14.7 Chapter Summary

Let us take stock of what you have learned.

Prompting is a skill, not a trick. The quality of AI output depends heavily on the quality of your input. Effective prompts are specific, contextualized, formatted, constrained, and audience-aware. This is not about finding magic words — it is about communicating clearly with a system that responds to statistical patterns.

Four named techniques give you a practical toolkit. Zero-shot prompting (no examples), few-shot prompting (provide examples), chain-of-thought prompting (ask for step-by-step reasoning), and role prompting (assign a persona) each work for different situations and can be combined.

Verification is the core skill of AI literacy. The VIBE Check — Verifiable, Internally consistent, Balanced, Evidence-backed — gives you a systematic framework for evaluating any AI output. Never skip verification, no matter how good the output looks.

Academic integrity requires thoughtful judgment. There is a spectrum from clearly acceptable uses (research assistance, concept explanation) to clearly problematic ones (submitting AI-generated work as your own). The Skill-Building Test — "Is this assignment designed to build a skill? Will I still build it if I use AI?" — helps you navigate the gray areas. When in doubt, disclose.

Professional AI use carries its own risks. Confidential data, fabricated citations, and accountability gaps are real concerns across fields. The consistent principle: you are responsible for the output, regardless of how it was generated.

A personal AI policy turns good intentions into durable commitments. Writing down your principles — and revisiting them regularly — protects you from making decisions you will regret when you are under pressure.

📋 Key Concepts Introduced in This Chapter

Concept	Definition
Prompt engineering	The practice of designing inputs to AI systems to get better, more useful outputs
Zero-shot prompting	Giving a model a task without examples
Few-shot prompting	Providing examples of desired input-output patterns before the actual task
Chain-of-thought prompting	Asking the model to show step-by-step reasoning
Role prompting	Assigning the model a persona or expertise to shape its response style
VIBE Check	A verification framework: Verifiable, Internally consistent, Balanced, Evidence-backed
Skill-Building Test	A heuristic for academic AI use: does using AI still allow the intended learning?
Personal AI policy	A written set of principles guiding when, how, and why you use AI tools

🎯 Project Checkpoint: AI Audit Report — Step 14

Your task: Use AI tools to assist with one section of your AI Audit Report, and then critically document the experience.

Instructions:

1. Choose one section of your audit report that could benefit from AI assistance — perhaps background research, outlining an analysis, or exploring counterarguments to your position.

2. Use an AI tool to help with that section. Apply the prompting techniques from this chapter. Save your prompts and the AI's responses.

3. Evaluate the output using the VIBE Check framework. Document: - What was accurate and useful? - What was inaccurate or misleading? - What was missing? - How did you verify the claims?

4. Reflect on the experience (200–300 words): - How did AI assistance change your workflow? - What did the AI do well? What did it do poorly? - Would you use AI for this task again? Why or why not? - How does this experience connect to the themes of this chapter?

5. Transparency note: Add a brief disclosure to your audit report describing how AI tools were used in this section.

Deliverable: Your AI-assisted section, the documentation of your evaluation, your reflection, and your transparency note. Add to your AI Audit portfolio.

What's Next

In Chapter 15: AI in Healthcare — Promise, Peril, and the Patient, we shift from the general to the specific. Healthcare is one of the domains where AI has the most potential to save lives — and the most potential to cause harm. We will return to MedAssist AI, our anchor case study, and examine it through the lenses of efficacy, equity, trust, and regulation. You will see how the evaluation skills you built in this chapter apply in a context where the stakes are measured in human health and human lives.