AIQ
AIQ
Part of the Deepfakes & Misinfo lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 🔧 Builder (8–10) · ⚡ Architect (15–18)
Open with a direct question: "How many of you think you could tell a real video from an AI-faked one, just by looking?" Take a quick poll — most will say yes. Then say:
"Deepfakes have gone from obvious fakes to nearly perfect copies. How do you know what's real?
Let's learn how to detect AI-generated content."
Push back gently on the confident poll answers: "Five years ago, that confidence would've been fair. Today, researchers who study this full-time struggle to catch the best examples by eye alone. We're going to look at why detection is getting harder, and why 'trust your eyes' isn't the safety net it used to be." Ask if anyone has heard the terms "deepfake," "voice cloning," or "GAN" before, and let students share what they know — this age group often has fragments of the right idea from social media or games, which is a good jumping-off point to correct and sharpen.
Open the lesson on the app and walk through the hook screen's four icons — 🎭 Deepfakes, 📰 Fake News, 🔍 Detection, 🛡️ Defense — framing them as the four sections of today's lesson: what these are, why they're a real societal problem, how they're technically detected, and what actually defends against them.
If a student brings up a specific deepfake video or meme they've seen — this age group frequently has — use it as a live example rather than steering away from it: ask what made it convincing or obviously fake, and note that you'll circle back to exactly that kind of judgment call in Scene 3. Avoid pulling the actual clip up on a classroom screen unless you've already vetted it yourself; treat "I've seen a similar example, let's talk about what made it work" as good enough.
Work through the app's three scenes, treating each fact as a jumping-off point for a slightly deeper technical or social explanation than the fact itself gives. Keep a running distinction on the board between the three techniques the lesson covers — face swap, voice clone, and full image/video generation — since students at this age sometimes lump all of "deepfake" into one undifferentiated idea, when the underlying mechanisms and defenses actually differ somewhat between them.
Face Swap: a neural network trained on many photos and video frames of a target person's face learns the patterns of that face well enough to generate new frames showing it on someone else's body, in real time or in post-production — the "nearly undetectable" part comes from how much training data and compute is now available for even a non-expert to use. Voice Clone: modern voice cloning can extract a usable voiceprint from about 3 seconds of audio, and generate new speech in that voice, saying anything typed into it. Real Photo: even genuinely unaltered photos can mislead through missing context — cropping out what's actually happening around the frame, or reusing an old photo as if it depicts a current event. AI Images: sites like thispersondoesnotexist.com generate a fully synthetic, photorealistic face on every page load — show it live if you have internet access, and ask the class what they'd need to check before trusting any single online photo of a person they don't already know. Note, too, that face-swap and voice-clone tools have legitimate creative and accessibility uses — dubbing film into other languages with lip-sync, restoring a voice for someone who has lost theirs to illness — so the technology itself isn't the problem; the problem is consent and disclosure. A dubbed film tells you it's dubbed. A scam call does not.
Fake News: AI can generate a fluent, convincingly formatted fake news article in seconds, with no visual artifacts to give it away, unlike a fake video or image. Election Threats: a deepfake video of a candidate saying something inflammatory, released close to an election, can spread faster than any correction or fact-check can catch up to it. AI Scams: recount the 2019 case where an AI-cloned CEO voice tricked an employee into wiring $243,000. Megaphone: contrast this with something that only amplifies a real voice — whatever it spreads, at least it's genuinely that person's voice. Then introduce the idea this age band's quiz will test directly: the "liar's dividend." Explain it as a two-sided problem — deepfakes don't just make it easier to fake evidence, they also make it easier for someone caught doing something real and damaging to claim the real footage is fake. Ask: "If deepfakes get good enough, could someone get away with something real just by saying 'that's AI-generated'?" Let the unsettling implication land — that's the liar's dividend.
Look Closely: visual tells like blurry hands, mismatched lighting, and unnatural blinking are still useful, but explicitly note they're a shrinking safety net as generation quality improves. Detection AI: some tools look for statistical artifacts in pixel patterns invisible to human perception. Source Check: verified account, corroboration from other trusted sources, timing of the post. Critical Thinking: the strongest, most durable defense — asking whether content seems designed to provoke outrage, and who benefits if you believe and share it. Close this scene by introducing non-consensual deepfake intimate imagery as a real and serious harm: fake explicit images or video made of a real person without consent, disproportionately targeting women and girls, with laws still catching up in most places — worth naming plainly, without dwelling on graphic detail, since students this age may already have encountered the term or a related incident. If it comes up, it's worth being direct about what to do: this kind of content is a form of abuse, not a prank, and a student who encounters it — whether as a target or as someone who's been shown it — should tell a trusted adult rather than share it further, even out of shock or curiosity. Sharing it, even without meaning harm, spreads the abuse.
Close by reading the app's summary together: "Deepfakes work by swapping faces using neural networks. Voice cloning copies someone's voice from just seconds of audio. Detection tools look for artifacts like unnatural blinking or lighting errors." Give the activity about 15 minutes.
Push students past their first answer on the liar's-dividend question in particular — it's common for this age group to initially say "well, people should just believe real evidence," without grappling with the actual problem, which is that once deepfakes are common knowledge, an audience has a genuine, reasonable-sounding excuse to doubt real footage. There isn't a clean fix for that, and it's fine to sit with that discomfort in the discussion rather than resolving it too quickly.
Two of these four questions use vocabulary introduced for the first time in this lesson — "liar's dividend" and "C2PA" — so it's worth pausing on the term itself before revealing the answer, rather than only explaining the correct choice after the fact.
Close with: "Deepfake detection is a moving target — every improvement in catching fakes tends to get used to train better fakes. That's exactly why this lesson keeps coming back to critical thinking and source-checking instead of any one tool: those are the defenses that still work even as the technology on both sides keeps changing." Revisit the opening poll — ask if anyone would answer differently now about how confident they'd feel spotting a deepfake by eye alone.
Extension activity (20 minutes): Split the class into small groups and assign each group one real, documented deepfake or AI-misinformation incident to research briefly (using school-approved sources) — for example, an AI voice-cloning scam case, a political deepfake that made news, or a fact-check of a viral AI-generated image. Each group presents in two minutes: what happened, how it was eventually caught or debunked, and what a "liar's dividend" angle on it might look like (could someone use the existence of deepfakes to cast doubt on something genuinely real connected to the story?). Close by asking whether platforms, schools, or governments should be doing more, and what specifically that might look like.
If you'd rather run a shorter, single-lesson version of the extension, skip the group research and instead have the whole class collaboratively draft a one-paragraph "deepfake policy" for a fictional school or social platform: what content must be labeled, who enforces it, and what happens when the label is missing. Reading a few of these aloud tends to surface, on its own, how hard actual enforcement is — a useful, honest note to end the lesson on rather than a false sense that this is an easy problem to legislate away.