Midterm Exam — Cumulative (Weeks 1–7) · Objectives 1–3

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Scope: Cumulative — Weeks 1–7, Objectives 1–3 (what generative AI is and how it works conceptually · effective prompting — four weeks · multimodal AI and tool choice).
Format: 20 items, 100 points (5 each) · concept- and scenario-based items · mixed types (multiple-choice, matching, multiple-answer, true/false). AI is not permitted on the midterm.
Points: 100 · Assignment group: Midterm (20% of the course grade) · Window: opens at the start of the Week 8 module; due 6 days later · allowed attempts: 1. The midterm replaces Week 8's quiz, assignment, and AI Build Studio.

This is the human-readable exam with its vetted answer key and one-line feedback. The import-ready Classic QTI 1.2 is in L-midterm-week-08-qti.xml (generated by a validated Python script — parses with 20 items, every single-answer item exactly one correct). The item-bank/coverage note and the Canvas placement block are at the bottom of this file.

This is the live exam. Its paired ungraded rehearsal — O-practice-exam-week-08.md — mirrors this blueprint with fresh variants and shares none of these items.

---

Blueprint (items → objective → source week)

Coverage is proportional to teaching time: Obj 1 = 6 · Obj 2 = 8 · Obj 3 = 6. No trick questions; every single-answer item has exactly one correct option; the matching items pair one-to-one; the multiple-answer items list every correct option. The midterm does not reach verification in depth, Cowork, ethics/privacy, or the capstone (Weeks 9–16), which are assessed on the cumulative final.

#	Type	Concept	Objective	Week
1	Multiple choice	Generative AI vs. search — correct tool for the task	1	1
2	Multiple choice	AGI does not exist today	1	1
3	Matching	AI vocabulary (genAI / LLM / AGI / context window)	1	1–2
4	Multiple choice	Hallucinated citations — what to do	1	2
5	True / False	Larger context window ≠ more accurate	1	2
6	Multiple choice	Sycophancy — agreeing with a contested premise	2	3
7	Matching	Emphasis techniques → what each signals	2	3
8	Multiple choice	"What's the prompting fix?" — structured components	2	3–4
9	Multiple choice	Evaluation component vs. Constraints	2	4
10	Multiple answer	Structured-prompt truths (Role ≠ accuracy; Evaluation; "use only what changes output")	2	4
11	Multiple choice	Few-shot definition (three examples = few-shot)	2	5
12	True / False	Regenerate does not fix fabricated citations	2	5
13	Matching	Simulation types → primary purpose	2	6
14	Multiple choice	Simulated historical quotes are not real records	2	6
15	Multiple choice	Reusable prompt = placeholder variables	2	6
16	Multiple choice	Multimodal AI definition	3	7
17	Matching	Task → best tool type (modality matching)	3	7
18	Multiple choice	AI fabrication in the analysis step	3	7
19	True / False	Chatbots-are-text-only misconception	3	7
20	Multiple choice	"What's the prompting fix?" — voice-mode transcription error	3	7

Objective totals: Obj 1 = 6 items (30 pts) · Obj 2 = 8 items (40 pts) · Obj 3 = 6 items (30 pts) → 20 items, 100 points.

---

Questions, key, and feedback

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

Q1 (MC). A student needs to find the exact wording of a bill currently before Congress. Which tool is better suited for this task, and why?
- A. A generative AI chatbot — because it can retrieve live government documents
- B. A generative AI chatbot — because it will write a clean, error-free summary of the bill
- C. A search engine — because it finds and links to real, existing documents the student can read directly ✅
- D. Either tool equally — they both search the internet and return reliable results

Feedback: Generative AI creates new text; it does not retrieve existing documents. A search engine finds and links to real, existing sources — exactly right for locating the actual legislative text. Options A and D wrongly treat the chatbot as a retrieval tool; option B gets the direction right (use AI) but wrong (it still generates, not retrieves).

Q2 (MC). A classmate says today's chatbot is basically artificial general intelligence because it can do anything. Which response is most accurate?
- A. They are correct — modern chatbots have achieved AGI
- B. Today's chatbots are powerful narrow tools; AGI — a hypothetical system that can do any human intellectual task — does not yet exist ✅
- C. AGI and LLM are two names for the same technology
- D. AGI exists but is not yet released to the public

Feedback: AGI does not exist today. Current LLMs are impressive but narrow — they predict text from patterns. AGI (a system with the flexible, general intelligence to do any human intellectual task) remains hypothetical. A, C, and D all assert something that is not true as of the course's knowledge base.

Q3 (Matching). Match each term to its correct description.

Term	Correct description
Generative AI	The slice of AI that creates new content — text, images, audio, code
Large language model (LLM)	The text-prediction engine inside a chatbot app
AGI	A hypothetical future AI that could perform any human intellectual task
Context window	The maximum amount of text the model can see at once during a conversation

Feedback: Keep these distinct: Generative AI is the category (creates content); the LLM is the engine inside the app (text-prediction); AGI is the hypothetical future goal (not today's reality); the context window is the real-time size limit in the current conversation (not the model's knowledge date, which is the training cutoff — a different concept).

Q4 (MC). An LLM is asked for three academic citations on a niche research topic and returns three perfectly formatted references. The student submits the paper without checking them. What is the most likely problem?
- A. The LLM will refuse to provide citations it cannot verify
- B. The citations may look real but be entirely fabricated — the LLM predicts plausible text, not verified facts ✅
- C. The citations are accurate because the LLM searched a verified academic database
- D. The citations will be accurate as long as the topic is within the model's training period

Feedback: Hallucination — the LLM generates plausible-looking citations. Format (journal name, volume, page numbers, DOI format) is easy to fabricate because the model has seen thousands of real citations and knows the pattern. The fix: verify each citation independently in a library database or Google Scholar before citing it. Options A, C, and D all falsely imply the model vets its own outputs.

Q5 (True / False). True or False: Upgrading to an AI plan with a larger context window makes the model's answers more factually accurate.
- False ✅

Feedback: False. A larger context window holds more text in the current conversation — nothing more. The model still predicts the next token from patterns learned in training. It does not become more accurate, less prone to hallucination, or more knowledgeable about events after its training cutoff. Context window (size) and training cutoff (knowledge date) are two independent limits; neither controls the other.

---

Objective 2 — Effective Prompting (Weeks 3–6)

Q6 (MC). A student tells an AI: My essay argues that social media has had zero negative effects on mental health — agree? The AI immediately responds that this is a well-supported position, even though the research is contested. This is best described as —
- A. Hallucination — the AI invented a false fact about research
- B. Providing content — the student gave the AI context to work with
- C. Sycophancy — the AI agreed with the student's premise rather than pushing back on a contested claim ✅
- D. Accurate factual retrieval — the AI found a reliable consensus

Feedback: Sycophancy is the AI's tendency to validate rather than push back — especially when the user's framing is assertive. The AI didn't invent a fact from nothing (that would be hallucination); it agreed with a claim that was handed to it. Counter move: ask "What are the strongest arguments against this position?" before asking for agreement.

Q7 (Matching). Match each emphasis technique to what it signals to the AI.

Emphasis technique	What it signals
Markdown heading (## Task)	Marks a section the AI should read as a labeled division
XML-style tag wrapped around content	Separates instruction, content, and constraints into named segments
ALL CAPS for a must-do constraint	Signals a priority the AI must not miss or overlook

Feedback: All three send structural signals, not motivational ones. Markdown organizes sections; XML tags label each part of the prompt by name; ALL CAPS flags non-negotiable constraints. None of these make the AI "try harder" — they give it clearer structural information about what's instruction, what's content, and what's a hard constraint.

Q8 (MC). A student sends this prompt: Help me write an email to my professor. The AI returns a generic, overly formal email that misses the actual situation. What is the BEST single prompting fix?
- A. Add "please" at the start and resend — politeness improves AI output
- B. Retype the same prompt in ALL CAPS to show emphasis
- C. Add structured components: specify the Goal (request an extension), Audience (an undergraduate professor), and Constraints (keep it under 100 words, respectful, no begging) ✅
- D. Ask the AI to regenerate until the output is better

Feedback: The original prompt gives the AI no information about the situation, the purpose, the recipient, or the constraints. The fix is structure and specificity — adding Goal, Audience, and Constraints from the structured-prompt framework. Politeness (A) gives no useful information. ALL CAPS (B) alone adds emphasis but not the missing substance. Regenerating (D) produces another generic email.

Q9 (MC). A prompt ends with: Before returning your draft, check: Is it under 150 words? Does it avoid jargon? Does it have a clear call to action? Which structured-prompt component is this?
- A. Constraints — it lists what the output must not do
- B. Goal — it states the primary task
- C. Evaluation — it asks the AI to self-check before returning the output ✅
- D. Voice/Format — it sets the tone and length

Feedback: The Evaluation component is the built-in self-check — it directs the AI to test its output against success criteria before returning it. This is different from Constraints (A), which say what NOT to do in the content (e.g., "avoid bullet points"). Evaluation says "check the output against these criteria before you deliver it." Both are important components; they do different jobs.

Q10 (Multiple answer — select all that apply). Which of the following statements about structured prompts are TRUE?
- A. You do not need all nine components for every prompt — use only the ones that change the output ✅
- B. The Role component makes the AI's output factually accurate by granting it expertise
- C. The Examples component shows the AI the style or format you want; the Constraints component tells it what to avoid ✅
- D. A prompt with more components is always better than a shorter one
- E. The Evaluation component is a built-in self-check the AI applies before returning its output ✅

Feedback: True: A (use only what's needed), C (Examples = show style; Constraints = what not to do), E (Evaluation = self-check before delivery). B is false — Role shapes style and framing; assigning "licensed attorney" does not make the legal content factually reliable — always verify. D is false — contradictory instructions degrade output; more components is not automatically better.

Q11 (MC). A student provides three sample social-media captions to the AI before asking it to write two more in the same style. This technique is called —
- A. Zero-shot prompting, because the student did not explain the rules in words
- B. One-shot prompting, because there was only one type of content
- C. Few-shot prompting, because several examples teach the AI the desired pattern ✅
- D. Meta-prompting, because the student is asking the AI to write about writing

Feedback: Few-shot = a few examples (typically two to five) before the task. Zero-shot = no examples. One-shot = exactly one. Three examples = few-shot. The distractor in A is common: students think "if I didn't explain the rule in words, it's not a technique" — but examples ARE the technique in few-shot prompting. The AI learns the pattern from the examples themselves.

Q12 (True / False). True or False: When an AI produces a list of citations that appear to be fabricated, clicking Regenerate will prompt the model to search for and return verified, accurate citations.
- False ✅

Feedback: False. Regenerating produces a different set of outputs — not a verified set. A new list of citations can be a new list of equally fabricated citations. The model generates plausible text each time; it doesn't gain access to a verification database between attempts. The fix for suspected hallucinated citations is to verify each one independently using a library database or search engine.

Q13 (Matching). Match each simulation type to its primary purpose.

Simulation type	Primary purpose
Difficult-customer simulation	Rehearse handling complaints or conflict before a real interaction
Pre-mortem simulation	Imagine a project has already failed and reason backward to find the cause
Decision role-play	Hear multiple stakeholder perspectives to stress-test a plan
Adaptive-tutor simulation	Get personalized teaching and feedback at your own pace on a specific topic

Feedback: Each type has a distinct purpose: difficult-customer = rehearse conflict; pre-mortem = fail first, then fix before you start (reasoning backward is more effective than brainstorming forward); decision role-play = stress-test from multiple stakeholder angles; adaptive tutor = personalized learning at your pace.

Q14 (MC). A student runs an AI simulation in which a famous historical figure says something about a current event. The student plans to cite this as a quote in a research paper. What is the critical problem?
- A. The quote is accurate if the AI was trained on that figure's historical writings
- B. The quote is a generated line of text — the AI invented it; it is not a verified historical record and must never be cited as real ✅
- C. The quote is acceptable if the student adds a disclaimer
- D. The quote is accurate as long as the topic is historically plausible

Feedback: Every word the AI puts in a historical figure's mouth is generated, not transcribed. Training on historical texts means the AI can imitate a figure's style; it does not mean the specific generated words appear in any verified historical record. This is a critical rule with no exceptions: AI-generated dialogue attributed to real people must never be cited as real. (B and C might feel close — a disclaimer does not make a fabricated quote citable.)

Q15 (MC). What is the key innovation that turns a one-time prompt into a reusable template?
- A. Saving it on the AI's servers so it never needs to be re-typed
- B. Using Markdown so the AI can find it in its memory later
- C. Adding clear placeholder variables (e.g., [TOPIC], [AUDIENCE]) that you fill in each time you use it ✅
- D. Sending it to the AI multiple times so it learns the pattern

Feedback: Placeholder variables are the innovation — they mark exactly what changes for each new use. You maintain the template yourself (not on AI servers, which don't store prompts between sessions). Markdown improves structure but doesn't enable reuse on its own. Option D misconceives how AI learning works — the model doesn't learn your prompts between independent sessions.

---

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

Q16 (MC). The term "multimodal AI" refers to AI systems that —
- A. Use multiple servers to respond faster to queries
- B. Can only process text typed in a chat box
- C. Can process and/or generate more than one type of data, such as text, audio, images, and documents ✅
- D. Are only available on mobile devices

Feedback: Multimodal = multiple modes of data — text, voice, audio, images, documents. The major chatbots (ChatGPT, Claude, Gemini, Copilot) are increasingly multimodal; they are not text-only. Options A and D are irrelevant to what "multimodal" means. Option B is the classic misconception the question is built to catch.

Q17 (Matching). Match each task to the best tool type for completing it.

Task	Best tool type
Convert a recorded meeting to a text file	Audio transcription tool (e.g., Whisper-class or built-in phone feature)
Generate a new illustration from a text description	Text-to-image generation tool (e.g., DALL·E, Midjourney, Adobe Firefly)
Extract data from a photo of a handwritten table	Multimodal chatbot with image upload (e.g., ChatGPT, Claude, Gemini)
Ask questions about the contents of an uploaded PDF	Multimodal chatbot with document upload (e.g., ChatGPT, Claude, Gemini)

Feedback: Direction matters. Transcription converts audio → text. Text-to-image tools (DALL·E, Midjourney, Firefly) go text → image; they cannot analyze photos. Multimodal chatbots with vision (ChatGPT, Claude, Gemini) go image/document → text (analysis). Matching the direction of data flow to the right tool is the core Week-7 skill.

Q18 (MC). In the record-transcribe-analyze workflow, a student notices the AI meeting summary mentions a decision that was never actually discussed. What is the most likely explanation?
- A. The transcription tool searched the internet for common meeting outcomes and added them
- B. The recording was corrupted, causing random text to appear
- C. The AI added details during the analysis step — fabricating content not present in the transcript ✅
- D. The AI cannot make errors during analysis if the transcription was accurate

Feedback: The AI analysis step is an error-entry point: the AI may smooth contradictions, fill gaps, or invent plausible details that were never said. A perfect transcript does not guarantee a perfect summary — the AI can still fabricate in step three. The fix: always verify the summary against the original transcript. Option A misunderstands what transcription tools do; option D states the opposite of the truth.

Q19 (True / False). True or False: Major AI chatbots such as ChatGPT, Claude, and Gemini are text-only tools that cannot process images or audio.
- False ✅

Feedback: False. These assistants now support image uploads and, in many modes, voice/audio input. "Chatbots are text-only" is the classic Week-7 misconception — it was once true of early chatbots, but the major platforms are now multimodal. Check the current capabilities of each tool, as they evolve.

Q20 (MC). A student uses voice mode to ask an AI to summarize their notes on a topic, but the AI responds with a general overview that ignores the notes entirely. What is the BEST first fix?
- A. Switch to a different AI tool — voice mode is broken on this one
- B. Check the transcribed text the AI received — it likely mis-heard a key word — then rephrase and speak again more clearly ✅
- C. Type the same request instead, because voice mode is always less accurate than typing
- D. Ask the AI to replay the audio and try again

Feedback: Voice mode has two steps: speech → text (transcription), then text → AI response. If the AI gave an off-topic response, the most likely cause is a transcription error in step one — a mis-heard word changed the meaning. The best first fix is to check the displayed transcript, identify the error, correct it, and re-prompt clearly. Abandoning voice mode (A, C) without diagnosing the problem skips the fix. The AI cannot replay audio (D).

---

Answer key (quick reference)

Q	Answer	Q	Answer
1	C (search engine for real docs)	11	C (few-shot = several examples)
2	B (AGI doesn't exist today)	12	False (regenerate doesn't verify)
3	genAI→creates content / LLM→engine inside app / AGI→hypothetical / context window→max text at once	13	difficult-customer→conflict rehearsal / pre-mortem→fail backward / decision role-play→stress-test / adaptive tutor→personalized learning
4	B (citations may be fabricated)	14	B (generated text, never cite as real)
5	False (size ≠ accuracy)	15	C (placeholder variables)
6	C (sycophancy — agreed with contested premise)	16	C (processes multiple types of data)
7	Markdown→labeled division / XML tags→named segments / CAPS→priority constraint	17	recorded meeting→transcription tool / illustration from text→text-to-image / photo of table→multimodal chatbot vision / PDF questions→multimodal chatbot doc upload
8	C (structured components: Goal, Audience, Constraints)	18	C (AI fabricated in analysis step)
9	C (Evaluation = self-check before delivery)	19	False (chatbots are not text-only)
10	A, C, E	20	B (check transcript, correct error, re-prompt)

---

Quality gate (self-checked)

• Structure: 20 items, 5 points each, 100 points total; coverage Obj 1 = 6 · Obj 2 = 8 · Obj 3 = 6 matches the blueprint. Item-type mix: 11 multiple-choice + 3 matching + 3 true/false + 1 multiple-answer + 2 scenario-based multiple-choice (Q8, Q20 are the "what's the prompting fix?" items) = 20.
• Single-answer integrity: every multiple-choice and true/false item (Q1–Q2, Q4–Q6, Q8–Q9, Q11, Q14–Q16, Q18–Q20) has exactly one correct option; the three matching items (Q3, Q7, Q13, Q17) pair one-to-one; the one multiple-answer item (Q10) keys A, C, E (B and D are false distractors targeting real misconceptions).
• "What's the prompting fix?" coverage: Q8 (structured components for a weak prompt) and Q20 (voice-mode transcription error diagnosis) — both required by the brief.
• Matching coverage: four matching items (Q3, Q7, Q13, Q17) — exceeds the minimum of one required.
• Product-accuracy gate: PASS. Tools named factually (ChatGPT, Claude, Gemini for multimodal; DALL·E, Midjourney, Adobe Firefly for image generation; Whisper-class for transcription; ElevenLabs in study guide only). No fabricated features, no version/price claims. All conceptual claims (AGI status, hallucination mechanism, sycophancy, context window vs. training cutoff, multimodal definition, voice two-step, image direction) are accurate and consistent with VERIFIED_FACTS.md. Turing test described factually (Turing, 1950 — behavioral test, not proof of consciousness). No fabricated quotes or statistics anywhere.
• No fabrication: all depicted "AI errors" (Q4: fabricated citations; Q12: regeneration doesn't verify; Q14: simulated historical quote; Q18: AI adds fabricated meeting decision) are clearly framed as errors the student should catch — not presented as true.
• Integrity vs. practice exam: 0 items are shared with O-practice-exam-week-08.md (verified by full stem comparison — every concept slot uses a different scenario, scenario direction, or framing).

---

Item-bank & coverage note

All 20 items are fresh variants assembled from the Week 1–7 item banks (changed scenarios and contexts to reduce answer-sharing with the weekly quizzes), tagged course=AI101 · exam=midterm · weeks=1–7 · objectives=1–3 and deposited back into the banks for future per-term ($39) regenerations:

Objective	Drawn from banks	Items
1	Weeks 1–2 (What AI Is; How It Works & Its Limits)	Q1–Q5
2	Weeks 3–6 (Prompting I–IV; Simulations)	Q6–Q15
3	Week 7 (Multimodal AI)	Q16–Q20

Each term's update regenerates fresh midterm variants from these same banks; the paired practice exam is regenerated alongside and continues to share none of the live items.

---

Canvas placement block

canvas_object             = Quizzes::Quiz
title                     = "Midterm Exam -- Cumulative (Weeks 1-7)"
assignment_group          = "Midterm"
points_possible           = 100
grading_type              = points
available_from_offset_days = 0        # opens at the start of the Week 8 module
due_offset_days           = 6        # 6 days after module start
published                 = true
allowed_attempts          = 1
shuffle_answers           = true
ai_permitted              = false     # AI is not permitted on the midterm
provenance                = "~ Prof. Quinn's edition · Fall 2026 · built with thecoursemaker.com"