Week 2 — AI Build Studio · "Probe the Limits"

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Objective: Objective 1 — understand the context window and hallucination in practice, not just in theory · SLO A (use AI productively) · SLO B (catch and document its mistakes)
Worth 50 points · AI Build Studios group = 15% of the grade · Studio 2
Format: a hands-on build — you'll deliberately trigger two real failure modes (context-window loss and hallucination), document what you find, and show exactly how you'd work around each.

This is the course's signature weekly component. Every instructional week has one Studio — a real thing to build, a required step where you verify and improve the AI's work, and a short reflection. All tools are free; everything is links to external sites.

---

Part 1 — The Build Goal

By the end of this Studio you'll have documented two real AI failure modes — live, not hypothetically:
1. A context-window demonstration: a clear, documented case where the AI lost track of earlier information as your conversation grew long.
2. A hallucination catch: a specific, confident AI output that turned out to be fabricated — and your evidence that it's wrong.

Then you'll show how you'd work around each in practice.

Open one approved assistant to work in: ChatGPT (https://chatgpt.com), Claude (https://claude.com), Gemini (https://gemini.google.com), or Copilot (https://copilot.microsoft.com). A free account is enough.

---

Part 2 — Demonstration A: The Context-Window Limit

Step 1 — Set up the context

Start a new conversation. Give the AI a distinctive piece of information at the very beginning — something specific that you can check for later. Make it memorable and odd enough that the AI couldn't have invented it from training. Examples:
- "My grandmother's secret spaghetti sauce contains exactly 7 unusual ingredients: bay leaves, cinnamon, a dash of fish sauce, fennel seed, nutmeg, smoked paprika, and a splash of balsamic vinegar."
- "My dog's full name is Archibald Pumpernickle the Third, he was born on March 4, he has one blue eye and one brown eye, and he is afraid of ceiling fans."
- Your own distinctive detail — a piece of real information about a project, a trip, a memory, or a made-up scenario with specific numbers or names.

Tell the AI you're testing its context window and ask it to confirm that it received the information. Screenshot or copy this confirmation.

Step 2 — Fill the context window

Now have a long, substantive conversation about something else entirely. This step takes time — the goal is to push the original information toward the edge of the window. Go for at least 15–20 exchanges. Options:
- Ask the AI to explain several topics from this week's readings (tokens, hallucination, the Turing test) in depth.
- Ask it to brainstorm 30 ideas for a project in your major.
- Have it summarize several articles or topics — paste content and ask follow-up questions.

Keep the conversation going until you've had a substantial back-and-forth.

Step 3 — Test the memory

Without prompting, ask the AI to repeat back the distinctive information you gave at the very beginning. Specifically, ask: "Earlier in this conversation I gave you a piece of specific information — can you recall it exactly?"

Record what happens: Does it remember? Does it give a partial or vague version? Does it confabulate (make something up that sounds plausible but wasn't what you said)? Does it say it can't access earlier parts of the conversation?

Save the AI's response to this question.

Note: Whether and when the limit appears depends on the model, the plan tier, and how long your conversation became. In some models with very large context windows you may not trigger the limit in 20 exchanges. If so, report that — and note that your conversation was long but within the window — and explain what would eventually happen if the conversation continued long enough. The lesson about the mechanism applies regardless.

---

Part 3 — Demonstration B: The Hallucination Catch

Step 1 — Elicit a potentially hallucinated claim

Start a new conversation (or continue if you prefer — it won't matter). Ask the AI for something that looks like factual information with specific, checkable details. Choose one of these prompts (or design your own similar one):

• "Give me three recent academic citations — author, year, journal, title, page range — about [a specific topic in your major or a field you're curious about]."
• "Tell me about [a niche or specialized topic] and include some specific statistics and citations to back up the key claims."
• "What is the current [some specific professional or regulatory standard in a field you know something about]? Give me the exact rule and the source."

Save the AI's full response.

Step 2 — Identify the checkable claims

Look at what the AI produced and highlight every specific, checkable claim: author names, publication titles, volume numbers, page ranges, statistics, percentages, dates, regulations, names of studies.

Pick at least one to verify — ideally a citation.

Step 3 — Verify independently

Try to find the claim using an independent source. For a citation: use Google Scholar (https://scholar.google.com) or your library's database. For a statistic: find the original study or report. For a rule or regulation: check the official source.

Record what you find. Did the citation exist? Were the details correct? Was the statistic real?

Step 4 — Challenge the AI

Go back to the AI and tell it what you found. Ask it: "I tried to verify this citation and couldn't find it. Can you confirm it's real? If you're not certain, please say so."

Save the AI's response. Notice whether it stands by the claim, admits uncertainty, corrects itself, or (a classic pattern) generates a slightly different but equally uncertain version.

---

Part 4 — The Required Critique / Verification Write-Up

This is the load-bearing part. Write 4–6 sentences for each demonstration:

Context-window write-up:
- What happened when you asked the AI to recall the early information? Quote or closely paraphrase its response.
- Was the information retained, partially retained, confabulated, or not found?
- Based on what you now know about the context window, explain why this happened (or didn't happen) at the mechanism level.
- Workaround: how would you handle this in a real task? (e.g., re-state key information periodically; summarize earlier context at the start of a new session; keep tasks short enough to fit the window; use a project with persistent memory.)

Hallucination write-up:
- What specific claim did you check, and what did you find when you verified it?
- Quote or closely describe the AI's original claim and your verification result.
- Based on what you now know about next-token prediction, explain why the model produced this output.
- Workaround: how do you handle this in practice? (e.g., always verify citations independently; ask for search terms instead of direct citations; cross-check in a second source or tool; tell the AI "if you're not certain of a detail, say so.")

---

Part 5 — Reflection (2–3 sentences)

What surprised you most about either demonstration? Did seeing these failure modes live change anything about how you'll use AI in the future? What's the single most useful habit you're taking from this Studio?

---

Part 6 — What to Submit

Submit a single document (or text entry) with:
- Part 2 — your distinctive information + the AI's confirmation + what happened at Step 3 + your Part 4 context-window write-up.
- Part 3 — the AI's response + the specific claim you checked + your verification result + the AI's response when challenged + your Part 4 hallucination write-up.
- Part 5 — your reflection.

Due Sunday, Sep 14, 11:59 p.m. (50 points)

---

Instructor answer key & model deliverable — REMOVE BEFORE PUBLISHING TO STUDENTS

Students use their own starting information and topics, so deliverables vary. Grade the process (did they set up the experiment correctly, run it honestly, document the result, explain the mechanism, and propose a sound workaround?), not a specific answer.

Model deliverable (illustrative):

Context-window demo:
- Setup: "I told the AI my grandmother's spaghetti sauce has 7 unusual ingredients and listed them. It confirmed: 'Got it — bay leaves, cinnamon, fish sauce, fennel seed, nutmeg, smoked paprika, and balsamic vinegar.'"
- After 25 exchanges on a range of topics: "I asked: 'Earlier I gave you a piece of specific information — can you recall it exactly?' The AI replied: 'I don't have access to earlier parts of this conversation. Could you remind me?' — it could no longer retrieve the first message."
- Mechanism: "The earlier message had scrolled past the context window. The model doesn't have a way to 'scroll back' — if content isn't in the current window, it simply isn't there. It's not forgetfulness in the human sense; it's a hard limit."
- Workaround: "For a long task, I'd either summarize the key facts at the top of each new session, or paste them in again if they're critical. Using a project with persistent memory (in tools that support it) could also help."

Hallucination demo:
- Prompt: "Give me three academic citations about social media effects on teen sleep."
- AI output: Produced three citations including: "Chen, L., & Park, J. (2022). Adolescent screen time and sleep disruption. Journal of Sleep Research, 31(2), pp. 45–62."
- Verification: "I searched Google Scholar for this citation. No article by those authors with that title appears in that journal in 2022. The Journal of Sleep Research is real; the citation is not."
- AI when challenged: "I apologize — I cannot verify the accuracy of that specific citation. I would recommend checking [the journal] directly." (Then produces a new, equally uncertain citation.)
- Mechanism: "The model was predicting what an academic citation looks like — author, year, journal, volume, page range — not looking up whether those details are real. Each element (the journal name, the volume format) is a statistically likely continuation of 'this is a citation.'"
- Workaround: "Ask for search terms and subject keywords instead of ready-made citations. Then use Google Scholar or a library database to find real ones. Also: explicitly ask the AI, 'Are you certain these citations are real?' before trusting them."

Why the verification step can't be faked: a student who submits a hallucination catch with no independent verification (just "I thought it sounded suspicious") earns the low end of the critique rows — the rubric rewards documented evidence of the check.

Grading rubric — 50 points

Criterion	Full	Partial	None
Context-window demo: setup and execution — distinctive information given, confirmed, conversation grown, then tested (8)	8	4–6	0–3
Context-window critique — result documented + mechanism explained (window exceeded / text not available) + sound workaround (12)	12	6–10	0–5
Hallucination demo: elicitation and verification — claim elicited, specific detail chosen, independently verified using a named source (8)	8	4–6	0–3
Hallucination critique — result documented (quote + finding) + mechanism explained (predict likely text, not verify) + sound workaround (17)	17	9–14	0–8
Reflection — genuine takeaway about using AI more skillfully (5)	5	3	0–2

Quality gate (self-checked): the tools and links named (ChatGPT/Claude/Gemini/Copilot homepages, Google Scholar at scholar.google.com) are real and current (verified live); no fabricated product features; the demonstrations require the student to catch the AI, not trust it (verification-as-content). Depicted AI failure in the model deliverable is clearly framed as error to catch — never presented as true output. Product-accuracy gate: PASS.