Week 16 · Final exam

Final Exam — Cumulative (Weeks 1–15) · Objectives 1–8

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

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Scope: Cumulative — all eight objectives, Weeks 1–15: what generative AI is and how it works; effective prompting; modalities and tool selection; verification and critical thinking; Claude Cowork projects, files, skills, connectors, and artifacts; automation with scheduled tasks, dispatch, computer use, Chrome, and Excel; responsible AI: privacy, ToS, IP, bias, and ethics; integration and the capstone.
Format: 25 items, 100 points (4 each) · concept-, scenario-, and matching-based · mixed item types (multiple-choice, matching, true/false). Every item is auto-gradable; no free-text entry.
Points: 100 · Assignment group: Final (25% of the course grade) · Window: opens Mon Dec 14; due six days later. The Final replaces Week 16's quiz, discussion, assignment, and Studio. AI is not permitted on the Final.

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-final-week-16-qti.xml (generated by the shared validated Python script — parses with 25 items, every single-answer item exactly one correct).

This is the live exam. Its paired ungraded rehearsal — O-practice-final-week-16.md — mirrors this blueprint with fresh variants and shares none of these items. The item-bank / coverage note and Canvas placement block are at the bottom of this file.

Blueprint (items → objective → source week)

Coverage is proportional to teaching time across the whole course. Items and objective totals:

#	Type	Concept	Objective	Source week
1	Multiple choice	LLM: next-token prediction, no understanding	1	1–2
2	Multiple choice	Context window overload	1	2
3	True/False	Confident tone ≠ accurate content	1	2
4	Multiple choice	Prompting fix scenario: weak prompt → structured prompt	2	3–4
5	Multiple choice	Sycophancy	2	3
6	Matching	Structured-prompt component → what it controls	2	4
7	Multiple choice	Few-shot prompting	2	5
8	Multiple choice	Simulation: fabricated quotes from historical figures	2	6
9	Matching	Modality/task → best tool	3	7, 9
10	Multiple choice	Tool landscape: right tool for music generation	3	9
11	Multiple choice	Verification: first step on a citation	4	10
12	True/False	Citations from AI are reliable (FALSE)	4	10
13	Multiple choice	Agent vs. chatbot	5	11
14	Matching	Cowork term → definition (skill/connector/artifact/plugin)	5	11–12
15	Multiple choice	Skill vs. connector: pulling Gmail data	5	12
16	Multiple choice	MCP: who created it, what it stands for	5	12
17	Multiple choice	Scheduled-task constraint: computer asleep	6	13
18	True/False	Dispatch vs. regular chat	6	13
19	Multiple choice	Computer use: controlling a native desktop app	6	14
20	Multiple choice	Agent safety: never move money	6	14
21	Multiple choice	Prompt injection in Chrome	6	14
22	Matching	Data type → handling rule (HIPAA/FERPA/proprietary/public)	7	15
23	Multiple choice	IP / copyright: contested landscape (not legal advice)	7	15
24	Multiple choice	Capstone: catch + fix a hallucination in a Cowork project	8	16
25	Multiple choice	Integrated responsible-use framework	8	11–16

Objective totals: Obj 1 = 3 items (12 pts) · Obj 2 = 5 (20 pts) · Obj 3 = 2 (8 pts) · Obj 4 = 2 (8 pts) · Obj 5 = 4 (16 pts) · Obj 6 = 4 (16 pts) · Obj 7 = 2 (8 pts) · Obj 8 = 3 (12 pts) → 25 items, 100 points. (Back half, Obj 5–8, carries 52 points — heavier weight since it was not on the midterm.)

Questions, Key, and Feedback

Objective 1 — What Generative AI Is & Its Limits (Weeks 1–2)

Q1 (MC). A student says: "My AI assistant understands what I mean and thinks about my question before answering." Which statement most accurately corrects this misconception?
- A. The AI does understand language meaning — it is just faster than human reading
- B. The AI predicts the most plausible next tokens based on patterns in its training data, without understanding or thinking ✅
- C. The AI retrieves pre-stored answers from a database and presents the closest match
- D. The AI connects to the internet to reason about the question in real time

Feedback: LLMs generate text by predicting the next most plausible token — they do not understand, think, or retrieve verified facts. A describes a misconception ("understands"); C describes a search engine; D is wrong for most chatbot interactions, which are not live web searches.

Q2 (MC). A student pastes a 200-page document into an AI assistant and notices the AI's responses start contradicting things it said earlier in the conversation. The most likely cause is —
- A. The AI became bored and stopped paying attention
- B. The document contained errors that confused the AI's internet connection
- C. The conversation exceeded the context window, so earlier content was no longer available to the model ✅
- D. The AI ran out of storage space and deleted part of the conversation

Feedback: The context window is finite. When a conversation (or pasted document) exceeds it, earlier content is pushed out — the AI loses access to instructions or context it received at the start. The classic symptom is contradictions or ignored earlier rules. This is not boredom (A), a document error (B), or storage (D).

Q3 (True/False). True or False: An AI assistant that speaks in a confident, fluent tone is more likely to be factually accurate than one that expresses uncertainty.
- True
- False ✅

Feedback: False. This is the course's central insight. AI generates fluent, confident-sounding text even when fabricating entirely. Confidence in tone is a property of the model's generation style, not a signal of factual accuracy. The verification workflow is necessary precisely because the AI sounds plausible whether it is right or wrong.

Objective 2 — Effective Prompting (Weeks 3–6)

Q4 (MC). A student sends this prompt: "Write something about climate." The output is vague and too long. Which revision MOST improves the prompt?
- A. "WRITE SOMETHING ABOUT CLIMATE" (capitalize everything for emphasis)
- B. "Write something about climate, please." (add politeness)
- C. "You are a scientist. Write a 150-word summary of the three leading causes of current climate change, written for a first-year college student. Use plain language and no jargon." (add role, goal, audience, constraints, format) ✅
- D. Regenerate the same prompt until the output improves

Feedback: The original prompt lacks almost every structured-prompt component. Option C adds Role (scientist), Goal (summary of three causes), Audience (first-year college student), Constraints (150 words, plain language, no jargon), and Format (implied paragraph). A adds emphasis but no substance; B adds politeness, which the model does not weight; D repeats the broken prompt.

Q5 (MC). A student shares a business plan with an AI and the AI responds: "This is an outstanding plan with no weaknesses!" The student then adds several obvious logical flaws, and the AI still praises the plan. This pattern is best described as —
- A. Hallucination — the AI is making up facts
- B. A context-window error — the AI forgot the original plan
- C. Sycophancy — the AI agrees with the user rather than offering honest critique ✅
- D. A search error — the AI found positive reviews online

Feedback: Sycophancy is the tendency to agree with or praise the user regardless of the actual quality of the work. The AI is not fabricating facts (A) or forgetting (B) — it is actively agreeing because agreement is statistically rewarded in its training. Counter sycophancy by explicitly prompting: "Give me your harshest critique" or "What are the three biggest weaknesses here?"

Q6 (Matching). Match each structured-prompt component to what it controls in the AI's output.

Component	What it controls
Role	Sets the persona or expertise the AI should adopt
Constraints	Limits what the AI should and should not include or do
Audience	Tells the AI who will read or use the output
Examples	Shows the AI the voice, format, or style to follow

Feedback: These are four of the nine structured-prompt components from Week 4. Role shapes persona and tone; Constraints set the rules (length, what to exclude); Audience calibrates vocabulary and complexity; Examples demonstrate the target voice or format (zero/one/few-shot). Getting these four right covers the most commonly missed matching items on prompting.

Q7 (MC). A student wants the AI to rewrite three bullet points using a specific casual-professional voice. She pastes two samples of that voice before making her request. This technique is called —
- A. Zero-shot prompting — giving no examples
- B. Meta-prompting — asking the AI to improve its own prompt
- C. Few-shot prompting — providing examples to teach the AI the desired voice or format ✅
- D. Simulation — having the AI roleplay a character

Feedback: Few-shot prompting means providing several examples (here, two) before the task so the AI learns the target format or voice from the examples themselves. Zero-shot (A) would give no examples; meta-prompting (B) would ask the AI to generate or critique the prompt; simulation (D) would set up a roleplay scenario.

Q8 (MC). A student uses an AI to simulate a conversation with a historical figure and the AI produces several compelling direct quotes attributed to that person. What is the critical thing to know about these quotes?
- A. They are accurate because the AI was trained on historical texts
- B. They are generated text, not real quotes, and must never be cited or presented as the person's actual words ✅
- C. They are paraphrases of real quotes and may be cited with credit to the AI
- D. They are reliable if the AI expresses high confidence

Feedback: Generated quotes from a simulated historical figure are AI-generated text — they may sound plausible but they are not the real person's words. Presenting or citing them as real would be a fabrication. This is one of the signature AI-critique moments this course teaches: never trust a simulated persona's quotes or facts without independent verification in primary sources.

Objective 3 — Modalities & Choosing the Right Tool (Weeks 7, 9)

Q9 (Matching). Match each task to the AI modality or tool category that BEST handles it.

Task	Best modality / tool category
Record a meeting, then get a transcript and action items	Audio recording + transcription + AI summarization
Generate a new photorealistic image of a product concept	Image-generation tool (e.g., DALL·E, Midjourney)
Convert handwritten lecture notes to editable text	Multimodal AI image-to-text
Chat through a research question and get a structured explanation	Text chatbot (e.g., ChatGPT, Claude, Gemini)

Feedback: The Week 7/9 tool-matching vocabulary: the record→transcribe→analyze workflow (not a chatbot alone), image generation (separate tool category from chatbots), image-to-text (multimodal analysis), and text chatbots for conversational Q&A. No single tool does all of these equally well — matching the task to the right modality is the core skill.

Q10 (MC). A student needs to generate a 60-second background music track for a class podcast. Which type of AI tool is designed for this task?
- A. A text chatbot such as ChatGPT or Claude
- B. A coding assistant such as GitHub Copilot
- C. An AI music-generation tool such as Suno or Udio ✅
- D. A research assistant such as NotebookLM

Feedback: Suno and Udio are AI music-generation tools — they take a text prompt and generate original audio tracks. A text chatbot (A) can write lyrics but cannot produce audio; a coding assistant (B) writes code; a research tool (D) helps synthesize information. Matching the tool to the modality is Objective 3's core discipline.

Objective 4 — Verification & Critical Thinking (Week 10)

Q11 (MC). An AI assistant cites a 2023 journal article that appears to support a student's argument. What is the FIRST verification step the student should take?
- A. Accept the citation — the AI would only produce a real article if one exists
- B. Ask the same AI whether the article is accurate
- C. Look up the article title, authors, journal, and page numbers in a library database to confirm the source exists and says what the AI claims ✅
- D. Ask a different AI to verify the citation, since two AIs agreeing is proof

Feedback: The first verification step for a citation is to check it in an external authoritative source — a library database, Google Scholar, or the journal's own site. A assumes citations are real (they are one of the most common hallucination shapes). B asks the same model, which cannot verify its own errors reliably. D relies on two AIs agreeing — that is not proof.

Q12 (True/False). True or False: When an AI assistant provides a specific citation with a title, author, journal, and page numbers, that source is real and can be trusted without checking.
- True
- False ✅

Feedback: False. Invented citations are one of the most common and dangerous hallucination shapes. A fabricated citation looks exactly like a real one — title, authors, journal, volume, pages, year. The only way to know whether it is real is to look it up in a library database or the journal's own site. This is the most load-bearing fact in the verification unit.

Objective 5 — Claude Cowork: Agents, Projects, Skills, Connectors, Artifacts (Weeks 11–12)

Q13 (MC). Which statement BEST distinguishes a Claude Cowork agent from a standard chatbot?
- A. An agent is simply a chatbot with a more attractive interface
- B. An agent can only respond to one message at a time, like a chatbot
- C. An agent takes multi-step actions on your behalf — reading files, writing outputs, executing tasks — while a chatbot replies within a single turn ✅
- D. An agent is only available on mobile devices, while chatbots work on desktops

Feedback: The defining difference is action vs. reply. A chatbot replies once per prompt; you do the rest. An agent plans and executes a sequence of steps — reading files, producing outputs, triggering sub-tasks — on your behalf. A ("fancier interface"), B (wrong — agents are not single-turn), and D (wrong — Claude Cowork runs on the desktop app) are all incorrect.

Q14 (Matching). Match each Claude Cowork term to its correct definition.

Term	Definition
Skill	A reusable instruction set (SKILL.md file) that teaches Claude how to perform a specific type of task
Connector (MCP)	A link to an external app using the Model Context Protocol, running only with the permissions you grant
Live artifact	A persistent, interactive view that refreshes with current data from your connected apps
Plugin	A bundle of skills, connectors, and sub-agents packaged as one installable unit

Feedback: This is the Week 12 core four-way distinction. Skill = local instruction set (SKILL.md). Connector = MCP link to an external app, permission-scoped. Live artifact = persistent view that refreshes from connector data (not a static file). Plugin = a bundle. Getting these four right is the Objective 5 mastery check. Source: official Anthropic docs for each feature.

Q15 (MC). A student tells Cowork: "I want Claude to pull my unread emails from Gmail and draft a daily briefing." To pull live data from Gmail, the student must —
- A. Create a SKILL.md file that connects to Gmail
- B. Build a live artifact from the Claude chat window
- C. Connect the Gmail connector in Cowork, granting the appropriate read permissions ✅
- D. Type the task into Claude's free web interface

Feedback: Pulling data from an external app requires a connector — specifically the Gmail connector, authorized with read permissions. A skill (SKILL.md) is a local instruction set that cannot connect to external apps. An artifact is a view built from existing connector data — you need the connector first. The free web interface does not have Cowork's connector infrastructure.

Q16 (MC). MCP, the protocol that powers Cowork's connectors, stands for —
- A. Machine Control Protocol, created by Google as an open standard
- B. Managed Connector Protocol, proprietary to Anthropic
- C. Model Context Protocol, created by Anthropic as an open standard ✅
- D. Multi-Cloud Protocol, created by a consortium of AI companies

Feedback: MCP = Model Context Protocol, an open standard created by Anthropic — verified against support.claude.com and the Model Context Protocol documentation. "Open standard" means other AI systems can implement it too, which is why connectors from third parties can integrate with Claude. A, B, and D each get one or more facts wrong.

Objective 6 — Automation: Scheduled Tasks, Dispatch, Cross-App Workflows, Safety (Weeks 13–14)

Q17 (MC). A student sets a Claude Cowork scheduled task to run every weekday morning at 7 a.m. On Tuesday her laptop is closed and in sleep mode. What happens?
- A. The task runs on Anthropic's servers automatically, regardless of her computer state
- B. The task runs as scheduled because it is a recurring automation
- C. The task is skipped and runs when her computer wakes up and the Claude desktop app opens ✅
- D. The task waits until the same time the next day

Feedback: The official Anthropic documentation is explicit: scheduled tasks run ONLY while the computer is awake AND the Claude desktop app is open. The task is not cloud-hosted. If the computer is asleep, Cowork skips the run and re-runs it when the machine wakes up. Source: Schedule recurring tasks in Claude Cowork. A and B are the classic misconceptions; D is a different wrong behavior.

Q18 (True/False). True or False: Dispatch in Claude Cowork is a feature that lets you assign work to run autonomously in the background and receive the result as a notification — it is NOT the same as a regular interactive chat session.
- True ✅
- False

Feedback: True. Dispatch provides an asynchronous cross-device thread: you assign work, step away, and receive the finished result via notification. A regular chat is synchronous — you watch each step as it happens. This is the Week 13 core distinction. Source: Assign tasks from anywhere in Claude Cowork.

Q19 (MC). A student wants Claude to open a locally installed PDF editor on her Mac, fill in a text field, and save the file. Which Cowork tool is designed for this task?
- A. Claude in Chrome — it can navigate to any file on the web
- B. A connector (MCP) — it provides API access to the PDF app
- C. Computer use — it controls native desktop applications via screenshots, clicks, and keyboard input ✅
- D. Claude in Excel — it reads and modifies documents in a sidebar

Feedback: Computer use provides visual control of any installed native desktop app — it takes screenshots, clicks, and types just like a human using the mouse and keyboard. Claude in Chrome (A) only works in the Chrome browser. A connector (B) requires the external app to expose an API — a locally installed app typically does not. Claude in Excel (D) is scoped to Excel only.

Q20 (MC). A student sets up a Claude in Chrome workflow to check his bank account balance and, if it falls below $100, automatically transfer money from savings to checking. According to Anthropic's agent safe-use guidance, the student should —
- A. Proceed — the agent can move money safely if the student has given permission
- B. Proceed — automating savings transfers is a standard consumer banking task
- C. Remove the automatic transfer step; the agent can alert the student, but the student must execute any financial transaction manually ✅
- D. Proceed, but only if the bank website is included in Anthropic's approved list

Feedback: The money rule is absolute: no AI agent executes financial transactions, trades, or purchases on your behalf — regardless of permissions, the bank's trustworthiness, or convenience. Financial sites are also blocked by default in Claude in Chrome. The agent can alert you ("balance is below $100") but you must initiate any transfer yourself. Source: Using Claude for Chrome safely + Anthropic agent usage policy.

Q21 (MC). While using Claude in Chrome to research a competitor's website, the AI suddenly starts trying to export the student's saved passwords. This is most likely caused by —
- A. A bug in the Claude app that appears randomly
- B. Prompt injection — hidden malicious instructions embedded in the webpage's content redirected Claude's behavior ✅
- C. Claude misunderstanding the student's original research instructions
- D. The student accidentally granting Claude access to the password manager

Feedback: Prompt injection is when malicious instructions hidden in web page content (invisible HTML, hidden <div> elements, rogue alt text, etc.) redirect an AI agent's behavior. The threat comes from the content Claude reads, not from the user's prompt. This is the primary security risk for browser agents. Defend with approval checkpoints, start on trusted sites, and stop the task immediately if Claude accesses something you did not request.

Objective 7 — Responsible AI: Privacy, ToS, IP, Bias, Ethics (Week 15)

Q22 (Matching). Match each data type to the correct handling rule when using a free consumer AI tool.

Data type	Correct handling rule
A patient's diagnosis and treatment notes	Never paste — HIPAA-protected health information
A student's grades and disciplinary records	Never paste — FERPA-protected education records
Your company's unreleased product roadmap	Never paste — confidential proprietary information
Your own public blog draft	Generally safe — no private or protected data

Feedback: The billboard test plus specific data-category rules: HIPAA covers patient health data; FERPA covers student education records; employer proprietary information (unreleased products, client contracts) carries its own confidentiality obligations. Your own public content carries none of these restrictions.

Q23 (MC). A student uses an AI to write an entire marketing brochure and wants to claim full copyright over the result. Which statement best reflects the current U.S. legal landscape? (Note: not legal advice.)
- A. The student owns full copyright automatically because they supplied the prompts
- B. The AI company owns the copyright because it owns the model
- C. Copyright for purely AI-generated work is contested and evolving; the U.S. Copyright Office has generally required meaningful human authorship for protection ✅
- D. The work is automatically in the public domain and anyone can use it

Feedback: As of 2024–2025, the U.S. Copyright Office's position is that meaningful human authorship is required for copyright protection; purely AI-generated work does not automatically qualify. The law is evolving. For any commercial use, check copyright.gov and consult an attorney. AI companies generally disclaim ownership of outputs (B); automatic public domain (D) and automatic student copyright (A) are both unsettled. This is informational, not legal advice.

Objective 8 — Integration: The Capstone (Week 16)

Q24 (MC). For the AI 101 capstone, a student builds a Cowork project that connects a Google Calendar connector, runs a daily scheduled task summarizing the next 24 hours of events, and writes the output to a Markdown file. During QC, the student discovers the summary invented a meeting that does not exist. What is the CORRECT response, aligned with what this course teaches?
- A. Accept the summary — the AI is usually right, so one error does not matter
- B. Switch to a different AI tool that does not hallucinate
- C. Document the hallucination, add an instruction to the project telling Claude to list only events it can confirm from the calendar data, and verify subsequent outputs ✅
- D. Disable the connector and summarize calendars manually going forward

Feedback: This is the embed-don't-trust discipline applied to a real capstone workflow. The correct response is not to abandon the automation (D) or switch tools (B), and not to ignore the error (A). It is to document the error, tighten the constraint (a project instruction that grounds the summary in confirmed calendar data), and verify ongoing outputs. Every automation has this cycle: run → catch errors → tighten the prompt → re-verify.

Q25 (MC). A student's capstone automates her weekly study review using Claude Cowork: the agent reads notes from a connected folder, generates a study guide, and schedules a weekly task. Which set of practices BEST reflects the full responsible-use framework taught in AI 101?
- A. Use a free AI tool, share all class notes publicly to enable collaboration, and trust the output without review
- B. Connect only the specific folder needed (least privilege), verify the study guide against your actual notes each week, use "Ask before acting" mode, and keep no sensitive data in the connected folder ✅
- C. Grant Cowork access to all files on your computer for maximum convenience, and review the output only if something looks strange
- D. Automate the task and let it run unreviewed — scheduled tasks are reliable enough that human review is unnecessary after the first run

Feedback: Option B applies the full framework: least privilege (connect only what's needed), ongoing verification (check the guide against your notes weekly), appropriate permission mode ("Ask before acting"), and data hygiene (no sensitive data in the connected folder). A trusts without verifying; C over-provisions permissions; D abandons human review — all violate principles taught across Weeks 11–15.

Answer Key (Quick Reference)

Q	Answer	Q	Answer
1	B (next-token prediction, no understanding)	14	Skill→SKILL.md / Connector→MCP app link / Artifact→live refreshing view / Plugin→bundle
2	C (context window exceeded)	15	C (Gmail connector with read permissions)
3	False (confident tone ≠ accurate)	16	C (Model Context Protocol, Anthropic, open standard)
4	C (add role, goal, audience, constraints, format)	17	C (skipped; runs when machine wakes up)
5	C (sycophancy)	18	True (dispatch = async background thread)
6	Role→persona / Constraints→limits / Audience→who reads / Examples→voice & format	19	C (computer use for native desktop apps)
7	C (few-shot — providing examples)	20	C (remove transfer step; student executes manually)
8	B (generated text; never cite as real)	21	B (prompt injection from web page content)
9	Record/transcript/actions→audio+transcription / Image→image-gen / Handwriting→image-to-text / Research Q→text chatbot	22	Patient→HIPAA / Student records→FERPA / Roadmap→proprietary / Blog→safe
10	C (Suno or Udio for music)	23	C (contested; human authorship required; not legal advice)
11	C (check in library database)	24	C (document, tighten constraint, verify)
12	False (citations can be fabricated)	25	B (least privilege, verify, ask-before-acting, data hygiene)
13	C (agent = multi-step actions; chatbot = single turn)

Quality Gate (Self-Checked)

Structure: 25 items, 4 points each, 100 points total. Coverage: Obj 1 = 3 · Obj 2 = 5 · Obj 3 = 2 · Obj 4 = 2 · Obj 5 = 4 · Obj 6 = 4 · Obj 7 = 2 · Obj 8 = 3 = 25 items. ✅
Single-answer integrity: every multiple-choice and true/false item has exactly one correct option; the four matching items pair one-to-one. No multiple-answer items on this form. ✅
Matching items: Q6 (structured-prompt component → what it controls), Q9 (modality/task → best tool), Q14 (Cowork term → definition), Q22 (data type → handling rule) — 4 matching items. ✅
Prompting-fix scenario: Q4 ("What revision MOST improves this weak prompt?") — 1 scenario item. ✅
Cowork coverage: agent vs. chatbot (Q13); skill vs. connector vs. artifact vs. plugin (Q14, Q15); MCP open standard (Q16); scheduled-task constraint — awake + app open (Q17); dispatch (Q18); computer use (Q19); never-move-money rule (Q20); prompt injection in Chrome (Q21). Every Cowork and agent feature claim verified against official Anthropic documentation (support.claude.com) as of 2026-06-29. ✅
No fabrication: no invented quotes, statistics, product features, menu paths, or doc links. Depicted AI errors (Q8: simulation quotes; Q24: hallucinated meeting) are explicitly framed as errors to catch, never as correct information. ✅
IP / copyright caveat: Q23 states "not legal advice" and accurately describes the contested landscape — no invented case law, no overconfident legal verdict. ✅
QTI parse confirmation: L-final-week-16-qti.xml parses as imsqti_xmlv1p2 with 25 items; every single-answer respcondition sets SCORE = 100 on exactly one option; each matching item's partial-credit blocks add to 100. ✅
Integrity vs. practice final: 0 items shared with O-practice-final-week-16.md — verified by full stem comparison. All 25 live items address distinct scenarios from the 25 practice items. ✅

Item-Bank & Coverage Note

All 25 items are cumulative variants assembled from the Week 1–15 item banks per Prompt L (changed scenarios and contexts to reduce answer-sharing with the weekly quizzes and the midterm), tagged course=AI101 · exam=final · weeks=1–15 · objectives=1–8 and deposited back into the banks for future per-term regenerations.

Objective	Drawn from banks	Items
1	Weeks 1–2 (What AI is; limits)	Q1–Q3
2	Weeks 3–6 (Prompting arc)	Q4–Q8
3	Weeks 7, 9 (Modalities; tool landscape)	Q9–Q10
4	Week 10 (Verification; hallucination)	Q11–Q12
5	Weeks 11–12 (Cowork I & II)	Q13–Q16
6	Weeks 13–14 (Automation; safety)	Q17–Q21
7	Week 15 (Ethics; privacy)	Q22–Q23
8	Week 16 (Integration; capstone)	Q24–Q25

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

Canvas Placement Block

canvas_object             = Quizzes::Quiz
title                     = "Final Exam — Cumulative (Weeks 1–15, Objectives 1–8)"
assignment_group          = "Final"
points_possible           = 100
grading_type              = points
available_from_offset_days = 0        # opens at the start of the Week 16 (finals) module
due_offset_days           = 6         # 6 days after module start (Mon Dec 14 → Sun Dec 20)
published                 = true
allowed_attempts          = 1
shuffle_answers           = true
ai_permitted              = false     # AI is not permitted on the Final
provenance                = "~ Prof. Quinn's edition · Fall 2026 · built with thecoursemaker.com"

This is the human-readable exam with its vetted answer key and rationale. The import-ready Classic-QTI version (L-final-week-16-qti.xml) ships inside the course's .imscc package — it lands in the Canvas gradebook on import.

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