Midterm Practice Exam (ungraded) · Weeks 1–7 (Objectives 1–3)

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
What this is: a low-stakes rehearsal for the cumulative midterm. It mirrors the real exam's blueprint — same coverage, item-type mix, length, and concept- and scenario-based difficulty — but is built from fresh item-bank variants and shares none of the live midterm's questions.
Settings: ungraded (0 points) · unlimited attempts · feedback shown after submission · opens before the exam window so you can prepare.

This is the human-readable practice exam with its vetted answer key and feedback. The import-ready Classic QTI 1.2 is in O-practice-exam-week-08-qti.xml (generated by a validated Python script — parses with 20 items). The Canvas placement block is at the bottom.

Integrity note for students. Every item here is a fresh variant — a new scenario and wording — with a pre-vetted answer. None of these are the live midterm questions. Working them builds the skill the midterm tests, honestly. The paired live exam is L-midterm-week-08.md.

---

Blueprint (mirrors the midterm)

Coverage is proportional to teaching time, matching the real exam: Obj 1 = 6 · Obj 2 = 8 · Obj 3 = 6. (The actual midterm items are not listed here — only the shared structure.)

#	Type	Concept	Objective	Week
1	Multiple choice	Generation vs. search — AI returns generated text, not the real article	1	1
2	Multiple choice	LLM (engine) vs. chatbot (app)	1	1
3	Multiple choice	Fluency ≠ truth — confident text must be verified	1	1
4	Matching	AI limits vocabulary (token / training cutoff / hallucination / sycophancy)	1	1–2
5	Multiple choice	Context window — long document, early content falls out	1	2
6	True / False	Turing test — behavioral benchmark, not proof of consciousness	1	2
7	Multiple choice	Providing content vs. asking blind	2	3
8	Multiple choice	"What's the prompting fix?" — emphasis + constraints	2	3
9	Multiple choice	Meta-prompting (clarifying questions + Markdown template)	2	4
10	Multiple answer	Meta-prompting and structured-prompt truths	2	4
11	Multiple choice	Zero/one/few-shot — correct definitions	2	5
12	Multiple choice	PII scrubbing with placeholder variables	2	5
13	True / False	AI-generated historical dialogue cannot be cited as accurate	2	6
14	Multiple choice	Simulation specificity fix	2	6
15	Multiple choice	Pre-mortem simulation — what it does	2	6
16	Multiple choice	Voice-mode two-step process	3	7
17	Multiple choice	DALL·E is creation, not analysis	3	7
18	Matching	AI tool to primary function	3	7
19	True / False	Transcription accuracy misconception	3	7
20	Multiple answer	Error-entry points in record-transcribe-analyze workflow	3	7

Objective totals: Obj 1 = 6 · Obj 2 = 8 · Obj 3 = 6 → 20 items (ungraded; mirrors the 100-point midterm's emphasis).

---

Questions, key, and feedback (feedback releases after you submit)

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

P1 (MC). A student asks a generative AI chatbot to find a recently published news article and paste its exact text. Which response best describes what will actually happen?
- A. The chatbot retrieves and pastes the real article verbatim from a news database
- B. The chatbot generates a plausible-sounding account of the event, which may not match any real article ✅
- C. The chatbot searches Google News and returns a verified link
- D. The chatbot refuses, because it cannot discuss recent events

Feedback: Generative AI generates — it does not retrieve. When asked for a specific article, it will produce a fluent, plausible-sounding text that may contain fabricated details, wrong dates, or a fictional headline. Use a search engine to find real, existing articles; use the AI to analyze or summarize an article you've already found and pasted.

P2 (MC). Which statement correctly distinguishes the chatbot app from the large language model (LLM)?
- A. The chatbot app and the LLM are the same thing with different names
- B. The LLM is the interface you see; the chatbot app is the engine that generates text
- C. The chatbot app is the interface you interact with; the LLM is the text-prediction engine inside it ✅
- D. The LLM is a search engine that the chatbot app queries for answers

Feedback: The chatbot is the app (the interface, the buttons, the conversation history, the subscription). The LLM is the engine inside — the model that predicts text. These are different things — a company might update the interface without changing the model, or swap the model without changing the interface. The LLM is not a search engine (D) and not the interface (B).

P3 (MC). An AI assistant writes a confident, fluent, three-paragraph analysis of a scientific study and cites specific statistics. What is the most important caution a careful reader should keep in mind?
- A. Fluent, well-structured text is reliable — the AI would only write this if the facts were correct
- B. Confidence and fluency do not guarantee factual accuracy — the statistics and citations should be independently verified ✅
- C. The analysis is reliable as long as the study it cites is within the model's training period
- D. The AI will add a disclaimer if any statistic is uncertain

Feedback: Fluency does not equal truth. An LLM generates plausible-sounding text from patterns — it can produce a confident, well-formatted analysis with fabricated statistics and invented citations. The fix is to check every specific claim independently: open each citation in a library database and verify each statistic at its source. The AI will not reliably add disclaimers when it's wrong (D).

P4 (Matching). Match each AI concept to its correct description.

AI concept	Correct description
Token	A small chunk of text — sometimes a full word, sometimes part of a word or punctuation
Training cutoff	The date after which no new information was included in the model's training data
Hallucination	AI output that is confident and fluent but factually wrong
Sycophancy	The tendency for an AI to agree with or validate the user rather than push back on errors

Feedback: Keep these distinct. Token = the unit of text the model processes. Training cutoff = the knowledge date (separate from the context window). Hallucination = confident and wrong output. Sycophancy = agreeing with the user rather than correcting an error — a different failure mode from hallucination (hallucination invents facts; sycophancy validates false claims the user presents).

P5 (MC). A student pastes a very long document into a chat and then asks about the beginning of the document. The AI's answer seems to ignore the opening sections. The most accurate explanation is —
- A. The AI gives extra weight to the earliest content and ignores recent parts
- B. The beginning of the document may have fallen outside the context window and is no longer visible to the model ✅
- C. The AI stores the full document in permanent memory and will always have access to it
- D. The AI understood the document but chose not to answer about the beginning

Feedback: The context window has a fixed size. When a document (plus conversation) exceeds the window, early content falls out — the model doesn't ignore it by choice; it literally can't see it. The fix: for very long documents, paste the relevant section separately, or use a tool designed for long-document analysis. The AI has no "permanent memory" between sessions (C is false).

P6 (True / False). True or False: An AI that passes the Turing test (Alan Turing, 1950) has proven it is genuinely conscious and understands language exactly the way a human does.
- False ✅

Feedback: False. The Turing test is a behavioral test — it asks whether a human evaluator can distinguish the machine from a human in a text exchange. Passing it means the AI's conversational performance is indistinguishable from a human's — it says nothing definitive about consciousness, inner experience, or genuine understanding. This has been debated since the paper was published, and the Turing test is widely considered a performance benchmark, not a consciousness test.

---

Objective 2 — Effective Prompting (Weeks 3–6)

P7 (MC). A student asks an AI to help improve a cover letter without pasting the letter. A classmate pastes the actual letter and makes the same request. Which approach is better, and why?
- A. Asking without the letter is better — it forces the AI to generate more creative suggestions
- B. Both approaches are equally useful — the AI writes well either way
- C. Providing the actual letter is better — the AI works from real material rather than inventing generic advice ✅
- D. Pasting the cover letter is risky because the AI will automatically share it publicly

Feedback: Providing content (pasting the actual letter) is far more effective. The AI can now react to specific word choices, gaps in experience, and tone — instead of producing a generic cover-letter template. Option D conflates a privacy-awareness caution (free tools may store inputs; don't paste highly sensitive material) with a false claim that pasting automatically shares publicly — that is not how these tools work.

P8 (MC). A student's prompt reads: Summarize this report for my boss. [pastes 5-page report]. The AI returns a six-paragraph summary that includes conclusions not in the report. The BEST single fix is —
- A. Add "please" and ask again
- B. Use a Markdown structure with a Constraints section: separate the Task, the Report content, and a DO NOT add information not in the report constraint ✅
- C. Type the request in ALL CAPS to show emphasis
- D. Regenerate until the summary is shorter

Feedback: The two problems are no length constraint and no instruction not to add outside information. Option B addresses both with Markdown structure (separate Task and Report sections) and an explicit CAPS constraint. Politeness (A) gives no structural information. ALL CAPS alone (C) without structure and content separation doesn't fix the root problem. Regenerating (D) may produce another six-paragraph summary with different fabricated conclusions.

P9 (MC). Which prompt best demonstrates the meta-prompting technique?
- A. "Write the best possible prompt for creating a study plan for my history final."
- B. "I need a reusable prompt for creating a study plan. Ask me clarifying questions one at a time — when you have enough, return a Markdown template I can copy and reuse." ✅
- C. "You are a study-plan expert. Create a study plan for my history final."
- D. "Give me tips on how to write a better study-plan prompt."

Feedback: Meta-prompting means asking the AI to help BUILD the prompt — specifically by interviewing you one question at a time and then returning a reusable Markdown template. Option A asks the AI to skip the interview and guess at what you need — it goes straight to output. Option C assigns a role and asks for output directly — that's a structured prompt, not meta-prompting. Option D asks for tips about prompting — useful, but not the meta-prompting technique.

P10 (Multiple answer — select all that apply). Which of the following are TRUE about meta-prompting and structured prompts?
- A. Meta-prompting asks the AI to help design the prompt by posing clarifying questions one at a time ✅
- B. Asking for a Markdown-formatted prompt is purely cosmetic and has no practical value
- C. A prompt with contradictory instructions (such as targeting experts and beginners simultaneously) will often produce a muddled output ✅
- D. Assigning a Role such as "licensed attorney" makes the AI's legal information factually reliable
- E. Over-engineering — adding so many conflicting components that the prompt works worse — is a real failure mode ✅

Feedback: True: A (meta-prompting = interview + Markdown template), C (contradictions degrade output), E (over-engineering is a real risk). B is false — requesting Markdown output makes the components visible and reusable; it's practical, not cosmetic. D is false — assigning a Role shapes style and framing; it does not grant factual expertise. Legal content (or any high-stakes content) assigned a professional Role still needs verification.

P11 (MC). Which statement about zero-shot, one-shot, and few-shot prompting is CORRECT?
- A. All three produce the same quality output — the number of examples does not matter
- B. Few-shot means providing exactly one example before the task
- C. Zero-shot gives no examples; one-shot gives exactly one; few-shot gives several (typically two to five) to teach a pattern ✅
- D. Few-shot and regeneration are the same technique with different names

Feedback: Zero-shot = no examples. One-shot = exactly one. Few-shot = two to five. Option B is the most common error — "few" sounds small and students often assume it means one. But one example is one-shot, not few-shot. The distinction matters because few-shot examples allow the AI to generalize a pattern; a single example may or may not generalize. Regeneration (D) is a different technique entirely.

P12 (MC). A student wants to paste a colleague's performance review into an AI to help draft feedback. What is the BEST way to protect the colleague's privacy?
- A. Use a paid AI plan — paid plans automatically strip all personally identifying information
- B. Ask the AI to detect and remove any sensitive details before processing the document
- C. Replace names, job titles, and identifying details with placeholders (e.g., [NAME], [ROLE]) before pasting ✅
- D. Paste the full review — the AI keeps all inputs completely private by default

Feedback: You do the scrubbing before pasting, using placeholders. Option A is false — paid plans may have better data policies, but they do not automatically detect and strip all PII. Option B transfers the responsibility to the AI, which may miss sensitive details or infer them from context. Option D is false — most free AI tools may store and, on some settings, process your inputs; never assume complete privacy.

P13 (True / False). True or False: When an AI role-plays a conversation between two historical figures, the dialogue it generates is drawn directly from verified historical records and can be cited as accurate.
- False ✅

Feedback: False. AI-generated historical dialogue is generated, not transcribed. The AI can imitate the style of historical figures based on training data — it does not reproduce verified words from actual records. The specific words the AI generates never appeared in any verified historical source. Such dialogue may be useful for exploring ideas, but it must never be cited or presented as accurate history.

P14 (MC). A student prompts: Pretend you are a customer service rep. The AI plays a generic character that does not match the student's industry. The BEST fix is —
- A. Add "please" and repeat the same prompt
- B. Specify the role in detail: company type, product, customer type, and the kind of complaint (e.g., "Act as a customer service rep at a software company handling a billing dispute with a frustrated small-business owner") ✅
- C. Ask the AI to search for real customer service transcripts online
- D. Regenerate until the character is more realistic

Feedback: This is a specificity fix — the original prompt gives the AI almost nothing to work with. Adding company type, product, customer profile, and the specific situation type transforms a generic response into a genuinely useful rehearsal. Politeness (A), searching (C, which AI chatbots don't do by default), and regenerating (D) all fail to address the root cause: insufficient role specification.

P15 (MC). In a pre-mortem simulation, what is the AI being asked to help with?
- A. Celebrate the strengths of a completed project
- B. Predict with high accuracy exactly what will go wrong
- C. Imagine a project has already failed and reason backward to surface the most likely causes before it starts ✅
- D. Search for historical examples of similar projects that failed

Feedback: The pre-mortem works by imagining failure first, then reasoning backward — which is more effective at surfacing risks than asking "what might go wrong?" forward. This is a deliberate cognitive technique, not a prediction (B) or a historical search (D). It was developed by Gary Klein and popularized in applied planning research.

---

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

P16 (MC). When a student speaks a request to an AI assistant in voice mode, what happens FIRST before the AI processes the words?
- A. The AI listens to the audio and understands it directly — there is no intermediate step
- B. The AI searches the internet for audio matching the student's voice
- C. The speech is converted to text through a transcription step, which can introduce errors ✅
- D. The AI compresses and permanently stores the student's audio file

Feedback: Voice mode has two steps: (1) speech → text (transcription, which can introduce errors); (2) AI processes the text and generates a response. This is why a bad voice-mode response should prompt you to check the displayed transcript first — the AI may have received mis-heard text, not the words you intended. There is no direct audio understanding (A), no internet search (B), and no permanent audio storage (D).

P17 (MC). A student says: I uploaded my photo to DALL·E to find out what is in it. What is wrong with this statement?
- A. Nothing — DALL·E analyzes photos and generates accurate descriptions
- B. DALL·E is a text-to-image generation tool; it is not designed to analyze photos by taking an image as input ✅
- C. DALL·E only works with black-and-white images
- D. DALL·E can analyze images, but only ones it generated itself

Feedback: Direction matters. DALL·E's direction is text in, image out — it creates new images from text descriptions. It is not a photo-analysis tool. For image analysis (image in, text out — describing what's in a photo), use a multimodal chatbot with vision enabled (ChatGPT with vision, Claude, Gemini). Options C and D are false.

P18 (Matching). Match each AI tool or modality to its primary function.

AI tool or modality	Primary function
DALL·E / Midjourney	Generate new images from text descriptions
Whisper-class transcription	Convert spoken audio to a text transcript
ChatGPT / Claude / Gemini with vision	Analyze uploaded images or documents and answer questions about them
ElevenLabs	Generate realistic synthetic voices from text

Feedback: Each tool has a primary direction: DALL·E / Midjourney = text → image (creation, not analysis); Whisper-class transcription = audio → text (step 2 of the record-transcribe-analyze workflow); Multimodal chatbots with vision = image/document → text (analysis, extraction, Q&A); ElevenLabs = text → audio (voice synthesis from text). These are the verified, current primary use cases of each tool.

P19 (True / False). True or False: AI-powered transcription tools reliably produce a perfect, error-free text transcript from any audio recording.
- False ✅

Feedback: False. Transcription accuracy varies with audio quality, accents, background noise, crosstalk, and technical vocabulary. This is error-entry point 1 in the record-transcribe-analyze workflow. Always review the transcript before passing it to an AI for analysis — a mis-heard word or phrase in the transcript will flow into the AI's summary as if it were accurate.

P20 (Multiple answer — select all that apply). In the record-transcribe-analyze workflow, which of the following are places where errors can enter the final output?
- A. During the transcription step, when audio is converted to text and words may be mis-heard or dropped ✅
- B. During the playback step, when the student listens to the recording again
- C. During the AI's analysis step, when the AI may add details, smooth contradictions, or invent conclusions not in the transcript ✅
- D. During the file-naming step, when the student saves the audio recording
- E. During the recording step itself, if loud background noise reduces the audio quality fed into transcription

Feedback: The two AI-error-entry points in the workflow are A (transcription step: the audio-to-text conversion introduces mis-heard words and dropped phrases) and C (AI analysis step: the model can fabricate details not in the transcript). Options B, D, and E may affect audio quality or human understanding, but they are not AI error-entry points in the workflow. Recording quality (E) affects the transcript's accuracy but the error enters at the transcription step (A), not the recording step itself.

---

Answer key (quick reference)

Q	Answer	Q	Answer
P1	B (generates account, not real article)	P11	C (zero=none, one=one, few=two-five)
P2	C (chatbot=interface, LLM=engine inside)	P12	C (placeholder scrubbing before pasting)
P3	B (verify; fluency ≠ truth)	P13	False (generated, not verified records)
P4	Token→chunk / cutoff→knowledge date / hallucination→confident+wrong / sycophancy→agrees with user	P14	B (specify role, company, situation in detail)
P5	B (context window exceeded; early content lost)	P15	C (imagine failure, reason backward)
P6	False (behavioral benchmark; not consciousness)	P16	C (speech→transcription first; then AI processes text)
P7	C (providing content = real material)	P17	B (DALL-E creates; doesn't analyze photos)
P8	B (Markdown structure + Constraints)	P18	DALL-E/MJ→create images / Whisper→audio-to-text / ChatGPT/Claude/Gemini vision→analyze images+docs / ElevenLabs→synthetic voice
P9	B (clarify one at a time + Markdown template)	P19	False (transcription is not always accurate)
P10	A, C, E	P20	A, C

---

Canvas placement block

canvas_object             = Quizzes::Quiz
title                     = "Midterm Practice Exam (ungraded) -- Weeks 1-7"
module                    = "Week 8 -- Midterm Review & Exam"
assignment_group          = "Practice exercises"
points_possible           = 0
grading_type              = not_graded
available_from_offset_days = 0
due_offset_days           = 6
published                 = true
allowed_attempts          = 0             # unlimited attempts
show_correct_answers      = true          # feedback after submission
shuffle_answers           = true
ai_permitted              = true          # studying can use AI; the live midterm cannot
provenance                = "~ Prof. Quinn's edition · Fall 2026 · built with thecoursemaker.com"

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