Week 8 · Study guide

Midterm Study Guide · Weeks 1–7 (Objectives 1–3)

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

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
This is a student-facing review page. Work through it before running the Exam-Prep Tutorial and taking the Practice Exam. (This guide points to those two — it doesn't repeat them.)

Integrity note for students. Every practice item on this page is a fresh variant — a new scenario, different wording — with a vetted answer. None of these are the live midterm questions. Working them builds the skill the midterm tests, the honest way.

What the midterm covers (read this first)


Exam	Midterm — cumulative, Weeks 1–7, Objectives 1–3
Format	20 items, 100 points (5 each). Concept- and scenario-based: most items hand you a short situation and ask you to classify, identify, choose, or diagnose. Expect a mix of multiple-choice, three matching items (vocabulary, emphasis techniques, simulation types, modality-to-task), one "select all that apply", and three true/false. Two items are scenario-based "what's the prompting fix?" questions. AI is not permitted on the midterm.
Coverage	Obj 1 = 6 items (AI vocabulary + how LLMs work + limits) · Obj 2 = 8 items (all four prompting weeks) · Obj 3 = 6 items (multimodal AI + tool choice). Study Objective 2 hardest — it's the biggest slice — and make sure Objective 1's misconception cures are automatic.
Weight	The midterm is 20% of your course grade.
When / where	Opens in the Week 8 module (Monday Oct 19); window closes Sunday Oct 25, 11:59 p.m.; one attempt. This guide and the exam-prep tutorial post before the window so you can prepare. There is no weekly quiz, assignment, or AI Build Studio in Week 8 — the midterm replaces them. Discussion 8 (the midterm debrief) still runs.

How to use this guide. Each objective has the same four parts: (A) the key ideas in plain language, (B) the definitions / terms / techniques, (C) the predictable mistakes and their cures, and (D) where to review. After all three objectives come fresh self-check items (with answers), a dated study plan, and test-taking strategy.

Objective 1 — What Generative AI Is and How It Works (Weeks 1–2) · 6 items

(A) Key ideas, plain language

Week 1 asks: what is this thing, and can you trust it? Generative AI creates new content — it doesn't search or retrieve. The vocabulary matters: AI / generative AI / LLM / AGI are nested, related but distinct concepts. The working mindset is "general to specific, iterate, human stays the judge." Week 2 asks: why can't you always trust it? Because it predicts plausible text from patterns, not verified facts. Four limits constrain every LLM: the token, the context window, the training cutoff, and hallucination. These are the vocabulary of Week 2, and knowing them cold is the payoff of the whole first arc.

(B) Definitions, terms, techniques

Generative AI — the slice of the AI field that creates new content (text, images, audio, code) from a request. It generates; it does not search or retrieve.
LLM (large language model) — the text-prediction engine inside a chatbot app. The chatbot is the app; the LLM is the engine. These are different things.
AGI (artificial general intelligence) — a hypothetical future AI that could do any human intellectual task. It does not exist today. Today's tools are powerful but narrow.
Token — a small chunk of text the model processes: sometimes a full word, sometimes part of a word, sometimes punctuation. LLMs generate one token at a time.
Context window — the maximum amount of text the model can "see" at once in a single conversation. When the conversation exceeds the window, early content falls out. A larger context window holds more text — it does not make the model more accurate. Context window (size, current session) ≠ training cutoff (knowledge date).
Training cutoff — the date after which no new information was included in the model's training data. The model can be confidently wrong about events after this date. Separate from the context window.
Hallucination — AI output that is confident and fluent but factually wrong. Classic shapes: invented citations, fabricated statistics, confident but wrong summaries, simulated quotes presented as real. Fluency ≠ truth.
The Turing test (Alan Turing, 1950, "Computing Machinery and Intelligence") — a behavioral test asking whether a human evaluator can distinguish a machine from a human in text exchange. Passing it is a meaningful conversational benchmark — not proof of consciousness or genuine human-style understanding.
Search engine vs. AI chatbot — a search engine finds and links to real, existing documents; a generative AI chatbot creates new text from patterns. Use the right tool for the task.
The working mindset — general → specific, iterate; the machine has no brain, use your own; be the judge.

(C) Predictable mistakes → cures

❌ "AI is basically a search engine that finds the right answer." → ✅ AI generates new text. A search engine retrieves real, existing documents. Different tools, different tasks.
❌ "If the AI's answer is fluent and confident, it's probably right." → ✅ Fluency ≠ truth. The model predicts plausible text; it can be confidently and specifically wrong.
❌ "A bigger context window makes the AI more accurate." → ✅ It holds more text in the conversation — accuracy is unchanged. The model still predicts tokens from patterns.
❌ "AGI is here — it's what powers ChatGPT." → ✅ AGI does not exist. Today's tools are narrow; AGI remains hypothetical.
❌ "The Turing test proves an AI is conscious." → ✅ It's a behavioral benchmark — it tests whether the evaluator can tell the machine from a human. It says nothing definitive about consciousness.
❌ "The context window is the same as the training cutoff." → ✅ Two independent limits: context window = real-time size (current conversation); training cutoff = knowledge date (what was in training data).

(D) Review in the module

Week 1 → Lecture Outline B1, Slides (Deck 1), Readings H1, Tutorial C1, Practice D1, Quiz F1. Week 2 → Lecture Outline B2, Slides (Deck 2), Tutorial C2, Practice D2, Quiz F2.

Objective 2 — Effective Prompting (Weeks 3–6) · 8 items — the biggest slice

(A) Key ideas, plain language

Four weeks of building the skill of prompting well. Week 3: three skills — have a conversation (and catch sycophancy), provide content, use structural emphasis. Week 4: two skills — meta-prompt to build a better prompt, use the nine structured-prompt components. Week 5: one skill — use zero/one/few-shot examples and a control toolkit (count, format, constraints, expansion). Week 6: one skill — run a simulation; build a reusable template. The thread through all four weeks: structure and specificity produce better outputs; regenerating doesn't fix facts; the AI's default behavior includes real traps (sycophancy, fabricated citations, over-agreement); and nothing the AI generates from a role-play or simulation should be cited as a verified fact.

(B) Definitions, terms, techniques

Week 3 — Conversation, Content & Emphasis:
- Sycophancy — the AI's tendency to agree with or validate the user rather than push back on errors or contested claims. Counter move: ask for the strongest objections or weaknesses before asking for agreement.
- Providing content — pasting your actual document/notes into the prompt so the AI works from your material, not from guesses drawn from training data. Contrast: asking blind — asking without pasting the material.
- Context-window caution when providing content — very long documents may lose attention to early sections; always verify the output came from your material.
- Privacy of pasted content — on most free AI tools, inputs may be stored; never paste client data, employee data, medical records, or anything you wouldn't want in cloud storage.
- Markdown heading (## Task) — structural signal: labels a section as a division.
- XML-style tag (...) — structural signal: separates instruction, content, and constraints into named segments.
- ALL CAPS for must-do constraints (DO NOT...) — priority signal: flags something the AI must not overlook.
- Emphasis is structural, not motivational. "Please" and urgency phrases are not emphasis techniques.

Week 4 — Meta-Prompting & Structured Prompts:
- Meta-prompting — asking the AI to help you write the prompt. The move: "Ask me clarifying questions one at a time; when you have enough information, return a Markdown prompt template I can copy and reuse."
- The nine structured-prompt components: Context · Role · Goal · Audience · Constraints · Voice/Format · Data/Logic · Examples · Evaluation. Use only the components that change the output for your specific task.
- Role — shapes style and framing; does NOT grant factual expertise or accuracy. "You are a licensed attorney" changes how the output sounds; the legal content is still generated text and must be verified.
- Constraints — what the output must NOT do (e.g., "do not use jargon"; "keep each item under 15 words").
- Evaluation — the built-in self-check: "Before returning, check: is this under 200 words? Does it avoid bullet points?" Different from Constraints — Evaluation tests the output after it's written; Constraints constrain what's written.
- Over-engineering — too many contradictory components degrade output. Example: "target experts AND beginners; be persuasive AND not preachy." The fix: decide the audience and goal before adding voice/constraint details.

Week 5 — Examples, Structure & Control:
- Zero-shot — no examples before the task; the AI uses only the instruction.
- One-shot — exactly one example to set the pattern.
- Few-shot — several examples (typically two to five) to teach a format, voice, or pattern. Few ≠ one. This is the most important misconception-cure in Week 5.
- Regenerating — produces a different output; does NOT produce a verified one. A new list of citations can be a new list of fabricated citations. Regeneration ≠ fact-checking.
- Control toolkit — specifying a count, requesting a format, setting a constraint, asking for expansion.
- PII scrubbing before pasting — replace names, IDs, and identifying details with placeholders ([NAME], [ROLE], [DATE]) before pasting any sensitive document into an AI tool. You do the scrubbing; the AI cannot reliably do it for you.

Week 6 — Simulations & Reusable Prompts:
- Simulation types — difficult-customer (rehearse conflict), pre-mortem (fail backward to find causes), decision role-play (stress-test from multiple stakeholder perspectives), adaptive tutor (personalized learning at your own pace).
- AI-generated historical dialogue is NOT a verified historical record — training on historical texts allows the AI to imitate a figure's style; it does not mean the specific generated words appear in any verified source. Never cite a simulated quote as real.
- Reusable prompt template — a prompt saved with placeholder variables ([TOPIC], [AUDIENCE], [LENGTH]) that you fill in each use. The AI does not store prompts between sessions; you maintain the template yourself.
- Placeholder variables — the key innovation that converts a one-time prompt into a reusable tool.

(C) Predictable mistakes → cures

❌ "The AI will push back if I'm wrong." → ✅ Sycophancy is the opposite tendency. Ask for weaknesses explicitly.
❌ "Adding 'please' is an emphasis technique." → ✅ Emphasis is structural (Markdown / XML tags / CAPS). Politeness gives no structural information.
❌ "The Role component makes the AI factually accurate." → ✅ Role shapes style and framing only. Always verify; never trust a role-assigned output on high-stakes facts.
❌ "Few-shot means exactly one example." → ✅ One example = one-shot. Few-shot = two to five.
❌ "Regenerating fixes fabricated citations." → ✅ Regenerating produces a different output — still generated, not verified.
❌ "The AI-generated historical quote can be cited with a disclaimer." → ✅ A generated quote is not a historical record regardless of disclaimers. Never cite it as real.
❌ "More structured-prompt components always = better prompt." → ✅ Contradictory instructions degrade output. Use only what changes the output for this task.
❌ "AI will detect and strip PII from what I paste." → ✅ You must scrub PII before pasting, using placeholder variables.

(D) Review in the module

Week 3 → B3, Deck 3, C3, D3, F3. Week 4 → B4, Deck 4, C4, D4, F4. Week 5 → B5, Deck 5, C5, D5, F5. Week 6 → B6, Deck 6, C6, D6, F6.

Objective 3 — Multimodal AI and Tool Choice (Week 7) · 6 items

(A) Key ideas, plain language

Week 7 asks: what if the input or output isn't text? Multimodal AI handles multiple types of data — text, voice, audio, images, documents. The major chatbots are NOT text-only anymore. Three skills: voice prompting (two-step: speech → text → AI), the record-transcribe-analyze workflow (with two real error-entry points), and matching the right tool to the task (especially the direction of data flow: text-to-image vs. image-to-text).

(B) Definitions, terms, techniques

Multimodal AI — AI systems that can process and/or generate more than one type of data (text, audio, images, documents).
Voice prompting (Skill 8) — two steps: (1) speech converted to text via a transcription step (can introduce errors); (2) AI processes the text and generates a response. Always check the displayed transcript if the AI's response seems off.
Record → transcribe → analyze workflow (Skill 9):
Step 1: Record — capture the audio.
Step 2: Transcribe — convert audio to text using a transcription tool. Error-entry point 1: mis-heard words, dropped phrases, background noise.
Step 3: Analyze — paste the text into an AI assistant for analysis. Error-entry point 2: the AI may add details, smooth contradictions, or invent conclusions not in the transcript.
Always verify the summary against the original transcript.
Text-to-image generation — text in, image out. Tools: DALL·E (OpenAI), Midjourney, Adobe Firefly. These tools do NOT analyze photos.
Image analysis — image in, text out. Tools: multimodal chatbots with vision (ChatGPT, Claude, Gemini). These tools can describe images, extract text, and answer questions about them.
Document analysis — upload a PDF, spreadsheet, or document; multimodal chatbots can read it and answer questions about it (within the context window).
Direction matters: the key is knowing which way data flows. DALL·E creates; it doesn't analyze. A multimodal chatbot with vision analyzes; it doesn't create illustrations on its own.
Audio transcription tools — Whisper-class tools and phone built-ins convert audio to text. They are not AI chatbots; they are a step in the workflow.
Voice synthesis — tools like ElevenLabs generate realistic synthetic speech from text (text → audio).

(C) Predictable mistakes → cures

❌ "ChatGPT, Claude, and Gemini are text-only." → ✅ False. These assistants now handle images and audio.
❌ "Transcription is always accurate." → ✅ False. Accents, background noise, crosstalk, and technical terms cause errors. Always review the transcript.
❌ "AI image analysis is exactly like human seeing." → ✅ AI performs pattern recognition on pixel data — confident descriptions can be wrong. Use the result as a starting point, not a ground truth.
❌ "I can upload a photo to DALL·E and it will tell me what's in it." → ✅ DALL·E creates images from text; it doesn't analyze photos. Use a multimodal chatbot with vision for analysis.
❌ "If my voice-mode answer is wrong, the AI misunderstood me." → ✅ First check the transcript — a mis-heard word in the transcription step is the most common culprit.
❌ "The AI analysis step only uses the transcript — it never adds content." → ✅ False. The AI may fabricate in the summary. Verify against the original transcript.

(D) Review in the module

Week 7 → B7, Deck 7, C7, D7, F7.

Self-check items (all fresh — vetted answers)

None of these are live midterm items. New scenarios, new wording. Cover the answers, work each one, check.

Obj 1 — Self-check

True/False: A student asks an AI for the full text of a recent Supreme Court decision. The AI returns a fluent, well-cited response. Should the student cite it in a law brief?
→ False. The AI generated plausible text; the actual decision may differ. Use the official court record from a legal database.
What are the two limits that are most commonly confused — and how do you tell them apart?
→ Context window (real-time size in the current session) vs. training cutoff (knowledge date — what the model was trained on). One is about now; one is about history.
True/False: AGI is currently available on paid plans of the major AI platforms.
→ False. AGI doesn't exist. Today's tools are narrow; AGI is hypothetical.
What does the Turing test measure, and what does it NOT prove?
→ It measures whether a human evaluator can distinguish a machine from a human in text exchange (a conversational performance benchmark). It does NOT prove consciousness or genuine human-style understanding.

Obj 2 — Self-check

A student assigns "Role: Olympic marathon coach" in a training-plan prompt. Will the output be factually accurate about periodization and injury prevention?
→ Not necessarily. Role shapes style and framing; it doesn't grant factual accuracy. Verify anything high-stakes against expert sources.
What's the difference between Constraints and Evaluation in a structured prompt?
→ Constraints say what the output must NOT do (e.g., "no jargon"). Evaluation says what to CHECK before returning the output (e.g., "verify: is this under 200 words?"). Constraints shape content; Evaluation tests the finished output.
Three social-media captions are pasted before a task instruction. Is this zero-shot, one-shot, or few-shot?
→ Few-shot (three examples = several). One example would be one-shot; no examples would be zero-shot.
A student asks the AI to play Thomas Jefferson and "Jefferson" claims something about modern climate science. Can the student cite this quote?
→ No. It's generated text, not a historical record. AI-generated dialogue attributed to real people must never be cited as real.

Obj 3 — Self-check

In the record-transcribe-analyze workflow, name the two error-entry points.
→ (1) The transcription step (mis-heard words, dropped phrases); (2) the AI analysis step (fabricated or invented details in the summary).
A student wants to generate a product illustration from a text description. Which type of tool should they use?
→ A text-to-image generation tool (e.g., DALL·E, Midjourney, Adobe Firefly) — text in, image out.
A student wants to extract key data from a photo of a handwritten survey form. Which type of tool should they use?
→ A multimodal chatbot with image upload (e.g., ChatGPT with vision, Claude, Gemini) — image in, text out.
True/False: Transcription tools and AI chatbots are interchangeable in the record-transcribe-analyze workflow.
→ False. They serve different steps: the transcription tool converts audio to text (step 2); the AI chatbot analyzes and summarizes the text (step 3). They are not interchangeable.

Study plan — a dated countdown

Built for the Week 8 midterm. Adjust exact dates to your section's posted exam day; the rhythm is what matters.

When	Do this (≈45–75 min)
~7 days out (Week 7, after class)	Read this guide's Obj 1 & 2 sections. Work the Obj 1 & 2 self-checks (especially the misconception cures). Build your own one-page list of the terms and traps you find hardest — the six Obj-1 cures and the eight Obj-2 cures.
~5 days out	Read Objective 3 carefully. Work the Obj 3 self-checks. Make sure you can distinguish text-to-image from image-to-text, name the two error-entry points in the workflow, and name the first fix for a bad voice-mode response.
~3 days out	Run the paired Exam-Prep Tutorial (`N-exam-prep-tutorial-week-08`) in an approved chatbot (Gemini, Claude, or ChatGPT) — it diagnoses your weak spots across all three objectives and drills them with fresh scenarios. Submit the share link for Tutorial credit.
~2 days out	Take the Practice Exam (`O-practice-exam-week-08`) under timed, closed-note conditions. Score it; list every concept you missed.
~1 day out	Re-teach only the topics you missed on the practice exam. Use this guide's mistake-cures and the relevant Lecture Tutorial. Sleep — memory consolidates overnight.
Exam day	Skim your one-page trap/cure list. Read each item twice; answer the question actually asked. AI is not permitted — bring your understanding.

Two paired tools — use both (don't skip):
- Exam-Prep Tutorial (N-exam-prep-tutorial-week-08) — adaptive AI tutor that diagnoses, re-teaches, and drills across all three objectives. Best for active recall and finding weak spots.
- Practice Exam (O-practice-exam-week-08) — full-length, mirror-format run. Best for pacing and a final readiness check.

(This guide points to both on purpose — it doesn't duplicate them.)

How the midterm is graded + test-taking strategy

How it's graded.
- 100 points across 20 items, 5 points each, weighted toward application (read a scenario; identify the technique, name the trap, choose the fix) rather than bare vocabulary recall. Matching items: one-to-one pairings — each pair is graded. Multiple-answer items: each option graded independently.
- The midterm is 20% of your course grade. It replaces Week 8's quiz, assignment, and AI Build Studio. One attempt; AI not permitted.
- Coverage: Obj 1 = 6 · Obj 2 = 8 · Obj 3 = 6. Study Objective 2 hardest; it's the biggest slice.

Honest test-taking strategies for this material.
1. Translate the scenario into its concept first. Underline cue words: creates / retrieves · agrees without pushback · structural heading / named segment / priority · interview one-at-a-time / nine components / self-check before delivery · zero / one / few examples · text-in → image-out / image-in → text-out.
2. For "select all that apply," judge each option independently. The false option is almost always a named misconception (Role = accuracy; more components = better; regenerating verifies facts).
3. For matching, confirm every pair before submitting. One wrong pairing can cascade into multiple lost points.
4. On prompting-fix scenarios, look for the structural change. Politeness, urgency, and shouting in all-caps are not the right answer; the fix is almost always adding the missing Goal/Audience/Constraint, or switching to few-shot examples, or checking the transcript.
5. Watch the six Objective-1 reversals. These are the most common source of error: fluency ≠ truth; bigger context window ≠ more accurate; AGI doesn't exist; context window ≠ training cutoff; Turing test ≠ consciousness proof.
6. Do easy items first, flag hard ones, and budget time. 20 items, 5 points each — don't sink 15 minutes into one item.
7. On matching items, use process of elimination. If you're sure about three pairs, the fourth follows automatically.

Canvas placement block

canvas_object   = Page
title           = "Midterm Study Guide -- Weeks 1-7 (Objectives 1-3)"
module          = "Week 8 -- Midterm Review & Exam"
grading_type    = not_graded
available_from  = 2026-10-17      # posts before the Week 8 exam window opens
published       = true
provenance      = "~ Prof. Quinn's edition · Fall 2026 · built with thecoursemaker.com"

Term-update note: each term's $39 update regenerates fresh self-check variants from this same scope — the live midterm is never reproduced here.

The per-term $39 update (fresh assessment variants, re-paced to your next calendar) referenced above is on the roadmap — coming soon. Today's download is yours to keep, but it doesn't refresh itself.

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