Week 2 Quiz — How AI Actually Works (Conceptually) & Its Limits

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Covers: tokens · context window · training cutoff · hallucination · search vs. AI · the Turing test · capabilities vs. limits
Format: 10 auto-graded items (multiple-choice, multiple-answer, matching, true/false) · 10 points (1 each) · allowed attempts: 1 · No AI on this quiz.

This is the human-readable quiz with its vetted answer key and one-line feedback. The import-ready Classic QTI 1.2 is in F-quiz-week-02-qti.xml (generated by a validated Python script — parses with 10 items, every single-answer item exactly one correct). Reminder: AI is not permitted on quizzes — this checks that you understand the Week 2 ideas.

Questions, key, and feedback

Q1 (MC). A large language model processes text by —
- A. looking up words in a verified database and returning the matching entry
- B. predicting the next token from patterns learned during training ✅
- C. searching the internet for the most relevant page and paraphrasing it
- D. translating your words into code and executing a lookup function
Feedback: The fundamental mechanism is next-token prediction from learned patterns (B). There is no database lookup (A), no real-time web search unless a tool is active (C), and no code execution in the base generation loop (D).

Q2 (MC). Which of the following BEST describes a "token" as used in large language models?
- A. A single complete word in the model's vocabulary
- B. A small chunk of text — sometimes a full word, sometimes part of a word, sometimes punctuation ✅
- C. An encrypted version of the user's message that the model decodes
- D. A unit equal to exactly one sentence
Feedback: A token is a chunk (B) — not always a full word (A); long or rare words may be split into multiple tokens. Tokens are not encrypted (C) and definitely not sentence-sized (D).

Q3 (MC). Your conversation with an AI assistant has grown very long — over a hundred messages. The most accurate description of what happens to early messages is that —
- A. the model reads everything and gives extra weight to the oldest messages
- B. the model automatically saves the conversation to permanent memory
- C. early messages may fall outside the context window and are no longer available to the model ✅
- D. the model deletes early messages but alerts you before doing so
Feedback: The context window is a fixed size limit (C). When the conversation exceeds it, earlier content is simply no longer in the model's visible context — no special weighting (A), no permanent save (B), and no alert (D).

Q4 (Matching). Match each term to its correct description.

Feedback: These four terms describe the four key limits of language models: the unit (token), the current-conversation limit (context window), the knowledge limit (training cutoff), and the failure mode (hallucination). Keeping them distinct is the Week 2 objective.

Q5 (MC). You upgrade to a plan that gives you access to an AI model with a much larger context window. This means —
- A. the model will give you more accurate, truthful answers because it has more "room to think"
- B. the model will never lose track of earlier parts of your conversation
- C. the model can hold more text in a single conversation before earlier content falls out of view ✅
- D. the model's training data is automatically updated to include more recent information
Feedback: A larger context window means more text fits (C) — not greater accuracy (A, still predicts tokens), not infinite memory (B, it still has a limit), and certainly not updated training data (D, those are separate entirely).

Q6 (Multiple answer — select all that apply). Which of the following are recognized shapes that AI hallucination can take? Select ALL that apply.
- A. An invented citation that looks like a real academic paper ✅
- B. A fabricated statistic (e.g., "studies show that 73% of…" where no such study exists) ✅
- C. A response that says "I'm not sure — you should check a reliable source"
- D. A confidently stated quote attributed to a real person who never said it ✅
- E. A response that refuses to answer because the topic is outside the model's training data
Feedback: Hallucination takes many shapes: invented citations (A), fabricated statistics (B), and fabricated quotes (D) are three classic ones. Option C — honest uncertainty — is the opposite of hallucination. Option E — refusal — is also not hallucination (it's the model acknowledging a limit). Select A, B, D.

Q7 (True / False). The Turing test (Alan Turing, 1950) proves that an AI that passes it is truly conscious and genuinely understands language the same way a human does.
- True
- False ✅
Feedback: False. The Turing test is a behavioral test — it checks whether a human evaluator can distinguish a machine from a human in a text exchange. Passing it is a meaningful benchmark about conversational performance, but it says nothing definitive about consciousness, inner experience, or genuine human-style understanding. Philosophers and scientists have debated this point since the paper was published.

Q8 (MC). A student needs to find the text of the current U.S. federal budget legislation. Which tool is better suited to this task, and why?
- A. A generative AI chatbot — because it can access and read any government document
- B. A search engine — because it finds and links to real, existing documents that the student can read directly ✅
- C. Either one equally — they both search the internet
- D. A generative AI chatbot — because it will write a clean summary without errors
Feedback: A search engine finds real, existing, linkable documents (B) — exactly right for "find the actual text." A chatbot generates new text and may produce a confident summary that is outdated, incomplete, or wrong (D is tempting but wrong for this task). They are not equivalent (C), and a base chatbot without a search tool doesn't access government documents live (A).

Q9 (MC). Here is a weak scenario and one prompting fix. Which response demonstrates the BEST way to handle an AI that just gave you a confident-sounding citation?
- A. Trust the citation — the AI would not fabricate a source
- B. Ask the AI to "double-check" its own citation — it will correct any errors
- C. Search for the citation independently using a library database or Google Scholar to verify it exists and the details are accurate ✅
- D. Ask the AI for a second citation on the same topic and average them together
Feedback: Independent verification (C) is the correct move — because the AI can be confidently wrong. You cannot rely on the AI to self-correct its own fabricated citations (B — the model can regenerate an equally confident wrong answer). Trusting without checking (A) is the failure mode the course exists to prevent. "Averaging" citations (D) is not a verification method.

Q10 (MC). Which statement about the relationship between the context window and the training cutoff is accurate?
- A. They are the same thing — both refer to how much text the model can process
- B. They are two different limits: the context window is how much text fits in a current conversation; the training cutoff is the date past which training data was not included ✅
- C. A larger context window extends the training cutoff to more recent dates
- D. The training cutoff sets the context window size
Feedback: These are two separate, independent limits (B). The context window governs the current conversation (a size). The training cutoff governs the knowledge boundary (a date). They do not interact the way options A, C, or D imply.

Answer key (quick reference)

Blueprint & item-bank note

All 10 items are tagged course=AI101 · week=2 · objective=1 and deposited into the item bank for future per-term regenerations. Distractors target the week's classic misconceptions (bigger context window = more truthful; AI has a live verified database; the Turing test proves consciousness; AI never makes things up; context window and training cutoff are the same).

Quality gate (self-checked)

• Structure: 10 items, 1 point each; types = 7 multiple-choice + 1 matching + 1 multiple-answer + 1 true/false.
• Single-answer integrity: every MC and the true/false item has exactly one correct option; the matching item pairs one-to-one; the multiple-answer item keys A, B, D (C and E must be left unselected).
• Product-accuracy gate: PASS. All conceptual claims (next-token prediction, context window, training cutoff, hallucination, search-vs-AI, Turing test) are accurate, non-controversial, and consistent with VERIFIED_FACTS §7. No fabricated features or statistics. The Turing test description is factual — no invented quote from the 1950 paper.
• QTI parse confirmation: F-quiz-week-02-qti.xml parses as imsqti_xmlv1p2 with 10 items; each single-answer respcondition sets SCORE = 100 on exactly one option; the matching item's partial-credit blocks sum to 100; the multiple-answer item requires the exact A/B/D set.

Canvas placement block

canvas_object    = Quizzes::Quiz
title            = "Week 2 Quiz — How AI Actually Works (Conceptually) & Its Limits"
assignment_group = "Quizzes"
points_possible  = 10
grading_type     = points
available_from_offset_days = 7
due_offset_days  = 13
published        = true
allowed_attempts = 1
shuffle_answers  = true
ai_permitted     = false
provenance       = "~ Prof. Quinn's edition · Fall 2026 · built with thecoursemaker.com"