Class 03 — Prompting, Deepfakes, and AI Tutoring
Date: Thursday, Oct 8, 2026
Objectives
Section titled “Objectives”- Spot current visual and audio tells in AI-generated images, video (Veo 3), and voice cloning.
- Understand provenance and authentication methods: Google’s SynthID and C2PA content credentials.
- Extract requirements and write prompts with goal, audience, constraints, tone, and format.
- Iterate live on the shared screen to turn vague requests into structured tutor dialogues and study aids.
- Use AI as a Socratic study partner (quizzing and tutoring) to understand a mechanism (fission and control rods).
The assignment (projected on screen)
Section titled “The assignment (projected on screen)”Fake Assignment — “How Does a Nuclear Reactor Work?” Imagine your science teacher has asked the class to write an explainer for the school’s family science night — about 500 words — that a curious 10-year-old could follow.
Your explainer must cover:
- The core idea: what fission is, and how it turns water into electricity.
- What the control rods do, and why they’re the safety story.
- One myth you bust with a source (e.g. “a reactor can’t explode like a nuclear bomb” — but prove it).
- At least one cited source.
Graded on: accuracy · kid-friendly language · clear structure · honest sources.
AI rule for this assignment (read it twice): AI may help you brainstorm, clarify your wording, and quiz you. AI may suggest sources — but you must open and review every source yourself before it appears anywhere in your report. AI may not write your explainer. Every AI use on this assignment must appear in your prompt log.
Agenda (60 min)
Section titled “Agenda (60 min)”| Time | Block |
|---|---|
| 0–12 | Warm-up: synthetic media & deepfake showcase: student homework examples, AI-generated image artifacts (six fingers, unreadable text squiggles), news segment on video generation (Google Veo 3) and voice cloning scams, plus detection (SynthID, C2PA). |
| 12–20 | Mini-lesson: prompt anatomy — goal, context/role, negative constraints, examples, format; asking AI for questions before answers. |
| 20–35 | Demonstration — vague → strong: start from “write about nuclear reactors” and refine live on screen: set 500-word limit, 10-year-old audience, reject bad analogies (e.g. “kettle”), and test persona/tone shifts. |
| 35–52 | Live iteration & AI tutor drill (projected screen): prompt relay with student-directed constraints: - Round 1 (Role & Goal): set persona (e.g. science communicator or duck persona) and target audience. - Round 2 (Negative Constraints): outline only, no generated paragraphs, custom analogies. - Round 3 (AI as Tutor / Quizzer): instruct AI to act as a Socratic tutor guiding the class through how control rods absorb neutrons and regulate fission. |
| 52–57 | Artifact audit: compare prompt 1 vs. prompt 3 and the resulting drafts on screen. What changed — the model’s intelligence or our instructions? |
| 57–60 | Exit ticket: on paper slips, write the one prompt component or constraint you will never skip. |
Key terms
Section titled “Key terms”- Deepfake / Synthetic media — AI-generated video, voice, or imagery designed to mimic real people or events.
- SynthID & C2PA — watermarking (SynthID) and content-credential tracking (C2PA) designed to verify digital provenance.
- Assignment anatomy — pulling requirements out of an assignment sheet before writing a prompt.
- Iteration — treating the prompt as a draft to refine, not a one-shot command.
- Negative constraints — telling the AI what NOT to do (e.g. “don’t write paragraphs”, “don’t use a kettle analogy”).
- Socratic tutor pattern — prompting AI to ask questions and guide your thinking rather than handing over answers.
Seven-Question connection
Section titled “Seven-Question connection”Q1 (purpose) and Q6 (real value) — strong prompts and tutor patterns make AI improve your understanding rather than bypass learning. Q2 (is it allowed?) — the assignment’s AI rule reinforces that you must own the writing and verify citations. Q5 (ethics) — deepfake awareness highlights why verifying sources matters in real life.
What we actually covered (Oct 8)
Section titled “What we actually covered (Oct 8)”The record of the third class, from the session notes:
Trick puzzles & model differences
Section titled “Trick puzzles & model differences”- The Caged Animals puzzle: We tested a variation of the classic wolf, goat, and cabbage river-crossing puzzle where the animals were already in separate cages on one side. We compared how models (ChatGPT vs. Gemini 4) handled it: models frequently reverted to the classic multi-step crossing algorithm because pattern-matching overpowered the changed constraint.
- Getting a second opinion: Routing queries to different models (e.g., GPT vs. Gemini) produces noticeably different interpretations, but both share similar predictive blind spots.
Deepfakes, synthetic media, & detection
Section titled “Deepfakes, synthetic media, & detection”We examined the current state of synthetic media and real-world disinformation:
- AI image flaws: We looked at common giveaways—malformed hands/extra fingers (Stable Diffusion 1.5), pseudoword handwriting squiggles on letters, and unnatural lighting or melting backgrounds in portraits and food shots.
- Video generation (Google Veo 3): We watched an NBC News report (featuring Emilie Ikeda) demonstrating realistic synthetic video stand-ups, highlighting how hard it is becoming to distinguish real news footage from synthetic clips.
- Real-world misinformation: We discussed a fake video depicting an Israeli strike on Iran’s Evin Prison that fooled news broadcasts, which was later traced to an older still photo.
- Voice cloning scams & safe words: Audio cloning requires only a few seconds of audio. We discussed the practical defense: having a family “safe word” (like checking identities in Harry Potter against Polyjuice Potion) to defeat scam calls.
- Provenance & watermarks:
- SynthID (Google): Imperceptible watermarks embedded into pixel values, audio waves, and token selection.
- C2PA: Digital “nutrition labels” (content credentials) that track an asset’s provenance from camera to edit, though they often get stripped by social media compression.
Practical prompting & AI as a tutor
Section titled “Practical prompting & AI as a tutor”We used the nuclear reactor and control rods explainer as our live working lab on the projected screen:
- Vague → Strong: We started with the lazy prompt “write about nuclear reactors” and watched the AI spit out a generic, college-level textbook dump.
- Controlling tone & analogies: We tested persona shifts (including a duck persona) and pushed back on flawed analogies—the AI compared a reactor to a “controlled kettle,” which the class rejected because kettles don’t spin steam turbines.
- Socratic tutor & quiz mode: Instead of letting AI write the explanation, we instructed it: “Act as a tutor. Walk us through how control rods work by asking us questions one at a time.” The room debated answers to each question before submitting them.
- Nuclear fundamentals learned: Fission splits atoms → generates heat → boils water into steam → high-pressure steam spins a turbine → spins a generator to create electricity (like a hydroelectric dam, but powered by heat). Control rods absorb free neutrons to throttle or halt the chain reaction.
- Philosophical reflection: We closed on the nature of AI risk: the danger is human misuse of tools and deception (voice scams, deepfakes, unverified claims), not sentient machines going rogue.
Homework / log entry
Section titled “Homework / log entry”Continue your reactor explainer at home (the report is due Class 5 — that’s when we fact-check our own drafts). Log: today’s prompts, one revision, and verify one factual claim you plan to use — Class 2’s loop. If a teacher has explicitly allowed AI on a real assignment, you may log that instead.