← Back to the Using Artificial Intelligence outline The Course Maker
← Using Artificial Intelligence outline
Week 8 · Lecture outline

Lecture Outline — Week 8: Midterm Review & Exam

Using Artificial Intelligence · AI 101 Fall 2026 · Prof. Quinn Fictional sample

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Week: 8 of 16 · Fall 2026 · Midterm week — cumulative review (Weeks 1–7)
Objectives reviewed: 1 — What genAI/LLMs are and how they work · 2 — Effective prompting · 3 — Modalities and choosing the right tool
Format: Two 75-minute sessions; both are cumulative review. There is no new content this week.

This is the instructor lecture outline for Week 8. The two sessions together sweep the entire Objectives 1–3 arc (Weeks 1–7) before the Midterm window closes Sunday. Slides are in E-slides-week-08.pptx. The student prep kit (Study Guide, Exam-Prep Tutorial, Practice Exam) and the Midterm itself are the graded work this week.

Session 1 (Tuesday, Oct 20) — Objectives 1 & 2: What AI Is + All Four Prompting Weeks

Total: 75 minutes. No new content — active review and practice.

Segment 1 — Hook: "What Would You Tell a Friend?" (8 min)

Opening move: Before showing any slides, ask students to take 90 seconds and write down, in plain language, what they would tell a friend who asked: "What is generative AI and why can't I always trust it?"

Reveal: Have three or four students share. Use their words to launch the review. The goal: surface what's automatic, what's shaky, and what's still confused. The misconceptions that come up here are exactly the ones that cost points on the midterm.

Restate the spine of Objective 1:
- Generative AI creates new content (generates); it does not retrieve or search.
- The LLM is the engine inside the chatbot app — they are not the same thing.
- AGI is hypothetical and does not exist today. Today's tools are powerful but narrow.
- The fundamental mechanism: next-token prediction from patterns in training data.
- Fluency does not equal accuracy. Confident, specific, well-formatted text can be entirely wrong.

Misconception check: write on the board: "If the AI is fluent and confident, it's probably right." — Ask the class: True or False? Why?

Segment 2 — Objective 1 Deep Dive: The Four Limits (18 min)

Core vocabulary sweep — four limits that constrain everything the LLM does:

1. Token
- A token is a chunk of text — sometimes a full word, often part of a word, sometimes punctuation. The model generates one token at a time.
- Why it matters: unusual words may be split into multiple tokens; the model has a token budget per context.

2. Context window
- The total amount of text the model can "see" at once in a conversation.
- When the conversation grows beyond the window, early content falls out — the model is not ignoring it; it literally cannot see it.
- Classic trap: a larger context window does NOT make the model more accurate. It holds more text; the model still predicts tokens from patterns.
- Distinguish from the training cutoff: context window = real-time size (current conversation); training cutoff = knowledge date (what the model was trained on). These are two different, independent limits.

3. Training cutoff
- The date after which no new information was in the training data.
- The model can be confident about events after its cutoff and still be fabricating — it extrapolates from prior patterns.

4. Hallucination
- The AI generates plausible-sounding text that is factually wrong.
- Classic shapes: invented citations, fabricated statistics, confident but wrong summaries, simulated quotes presented as real.
- The fix is not to regenerate — it's to verify independently.

Worked example — live walkthrough:
- "Here's a scenario: I ask a chatbot for three academic citations on a niche topic. It returns three perfectly formatted references. What should I do before I cite them in my paper?"
- Walk through the correct answer: check each one in a library database or Google Scholar. Don't trust format; format is easy to generate and says nothing about whether the source is real.

The Turing test (brief):
- Alan Turing, 1950, "Computing Machinery and Intelligence" — a behavioral benchmark about conversational performance, not a proof of consciousness.
- Classic trap: an AI that passes the Turing test has NOT proven it is conscious or that it understands language the way a human does.

Quick interaction (2 min): "Your neighbor says: 'I upgraded to the bigger context window plan, so now the AI is more accurate.' What do you say?" — pair-share, then call on two or three pairs.

Segment 3 — Objective 2: Prompting Weeks 3–4 (20 min)

Week 3 — Conversation, Content & Emphasis (Skills 1–3):

Skill 1: Have a conversation and counter sycophancy.
- The AI tends to agree with the user's premise rather than push back. Classic example: a student says "My thesis is X, right?" and the AI says "Yes, absolutely."
- Counter move: ask for weaknesses before validation: "What are the three strongest objections to this argument?"
- Misconception to kill: "The AI will push back if you're wrong." It often won't.

Skill 2: Provide content.
- Pasting your actual document is far more effective than asking the AI to guess from training data.
- Privacy note: free tools may store inputs. Don't paste client data, employee data, or anything you wouldn't want stored in a cloud service.
- Verify outputs: if the AI adds a statistic or citation that wasn't in your document, it came from training data — or it was hallucinated. Always check against the original.

Skill 3: Emphasis.
- Markdown heading (## Task) = structure, labels a section
- XML-style tags (<task>...</task>) = separates instruction, content, constraints into named segments
- ALL CAPS for must-dos = signals a priority constraint
- Misconception to kill: politeness and urgency ("please" / urgency phrases / typing in all-caps without structure) are not the same as emphasis. Emphasis is structural, not motivational.

Week 4 — Meta-Prompting & Structured Prompts (Skills 4–5):

Skill 4: Meta-prompting.
- Ask the AI to interview you: "Ask me clarifying questions one at a time; when you have enough information, return a Markdown prompt template I can copy and reuse."
- Why it works: the AI surfaces what information it needs to do the job well.

Skill 5: The nine structured-prompt components.
- Context · Role · Goal · Audience · Constraints · Voice/Format · Data/Logic · Examples · Evaluation
- Role misconception (high-stakes): assigning a role shapes style and framing; it does not make the output factually accurate. "You are a licensed attorney" changes how the output sounds; it does not make the legal information reliable. Verify everything.
- Over-engineering: contradictory instructions (target experts AND beginners; be persuasive AND not preachy) degrade output. Use only the components that change the output for this specific task.
- Evaluation component: the built-in self-check — "Before returning, verify: Is this under 200 words? Is it in plain language? Does it avoid bullet points?" This is different from Constraints (what not to do in the content).

Worked example — "what's the prompting fix?" scenario:
- Weak prompt: "Help me write an email to my professor."
- Walk through the structured fix: add Goal (request an extension), Audience (a professor — formal register), Constraints (under 100 words, respectful, no begging). The resulting prompt is much more likely to produce a usable draft.

Segment 4 — Objective 2: Prompting Weeks 5–6 (20 min)

Week 5 — Examples, Structure & Control (Skill 6):

The shot vocabulary:
- Zero-shot: no examples — the AI uses only the instruction.
- One-shot: exactly one example.
- Few-shot: several examples (typically two to five) that teach the AI a specific format, voice, or pattern.
- Classic trap (high-stakes): "few-shot" does NOT mean exactly one example. One example is one-shot. Two to five is few-shot. This is engineered as a distractor on almost every prompting midterm.

Providing examples for voice and format:
- Paste two or three samples of your writing style → the AI writes in that voice.
- More targeted than describing the voice in words.

PII scrubbing:
- Replace names, IDs, and identifying details with placeholders ([NAME], [DATE], [COMPANY]) before pasting any sensitive document. The AI cannot reliably detect or strip PII on its own.

The control toolkit:
- Specify a count ("give me exactly five bullet points")
- Request a specific format ("return a Markdown table")
- Set a constraint ("do not use jargon; each item under 15 words")
- Ask for expansion ("take point 3 and write a full paragraph")

Regeneration misconception:
- Clicking Regenerate produces a different output — not a verified output. A new set of fabricated citations is still fabricated. Regeneration is a variety tool, not a fact-checking tool.

Week 6 — Simulations & Reusable Prompts (Skill 7):

The four simulation types:
1. Difficult-customer simulation — rehearse conflict or complaint handling before the real interaction.
2. Pre-mortem simulation — imagine the project has already failed; reason backward to surface causes before you start.
3. Decision role-play — hear multiple stakeholder perspectives to stress-test a plan.
4. Adaptive-tutor simulation — get personalized instruction at your own pace on a specific topic.

The non-negotiable rule (always on exams):
- AI-generated dialogue attributed to a historical figure is GENERATED, not transcribed from historical records.
- It must never be cited as a real quote. The AI imitates historical style from training data; it does not reproduce verified words.

Reusable prompt templates:
- Placeholder variables ([TOPIC], [AUDIENCE], [LENGTH]) are the innovation that turns a one-time prompt into a reusable template.
- The AI does not store or learn from prompts between sessions.

Quick interaction (3 min): "I run a simulation where the AI plays Abraham Lincoln and he says something about climate change. True or False: I can cite this in a history essay?" — pair-share.

Segment 5 — Session 1 Wrap + Student Questions (9 min)

Callback: rapid-fire review of the six Objective-1 and six Objective-2 classic traps (slides). Students call out the cure for each before it's revealed.

Tease Session 2: "Thursday we finish with Objective 3 — multimodal AI and tool choice — and then I'll walk you through the exam logistics and what the prep kit looks like in practice."

Point to the prep kit: Study Guide is available now. Open it tonight. The Exam-Prep Tutorial is the adaptive AI tutor that finds your weak spots; start there. The Practice Exam is the timed full-length run.

Session 2 (Thursday, Oct 22) — Objective 3: Multimodal AI + Exam Logistics

Total: 75 minutes. Objective 3 review + exam prep walkthrough + Q&A.

Segment 1 — Hook: The Modality Mismatch (8 min)

Opening scenario: "A student uploads a photo to DALL·E to find out what's in it. A second student uploads the same photo to Claude with a prompt asking it to describe the contents. Which student gets useful information, and why?"

Invite responses, then explain the direction principle:
- DALL·E is text-to-image: text in, image out. It generates new images from descriptions. It cannot analyze photos.
- Multimodal chatbots (ChatGPT with vision, Claude, Gemini) are image-to-text: image in, text out. They analyze photos and documents.
- Direction matters. The first student used the wrong tool for the task.

Segment 2 — Objective 3: Multimodal AI (Week 7) (22 min)

What "multimodal" means:
- AI systems that can process and/or generate more than one type of data — text, audio, images, documents.
- Misconception to kill: "ChatGPT, Claude, and Gemini are text-only." They are not. All three now support image uploads and voice input modes.

Skill 8: Voice prompting — the two-step process:
1. Your speech is converted to text by a transcription step (can introduce errors).
2. The AI processes the text and generates a response.

Errors in step one flow directly into step two. If the AI gives a strange or off-topic response in voice mode, the first fix is to check the displayed transcript — not to assume the AI misunderstood.

Skill 9: Record → transcribe → analyze workflow:
- Step 1: Record — capture the audio (meeting, lecture, voice memo).
- Step 2: Transcribe — convert audio to text using a transcription tool (e.g., Whisper-class, built-in phone recorder, a free web transcription app). This step can introduce transcription errors: mis-heard words, dropped phrases, background noise interference.
- Step 3: Analyze — paste the text transcript into an AI chatbot and prompt it to summarize, extract action items, identify themes, etc.

Two error-entry points:
- Transcription step: words may be mis-heard or dropped.
- AI analysis step: the AI may add details, smooth contradictions, or invent conclusions not present in the transcript. Always verify the summary against the original transcript.

Image analysis vs. image creation:
| Direction | Tool type | Example tools |
|---|---|---|
| Text → Image (creation) | Text-to-image generation | DALL·E, Midjourney, Adobe Firefly |
| Image → Text (analysis) | Multimodal chatbot with vision | ChatGPT (vision), Claude, Gemini |

Document/PDF analysis:
- Upload a PDF or spreadsheet to a multimodal chatbot and ask questions about its contents.
- The AI reads the document within the context window — for very long documents, the same context-window cautions apply.

Tool-to-task matching (Objective 3's core test skill):
- Convert a voice recording → text file: audio transcription tool
- Generate a new illustration from a description: text-to-image tool (DALL·E, Midjourney, Firefly)
- Extract data from a photo of a handwritten table: multimodal chatbot with image upload
- Ask questions about the contents of an uploaded PDF: multimodal chatbot with document upload
- Synthetic voice from text: voice generation tool (e.g., ElevenLabs)

Worked example — the verify-the-AI moment:
"Suppose you use the record-transcribe-analyze workflow to summarize a team meeting. The AI summary says the team decided to delay the project launch by three weeks. You don't remember that discussion. What do you do?"
- Walk through the correct response: compare the AI summary to the original transcript; then compare the transcript to the audio if needed. The AI may have fabricated a decision that was never made. Always verify.

Segment 3 — Objective 3 Misconception Cures (12 min)

Six Objective-3 traps (active review):
1. Chatbots are text-only → False. ChatGPT, Claude, Gemini handle images and audio.
2. Transcription is always accurate → False. Accents, background noise, and technical vocabulary cause errors. Always review the transcript.
3. AI image analysis = human seeing → False. Pattern recognition on pixel data; can produce confident, wrong descriptions.
4. DALL·E analyzes photos → False. DALL·E creates images; it doesn't analyze them.
5. Voice-mode off-answer means the AI misunderstood → The first fix is to check the transcript for a mis-heard word, not to blame the tool.
6. The AI analysis step never adds false content → False. The AI may fabricate in summaries; verify against the source transcript.

Live scenario practice (pair-share):
"A student uses voice mode and asks the AI to summarize notes on a specific paper. The AI gives a general overview of the field instead. What's the most likely cause and the best first fix?"
- Expected answer: the transcription step mis-heard the name of the paper or a key term. Check what text the AI received, correct it, and re-ask clearly.

Segment 4 — Exam Logistics & Strategy (15 min)

The exam, clearly stated:
- 20 items, 5 points each, 100 points total.
- Assignment group: Midterm (20% of the course grade).
- One attempt. AI is not permitted.
- Window: opens Mon Oct 19, due Sun Oct 25 at 11:59 p.m.

Coverage map (proportional):
- Objective 1 (Weeks 1–2): approximately 6 items — AI vocab, LLM mechanics, hallucination, context window, training cutoff, search vs. AI, Turing test.
- Objective 2 (Weeks 3–6): approximately 8 items — sycophancy, emphasis, structured prompts, meta-prompting, few-shot, regeneration, simulations, reusable templates.
- Objective 3 (Week 7): approximately 6 items — multimodal definition, voice two-step, record-transcribe-analyze, image creation vs. analysis, tool matching.

Item types: multiple-choice (single best answer) · multiple-answer ("select all that apply" — each option graded independently) · matching (one-to-one pairings) · true/false.

Test-taking strategy — this specific material:
1. Translate each scenario into its concept first. Underline cue words: creates / retrieves · shares / agrees · examples / instructions · text-in-image-out / image-in-text-out.
2. For "select all that apply" items, judge each option independently. The false option is almost always a named misconception.
3. For matching items, confirm every pair before submitting — one wrong pairing cascades.
4. If the stem asks "what's the best first fix?" — that's a prompting-fix scenario. Name the structural change (not politeness, not regeneration, not a different tool).
5. On every item, ask: "What classic trap is this distractor engineered around?" The wrong answers on this exam are not random — they're the specific misconceptions we've drilled all semester.

Segment 5 — The Prep Kit Walkthrough (10 min)

Three tools — use in order:

Study Guide (M-study-guide-week-08.md):
- A complete checklist by objective: every concept, every definition, every misconception cure.
- Read it first. Build your own one-page summary of the terms and traps you find hardest.

Exam-Prep Tutorial (N-exam-prep-tutorial-week-08.md):
- A boxed AI prompt you paste into any approved chatbot (Gemini, Claude, or ChatGPT).
- The AI tutors you adaptively: diagnoses weak spots, re-teaches, drills with fresh scenarios, ends with a Completion Summary.
- Submit the chat share link for Tutorial credit (graded, Lecture Tutorials group).
- Run this after the Study Guide, not before.

Practice Exam (O-practice-exam-week-08.md):
- 20 fresh items, same blueprint as the midterm, different stems. Zero overlap with the live exam.
- Sit it timed and closed-note (no Study Guide open) to simulate real conditions.
- Review every miss against the Study Guide before you sit the real exam.

Segment 6 — Callback + Tease the Back Half (8 min)

Callback: "Seven weeks ago we asked: what is AI, and can you trust it? You now have concrete answers to both — and you've practiced the skills that let you use AI well rather than get used by it."

Tease Week 9: "After the midterm, we open the second arc with Week 9 — the AI tool landscape. You'll match tools to tasks across chatbots, image generators, audio tools, video tools, research assistants, and coding tools. Then Week 10 goes deep on verification and hallucination — the course's central discipline, applied systematically for the first time. Then four weeks of Claude Cowork, the hands-on automation platform."

Hand-off: "The midterm window is open. Work the prep kit in order, sit the exam before Sunday, and then write an honest debrief in Discussion 8. The debrief is where you figure out what to do differently for the final. See you in Week 9."

Instructor FAQ

Question students commonly ask Instructor answer
"Is [specific concept] on the midterm?" "The midterm covers Objectives 1–3, which is everything from Weeks 1–7. If it was in a quiz, lecture, or Studio this term, it may appear."
"Can I use my notes on the midterm?" "No — one attempt, closed notes, AI not permitted. The exam is designed to be completeable with understanding, not memorization."
"What if I run out of time?" "20 items, roughly 5 minutes each across 100-minute window. Flag items you're unsure of and return. Don't sink 15 minutes into one item."
"Will the matching items have partial credit?" "Each matched pair is graded. A wrong pairing loses points; matching all four pairs correctly earns full credit."
"Is the practice exam items the same as the real exam?" "No — the practice exam shares the same blueprint (objectives, coverage proportions, item types) but has entirely different item stems. None of the practice items are the live exam questions."
"I submitted the exam — can I retake it?" "One attempt. The window is open for several days specifically so you can use the full prep kit before sitting it."

Scope flag

This week covers Objectives 1–3 (Weeks 1–7) only. Verification in depth (Objective 4, Week 10), Cowork (Objectives 5–6, Weeks 11–14), ethics/privacy (Objective 7, Week 15), and the capstone (Objective 8, Week 16) are not on the midterm. They are assessed on the cumulative final.

~ Prof. Quinn's edition · Fall 2026 · built with thecoursemaker.com