AIQ
AIQ
Part of the Deepfakes & Misinfo lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 🔧 Builder (8–10) · 💻 Hacker (11–14)
Open with a real, well-documented example: in January 2024, robocalls using an AI-cloned voice of a sitting U.S. president went out to New Hampshire primary voters, urging them not to vote — a genuine incident that led to an FCC investigation and a fine. Ask: "What does it take to build something like that? Who's actually liable — the person who ran the calls, the company that built the voice-cloning tool, or both?" Let the discussion run for a minute before framing the lesson:
"AI can now create fake faces, clone voices, and generate realistic video.
Let's examine how deepfakes work and how to detect them."
Set expectations for the level of this session: "We're going to go past 'AI can fake stuff' and into the actual mechanics — how these models are trained, what detection methods are technically looking for, and why this is genuinely an unsolved arms race rather than a problem with a clean fix." Walk through the hook screen's four icons — 🎭 Deepfakes, 📰 Fake News, 🔍 Detection, 🛡️ Defense — as the session's four sections.
Worth surfacing early: several students in this room may end up building tools adjacent to this space professionally — in computer vision, security, journalism technology, or policy. Frame the session accordingly: this isn't just "how to be a careful media consumer," it's also "here is a live, unresolved technical and policy problem that some of you may work on directly."
Work through the app's three scenes, using each fact as an entry point into the more technical explanation appropriate for this age band.
Face Swap: face-generation systems are trained on large sets of images or video frames of a target face and learn to synthesize new frames depicting that face in different poses, expressions, and lighting. Earlier systems relied heavily on GANs (generative adversarial networks) — a generator network producing fake frames while a discriminator network tries to tell them apart from real ones, with both improving through that competition. More recent systems increasingly use diffusion-based approaches, similar to the technology behind popular AI image generators, in addition to GAN-based methods. Voice Clone: modern zero-shot voice cloning can extract a speaker embedding from as little as a few seconds of reference audio and generate novel speech in that voice from arbitrary text — no fine-tuning on that specific voice required. Real Photo: note that "real but misleading" (an authentic photo stripped of context, cropped, or mis-dated) is a distinct problem from "synthetic," and often more common in actual misinformation campaigns than fully AI-generated fakes, because it's cheaper to produce and harder to technically disprove. AI Images: sites like thispersondoesnotexist.com demonstrate fully synthetic photorealistic faces generated on demand — worth showing live if possible, and worth noting that the same underlying generative technology extends to full video and audio, not just static images. Also worth naming as a distinct, legitimate category: the same face-swap and voice-clone research underlies real, disclosed uses — dubbing actors into other languages with matched lip movement, restoring the voice of someone who has lost it to illness or injury, or de-aging an actor on screen with informed consent and a credited process. The technology is dual-use; what separates a legitimate synthetic-media application from a harmful deepfake is consent and disclosure, not the underlying technique.
Fake News: LLM-generated fake articles can match the fluency and formatting of real journalism at essentially zero marginal cost, at any scale. Election Threats: reference the New Hampshire robocall incident from the hook as a concrete, already-happened case, not a hypothetical. AI Scams: the 2019 case where a cloned CEO voice led to a $243,000 fraudulent wire transfer illustrates zero-shot voice cloning's real-world social engineering risk. Megaphone: the contrast case — amplification of authentic speech, however biased or false its content, is a fundamentally different harm than fabricated attribution. Spend real time on the "liar's dividend" (a term from legal scholars Bobby Chesney and Danielle Citron): the existence of convincing deepfakes doesn't just enable new fabrications, it also lets someone caught on authentic, incriminating footage plausibly claim it's fake — a second-order harm to public discourse that may outlast any specific detection technology. Frame the adversarial arms race explicitly: detection improvements get incorporated into training better generators (this is structurally similar to how GANs themselves train, generator versus discriminator, escalating each other), so there's no stable endpoint where generation stops improving.
Look Closely and Detection AI: beyond the classic visual artifacts (blurring, lighting mismatches, unnatural blinking), cover biological signal detection — some detection research looks for the absence of subtle physiological signals that real video preserves and synthetic video tends to lose, such as pulse-related micro-changes in skin color (remote photoplethysmography) or natural, involuntary blink-rate patterns. Be precise that these are active research techniques with real but imperfect accuracy, not solved problems. Source Check and Critical Thinking: contrast after-the-fact detection with content provenance — C2PA (Coalition for Content Provenance and Authenticity, backed by organizations including Adobe, Microsoft, and the BBC) is a technical specification for cryptographically signing media metadata at the point of capture, creating a verifiable, tamper-evident chain of custody from camera to publication. Make the distinction explicit: detection tries to catch a fake after the fact by analyzing a suspect file; provenance tries to make authenticity provable from the start for media that was signed at capture. Neither covers everything — provenance is useless for the vast majority of existing, unsigned media, and detection keeps losing ground to generation quality. Close this scene by naming non-consensual deepfake intimate imagery as a serious, documented harm disproportionately affecting women and girls, where legal frameworks are still developing in most jurisdictions. It's worth noting for this age band specifically that laws are starting to catch up unevenly — some jurisdictions now criminalize creating or distributing this content, others treat it as a gap between existing harassment and revenge-pornography statutes, and platform policy in this area frequently moves faster than legislation, which is itself worth discussing as a pattern that recurs across most AI-harm categories.
Close by reading the app's summary together: "Deepfake technology includes face swapping, voice cloning, and video generation. Detection methods analyze biological signals and statistical artifacts. But the arms race between generation and detection continues. Media literacy is the strongest defense." Give the activity about 15 minutes, leaving room for the discussion questions below to run longer than at younger age bands.
These questions don't have clean answers, and that's intentional — treat disagreement in the room as a sign the discussion is working, not a sign it needs to be resolved before moving on. If the class gravitates toward one easy answer (e.g. "just ban it" or "just detect it better"), it's worth pushing back with a concrete counterexample from earlier in the lesson to complicate it.
This quiz leans on precise technical vocabulary from the activity above — arms race, biological signal detection, C2PA, zero-shot cloning — rather than general awareness of the topic, so it's worth confirming the class can define each term in their own words before treating a correct answer as genuine understanding rather than pattern-matching on a familiar phrase.
Close with: "Neither detection nor provenance is a permanent fix on its own — this is a genuinely unsolved problem, and it will likely stay partially unsolved for the rest of your working lives. That's not a reason for cynicism; it's the reason media literacy — questioning sources, asking who benefits, verifying independently — has to be the default habit, not a backup plan for when the technical defenses fail."
Extension activity (20–25 minutes): Have students research one real organization or initiative working on this problem — for example, the C2PA coalition, a specific deepfake detection research lab, or a documented case where an AI-cloned voice or video led to legal action (such as the New Hampshire robocall case or a wire-fraud case). Each student or pair prepares a short briefing covering: what the technology or incident actually involved, what the current state of the defense (technical or legal) is, and where they think the gap still is. As a closing debate, split the class on the question: "Should governments require content provenance labeling (like C2PA) on all AI-generated media by law?" — have one side argue for enforceability and public trust, the other for practical limits and free-expression concerns, and end by asking each student, individually, which side they actually find more convincing and why.
For students considering a technical career, it's worth pointing out directly that this is not a hypothetical future problem to prepare for — the detection labs, provenance coalitions, and policy teams working on synthetic media today are actively hiring for exactly the skills covered in this world of the app: statistics, machine learning, and the ability to explain a technical tradeoff to a non-technical audience, which is arguably the rarer skill of the two.