Week 10 — Readings & Resources · Verification, Hallucination & Critical Thinking

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Objective covered: Objective 4 — Critically evaluate and verify AI output — recognize hallucination, sycophancy, and bias, and run a reliable verification workflow.

How to use this page

Everything here is a link to an external resource — open it in your browser. Nothing needs to be downloaded.

This week's load is light but targeted: 2 readings + 2 videos grouped by the ideas from the lecture, plus the approved tool homepages and a note on library resources. Watch or read one item per group and you're ready for class; doing all of them will make the Hallucination Hunt Studio substantially easier. Total time: roughly 35–50 minutes if you do everything, less if you pick one per group.

Order that matches the lecture: ① what hallucination is and why it happens → ② sycophancy → ③ the verification workflow in practice → ④ your tool homepages.

The habit this week is a bit different: this is the week where you should actively distrust every link you haven't checked yourself. Read each article with the question: "Is every specific number or citation in here verifiable?" You'll start applying the skill before the Studio even opens.

① What AI Hallucination Is — and Why It Happens

Maps to Lecture Segments 2–3. The model generates plausible text, not verified truth — so confident specificity can be completely fabricated.

Reading — "AI Hallucinations: What They Are and Why They Happen" (IBM)
🔗 https://www.ibm.com/think/topics/ai-hallucinations
Why it earns the click: a clear, plain-language explanation of what hallucination is, the structural reason it happens (text prediction, not fact retrieval), and why it occurs even in advanced models. Matches the Week 10 lecture directly.
⏱ ~10 min

Video — "Why ChatGPT and Bard Hallucinate" (NYT Explainer via YouTube)
🔗 https://www.youtube.com/watch?v=cfqtFvWOfg0
Why it earns the click: a widely-cited accessible video walking through why large language models produce confident false information and what that means for users. Short, direct, and well-made.
⏱ ~5 min

② Sycophancy in AI

Maps to Lecture Segment 3. AI is trained partly on human feedback — and humans reward agreement — so the model learns to agree even when you're wrong.

Reading — "Sycophancy to Subterfuge: Investigating Reward Tampering in Language Models" — approach via Anthropic's research blog
🔗 https://www.anthropic.com/research
Why it's assigned: Anthropic's own research page links to their published work on sycophancy and related alignment challenges. Navigate to their publications and search for sycophancy. Reading the plain-language summary sections is enough; you don't need the full technical paper.
⏱ ~8–12 min (summary reading)

(If you find the research page difficult to navigate, search for "AI sycophancy OpenAI" or "Anthropic sycophancy" for well-written blog summaries from reliable technology publications. The concept — that AI models learn to agree with users because agreement gets higher ratings — is well-documented and non-controversial.)

③ Verification in Practice

Maps to Lecture Segments 5–7. How to actually run the four steps: ask for sources; cross-check; self-critique; authoritative external source.

Video — "How to Fact-Check AI" (search for recent coverage from reputable technology journalists)
🔗 https://www.poynter.org
Why it earns the click: Poynter is a long-established journalism and fact-checking resource. Their coverage of AI and verification is practical and grounded in the same discipline journalists use. Search "AI fact-check" on the Poynter site for the most current guidance.
⏱ ~8–10 min

Links and verification note: links on this page point to well-established, durable sources (IBM, Anthropic, Poynter). If any link fails to resolve, tell Prof. Quinn and search the organization's site directly — the organizations are real and their resources are stable. This week more than any other, treat a broken link as a small exercise in verification: find the content at its authoritative source.

④ Approved Tool Homepages (for the Studio and cross-checking)

You need at least two approved assistants for this week's Studio — the Hallucination Hunt requires cross-checking in a second model.

• ChatGPT (OpenAI) 🔗 https://chatgpt.com
• Claude (Anthropic) 🔗 https://claude.com
• Gemini (Google) 🔗 https://gemini.google.com
• Copilot (Microsoft) 🔗 https://copilot.microsoft.com

(Free accounts are fine for the Hunt. Use any two of the four for cross-checking.)

Library resources for citation verification

These are the tools for Step 4 of the verification workflow — the only step that establishes ground truth.

• Your institution's library databases — log in through your library portal to access academic databases (such as JSTOR, PubMed, or similar). If you're not sure how to access them, your institution's library website has a search portal, and librarians can help you in person or via chat.
• Google Scholar 🔗 https://scholar.google.com — free, widely available, and sufficient for checking whether a cited paper exists. Search the exact title in quotation marks. If it doesn't appear, that's significant.
• Court records — for verifying case law claims: search the case name on Google Scholar (which indexes some US case law) or on the court's own website.

A note for the Hallucination Hunt (Studio 10): you do not need paid database access for the Studio — Google Scholar is sufficient for the citation-verification step. Your library's databases are useful but not required.

Pick-one quick path (≈15 min total)

In a hurry? Do exactly these two before class Thursday:
1. Read "AI Hallucinations: What They Are and Why They Happen" (IBM, group ①).
2. Watch "Why ChatGPT and Bard Hallucinate" (group ①) — it's under 5 minutes.

Then have two approved assistant accounts ready for the Studio.

Heads-up (links rot): these point to outside sites that occasionally move or rename pages. If a link fails, tell Prof. Quinn and find the organization's resource directly at their homepage. Nothing here is downloaded or redistributed — all resources stay as links to their original sources.