← Back to the Using Artificial Intelligence outline The Course Maker
← Using Artificial Intelligence outline
Week 16 · Practice final

Practice Final Exam (ungraded) · 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
What this is: a low-stakes rehearsal for the cumulative Final. It mirrors the real exam's blueprint — the same eight objectives, the same item-type mix, the same length — but is built from fresh item-bank variants and shares none of the live Final's questions.
Settings: ungraded (0 points) · up to 3 attempts · feedback shown after submission · opens before the exam window so you can prepare. (Practice; AI is not permitted on the real Final.)

This is the human-readable practice exam with its vetted answer key and feedback (released after submission). The import-ready Classic QTI 1.2 is in O-practice-final-week-16-qti.xml (generated by the shared validated Python script — parses with 25 items, every single-answer item exactly one correct). The Canvas placement block is at the bottom.

Integrity note for students. Every item here is a fresh variant — new scenarios, new wording — with a pre-vetted answer. None of these are the live Final questions. Working them builds the skill the Final tests, honestly. The paired live exam is L-final-week-16.md.

Blueprint (mirrors the Final)

Coverage matches the real exam: Obj 1 = 3 · Obj 2 = 5 · Obj 3 = 2 · Obj 4 = 2 · Obj 5 = 4 · Obj 6 = 4 · Obj 7 = 2 · Obj 8 = 3. (The actual Final items are not listed here — only the shared structure.)

# Type Concept Objective Source week
1 Multiple choice LLM text generation: next-token prediction 1 1–2
2 Multiple choice Training cutoff: AI doesn't know recent events 1 2
3 True/False AI chatbot vs. search engine (same? FALSE) 1 1–2
4 Multiple choice Emphasis technique: Markdown bold 2 3
5 Multiple choice Meta-prompting: asking the AI to help write the prompt 2 4
6 Matching AI tool → best use (tool landscape) 3 9
7 Multiple choice Few-shot: three examples before the task 2 5
8 Multiple choice Simulation type: difficult-conversation practice 2 6
9 Multiple choice Record→Transcribe→Analyze order 3 7
10 Multiple choice Citation hallucination 4 10
11 True/False Asking the same AI to verify itself (NOT reliable) 4 10
12 Multiple choice Cowork project definition 5 11
13 Multiple choice Connected folder: what it means 5 11
14 Matching Agent tool → surface controlled (computer use/Chrome/Excel/connector) 6 14
15 Multiple choice Plugin: bundle of skills + connectors + sub-agents 5 12
16 Multiple choice Live artifact definition 5 12
17 Multiple choice /schedule command: what it does 6 13
18 Multiple choice Never automate money movement 6 13–14
19 True/False Prompt injection (malicious web content → redirects Claude) 6 14
20 Multiple choice Approval checkpoint: before irreversible action 6 14
21 Multiple choice Billboard test 7 15
22 Multiple choice AI bias (not neutral because training data) 7 15
23 Multiple choice Troubleshooting: context window overload → fresh conversation 7 15
24 Multiple choice Capstone: catch + fix a hallucination in a briefing workflow 8 16
25 Multiple choice Core course discipline: embed-don't-trust 8 All

Objective totals: 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 (ungraded; mirrors the 100-point Final's emphasis).

Questions, Key, and Feedback

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

P1 (MC). Which of the following MOST accurately describes how a large language model (LLM) generates text?
- A. It retrieves pre-written answers from a verified database
- B. It reasons through a problem the way a human expert does
- C. It predicts the next most plausible token based on statistical patterns learned during training ✅
- D. It searches the web in real time and summarizes the top results

Feedback: Next-token prediction from training-data patterns — not retrieval (A), not reasoning (B), not live search (D). The model generates text; it does not look anything up.

P2 (MC). A student asks her AI assistant about the winner of a major sports championship held two months ago, but the AI gives incorrect information confidently. The most likely reason is —
- A. The AI dislikes sports and intentionally gives wrong answers on that topic
- B. The AI's training data has a cutoff date and it may not have information about recent events ✅
- C. The AI connected to a sports database that contained an error
- D. The student should have asked in a different language for better accuracy

Feedback: LLMs are trained on data up to a cutoff date — they may not know about events after that date, and can hallucinate plausible-sounding information about them. Language (D), intention (A), and a live database (C) are not the issue.

P3 (True/False). True or False: A traditional search engine and an AI chatbot work the same way — both look up facts in a live indexed database and return verified results.
- True
- False ✅

Feedback: False. A search engine indexes and retrieves real documents. An AI chatbot generates text based on statistical patterns in training data — it does not look anything up in a live database, and its outputs are not verified.

Objective 2 — Effective Prompting (Weeks 3–6)

P4 (MC). A student wants the AI to focus only on the three key points she has bolded in her message and ignore all other text. The prompting technique she is using to signal importance is —
- A. Few-shot prompting — she is giving examples of what matters
- B. Meta-prompting — she is asking the AI to fix its own prompt
- C. Emphasis using Markdown formatting (bold) to direct the AI's attention ✅
- D. A simulation — she is roleplaying as a writing coach

Feedback: Emphasis with Markdown (bold, headers, lists), XML-style tags, or CAPS for must-dos is a prompting technique that signals to the AI what to focus on. Few-shot (A) means providing task examples; meta-prompting (B) means asking the AI to help write a prompt; a simulation (D) is a roleplay scenario.

P5 (MC). A student types: "I need a prompt that will help me write a personal statement for grad school. Ask me clarifying questions one at a time, then return a structured Markdown prompt." This technique is called —
- A. Zero-shot prompting — giving no prior examples
- B. Meta-prompting — using the AI to help generate or improve a prompt ✅
- C. Simulation — roleplaying a graduate admissions scenario
- D. Providing content — pasting a document for the AI to analyze

Feedback: Meta-prompting is asking the AI to help you write or improve a prompt. "Ask me clarifying questions one at a time, then return a structured prompt" is the classic meta-prompting formula. Not zero-shot (no examples weren't the point), not simulation (no persona), not content-providing (no document pasted).

P7 (MC). A student gives an AI three examples of correctly anonymized paragraphs before asking it to anonymize a new paragraph. This is an example of —
- A. Zero-shot prompting — no examples were given
- B. One-shot prompting — exactly one example was given
- C. Few-shot prompting — several examples were provided to guide the output ✅
- D. Meta-prompting — the AI is being asked to prompt itself

Feedback: Few-shot = several examples (here, three). Zero-shot = no examples; one-shot = exactly one; meta-prompting = asking the AI to help write the prompt. A classic misconception is "few-shot means exactly one example" — it means a few (two or more).

P8 (MC). A project manager uses an AI to simulate a difficult client who challenges every deliverable, in order to practice responding under pressure before a real presentation. Which type of simulation is this?
- A. A pre-mortem — imagining what could go wrong
- B. A decision role-play — testing different strategic choices
- C. A difficult-conversation simulation — practicing a challenging interpersonal scenario ✅
- D. An adaptive-tutor simulation — being taught a concept step by step

Feedback: The four simulation types from Week 6: difficult-conversation (practice a hard interaction), pre-mortem (imagine what could go wrong before a project), decision role-play (test different strategic choices), and adaptive tutor (AI teaches you something step by step). This scenario is a difficult-conversation simulation.

P6 (Matching). (Objective 3) Match each AI tool category to the task it is BEST suited for.

AI tool / category Task best suited for
NotebookLM (Google) Research: ask questions grounded in a set of uploaded documents
Midjourney or DALL·E Generate a photorealistic image from a text description
ElevenLabs Generate a realistic voice narration from written text
Sora (OpenAI) Generate a short video clip from a text prompt

Feedback: NotebookLM is a research tool — it grounds answers in documents you upload (not general web search). Midjourney/DALL·E are image-generation tools. ElevenLabs is a voice/audio generation tool (text-to-speech). Sora is a video generation tool. Matching tool to modality is the Obj 3 core skill.

P9 (MC). A student records a 10-minute group discussion on her phone, then runs the audio through a transcription tool, and finally asks a chatbot to extract the three main decisions made. In the correct workflow order, the three steps are —
- A. Analyze → Transcribe → Record
- B. Transcribe → Record → Analyze
- C. Record → Analyze → Transcribe
- D. Record → Transcribe → Analyze ✅

Feedback: The workflow is always Record → Transcribe → Analyze — in that exact order. You must have the audio before you can transcribe; you must have the transcript before the AI can analyze it for decisions. Catch transcription errors between steps 1 and 2; catch summary fabrications between steps 2 and 3.

P10 (MC). A student asks an AI to cite three peer-reviewed studies supporting a claim. The AI returns three plausible-sounding citations with real-looking author names and journal names. When the student checks, none of the studies exist. This is an example of —
- A. Sycophancy — the AI is agreeing with the student
- B. A search engine error — the AI retrieved broken links
- C. Citation hallucination — the AI fabricated plausible-sounding but nonexistent sources ✅
- D. A context-window error — the AI forgot what sources it had seen

Feedback: Citation hallucination — inventing plausible-sounding but nonexistent sources — is the most common and most dangerous hallucination shape. The AI generates what a real citation looks like, then fills in convincing-sounding details. Always verify citations in a library database before using them.

P11 (True/False). True or False: Asking the same AI model to double-check its own answer is a fully reliable verification method because the model can identify all of its own errors.
- True
- False ✅

Feedback: False. The model has the same biases and patterns that produced the error — it cannot reliably detect its own hallucinations. Cross-check externally: library databases, authoritative websites, or a second independent source. Two AIs agreeing is also not proof — both may have learned the same incorrect pattern.

P12 (MC). In Claude Cowork, a "project" is best defined as —
- A. A single chat conversation that ends when the window is closed
- B. A public workspace shared with all Cowork users
- C. A persistent, self-contained workspace with its own files, instructions, and memory that carries context across multiple tasks ✅
- D. A scheduled task that runs automatically every morning

Feedback: A Cowork project is a persistent workspace — it persists across sessions, stores its own files and instructions, and carries memory from one task to the next within that project. Not a single chat (A), not public (B), not a scheduled task (D). Source: Organize your tasks with projects in Claude Cowork.

P13 (MC). When a student connects a local folder in Claude Cowork, this means —
- A. The folder is uploaded permanently to Anthropic's cloud servers
- B. Claude can read files from and write files to that folder on the student's computer ✅
- C. The folder is shared with other Cowork users in the same organization
- D. Claude can only view the files, not edit or create new ones

Feedback: A connected folder gives Claude read and write access to that specific local folder on your computer. Not uploaded to any server (A), not shared (C), and not read-only (D). This is why least-privilege matters: connect only the folder Claude needs for the task at hand.

P14 (Matching). Match each Cowork agent tool to the surface or system it controls.

Agent tool Surface or system
Computer use Native desktop applications (via screenshots, clicks, and keyboard input)
Claude in Chrome Chrome browser tabs (navigate, click, fill web forms)
Claude in Excel Microsoft Excel workbooks (sidebar; read, analyze, create)
Connector (MCP) A specific external app through its API with the permissions you grant

Feedback: Each tool has a distinct scope: computer use = any installed native desktop app (visual control). Chrome = browser tabs only (web). Excel = the Excel sidebar (one application). A connector = API-backed link to a specific service. These are not interchangeable — confusing them is the top Obj 6 error.

P15 (MC). A student installs one package in Cowork and immediately gets custom formatting instructions, a Slack connector, a Google Calendar connector, and a scheduling sub-agent — all bundled together. This package is called a —
- A. Skill — a reusable instruction set
- B. Connector — an MCP link to an external app
- C. Plugin — a bundle of skills, connectors, and sub-agents ✅
- D. Live artifact — a persistent refreshing view

Feedback: A plugin bundles skills + connectors + sub-agents into a single installable package. The tell is "all bundled together in one install." A skill is just the instruction set; a connector is just the app link; an artifact is a live view — none of those are bundles. Source: Use plugins in Claude.

P16 (MC). Which statement BEST describes a live artifact in Claude Cowork?
- A. A static file Claude generates once and saves to a folder
- B. A persistent, interactive view that refreshes with current data from your connected apps each time you open it ✅
- C. A skill file that teaches Claude how to format documents
- D. A connector that links Claude to a specific external database

Feedback: A live artifact is the opposite of a static file — it refreshes with current data from your connected apps every time you open it. It persists across tasks, keeps version history, and lives in its own tab. Not a SKILL.md (C) and not a connector (D). Source: Use live artifacts in Claude Cowork.

P17 (MC). In Claude Cowork, a student types "/schedule" in a task. This command is used to —
- A. Ask Claude to schedule a meeting on her behalf in Google Calendar
- B. Configure a task to run automatically at a recurring or one-time date and time ✅
- C. Open the connector directory to add a calendar app
- D. Switch Claude to its most powerful model for complex scheduling tasks

Feedback: /schedule is the command to create a scheduled task — set it to run automatically at a specific time, recurring or one-time. It does not book a calendar meeting (A), open the connector directory (C), or change the model (D). You can also create scheduled tasks through the Scheduled area in the sidebar.

P18 (MC). Which of the following is explicitly prohibited by Anthropic's agent safe-use guidance, regardless of user permissions?
- A. Using a scheduled task to generate a daily summary of your emails
- B. Connecting a Google Calendar connector to read upcoming events
- C. Having an agent automatically execute a stock trade or financial transfer on your behalf ✅
- D. Using computer use to fill in a web form that you review before submitting

Feedback: The money rule is absolute: no AI agent executes financial transactions, trades, purchases, or transfers on your behalf — regardless of what permissions you have granted or how trustworthy the site appears. A, B, and D are all legitimate, safe uses of automation. Only C is explicitly prohibited.

P19 (True/False). True or False: When using Claude in Chrome, malicious instructions can be hidden inside a web page's content to redirect Claude's actions — this risk is called prompt injection.
- True ✅
- False

Feedback: True. Prompt injection is when malicious instructions are embedded in web page content (invisible HTML, hidden <div> elements, rogue alt-text) that redirect Claude in Chrome's behavior — causing it to take actions the user did not request. This is the primary security risk for browser agents. Claude's filters reduce but do not eliminate this risk.

P20 (MC). A student builds a Claude in Chrome workflow: (1) read a job listing, (2) draft a cover letter, (3) click Submit on the application form. Where is an approval checkpoint MOST critical?
- A. Before Step 1 — reading a page is risky
- B. After Step 2 — you should only review the cover letter, not the submission
- C. Before Step 3 — submitting a job application is irreversible and requires your review ✅
- D. No checkpoint is needed — the workflow is low-risk throughout

Feedback: The most critical checkpoint is before Step 3 — submitting is irreversible (you cannot un-submit a job application). Irreversible or high-stakes actions always require a human review moment before the agent executes them. Reading a page (Step 1) is low-risk; the cover letter (Step 2) should also be reviewed, but the submission is the point of no return.

P21 (MC). Before pasting content into a free AI tool, Prof. Quinn's "billboard test" asks the student to consider —
- A. Whether the content is fewer than 1,000 words
- B. Whether the AI tool has signed a privacy agreement
- C. Whether they would be comfortable if that content became publicly visible to anyone ✅
- D. Whether the content relates to their major or career

Feedback: The billboard test is about public exposure: if you would not be comfortable with this content appearing on a public billboard, do not paste it into a free AI tool. Length (A), the AI's agreements (B), and topic relevance (D) are not what the test asks.

P22 (MC). A student notices that an AI tool consistently generates leadership-role images showing people of only one demographic. This is best explained by —
- A. The AI was deliberately programmed to show bias
- B. AI models are trained on human-generated data that reflects existing societal biases, which can be amplified in the model's outputs ✅
- C. The student's prompts were too vague; more specific prompts eliminate bias
- D. Larger datasets always produce unbiased AI — the model just needs more training

Feedback: AI models learn from human-generated data that contains existing biases — representation gaps, historical stereotypes, unequal coverage. Those patterns are reflected and can be amplified in outputs. This is not deliberate programming (A); more specific prompts help somewhat but do not eliminate systemic bias (C); and more data does not mean unbiased — it means more of the same patterns at scale (D).

P23 (MC). Midway through a very long AI conversation, the AI starts contradicting its earlier advice and ignoring the formatting rules the student set at the start. The BEST first troubleshooting step is —
- A. Permanently switch to a different AI tool
- B. Retype the original instructions in all capitals
- C. Start a fresh conversation — the context window is likely overloaded and earlier instructions have been pushed out ✅
- D. Delete your account and create a new one

Feedback: The most common cause of contradictory or rule-ignoring behavior in a long AI conversation is context-window overload — earlier instructions have been pushed out as the conversation grew. The fix is a fresh conversation (start over). All-caps (B) is an emphasis technique, not a troubleshooting fix; switching tools permanently (A) or deleting an account (D) are drastic overreactions to what is a routine issue.

P24 (MC). During the AI 101 capstone, a student's Cowork project generates a weekly research briefing. She notices the briefing cites a paper that does not appear to exist. Which action BEST reflects the course's embed-don't-trust discipline?
- A. Assume it is a legitimate paper she cannot find through her library access
- B. Ask the same AI to confirm whether the paper is real
- C. Search for the paper in a library database, document it as a hallucination if it does not exist, and add a project instruction telling Claude to include only sources it can verify from connected data ✅
- D. Remove the verification step to make the workflow faster

Feedback: The embed-don't-trust discipline requires: (1) check externally (library database — not the same AI, which cannot verify its own errors); (2) document the error; (3) tighten the constraint (add a project instruction grounding Claude in verified connected data). Assuming it is real (A), asking the same AI (B — unreliable), and removing verification (D) all violate the course's core discipline.

P25 (MC). Looking across all eight objectives in AI 101, which statement BEST captures the course's central discipline — the habit that applies in every week, every tool, and every workflow?
- A. Trust AI output fully — it is more accurate than human memory and less prone to error
- B. Use AI only when you are sure the output will be correct, and avoid it otherwise
- C. Embed verification into every AI interaction — use your own judgment to catch hallucinations, sycophancy, over-promises, and privacy risks, and improve the output before acting on it ✅
- D. Choose the most powerful AI tool available for every task, since newer always means more accurate

Feedback: Embed, don't trust. Use AI at every step — for speed, range, and capability — but apply your own judgment to verify the output, catch errors, and improve it before you act on it. Blind trust (A) contradicts the whole course; avoidance (B) misses AI's genuine value; newest-is-best (D) is contradicted by the tool-matching unit (Obj 3). The discipline is verification and judgment — not avoidance, not blind trust.

Answer Key (Quick Reference)

Q Answer Q Answer
P1 C (next-token prediction from training patterns) P14 Computer use→desktop / Chrome→browser / Excel→Excel sidebar / Connector→API
P2 B (training cutoff) P15 C (plugin = bundle)
P3 False (search indexes & retrieves; LLM generates) P16 B (live artifact refreshes)
P4 C (emphasis with Markdown bold) P17 B (configure task to run automatically)
P5 B (meta-prompting) P18 C (financial transactions — absolute prohibition)
P6 NotebookLM→research / Midjourney→image / ElevenLabs→voice / Sora→video P19 True (prompt injection via web content)
P7 C (few-shot — several examples) P20 C (before Step 3 — irreversible submission)
P8 C (difficult-conversation simulation) P21 C (billboard test: comfortable if public?)
P9 D (Record → Transcribe → Analyze) P22 B (training-data bias reflected and amplified)
P10 C (citation hallucination) P23 C (fresh conversation — context overload)
P11 False (same AI cannot verify itself reliably) P24 C (library database, document, tighten constraint)
P12 C (persistent workspace with files, instructions, memory) P25 C (embed, don't trust — verification + judgment)
P13 B (Claude reads and writes to local folder)

Quality Gate (Self-Checked)

Canvas Placement Block

canvas_object    = Quizzes::Quiz
title            = "Practice Final Exam (Ungraded) — Weeks 1–15 (Objectives 1–8)"
module           = "Week 16 — Final Review & Exam (+ Capstone)"
assignment_group = "Practice exercises"
points_possible  = 0
grading_type     = not_graded
available_from_offset_days = 0
due_offset_days  = 6
published        = true
allowed_attempts = 3
shuffle_answers  = true
show_correct_answers = true      # feedback shown after submission
ai_permitted     = false         # same rule as the real Final; get genuine practice
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 (O-practice-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