Brightlab
brightlab.Made for curious mindsbrightlab.Made for curious mindsBrightlab
A video looks completely real. How can you tell whether to believe it?
A deepfake is a picture, video or voice recording made or changed by AI so it shows something that never happened. Fakes are now so good that looking closely is often not enough. See how they are made, why hunting for glitches stops working, and why even a good detector raises false alarms. Then practise the checks that still work.
Why it matters: Fakes are used to scam people, bully them and spread false news. And when anything could be fake, people can brush off real evidence by calling it fake.
Starting knowledge: Percentages. Years 7–8 can focus on the verification desk; Years 10–12 can go further with the detector maths.
The animations are silent; captions carry the explanation. Pause whenever a caption changes and say what you expect next.
An image generator turns random noise into a photo of something that never happened. A voice model turns a few seconds of someone’s voice into words they never said.
Some tells, like faces that never blink, really worked, until the fakes improved. People sorting real faces from AI-made ones did no better than tossing a coin, and real photos can look strange too.
Work through six posts on the verification desk, choosing which checks to run. Then use detector maths to find out what a “fake” flag really tells you.
A feed of 1,000 videos, 20 of them fake. A detector that catches 95% of fakes and wrongly flags 5% of real videos raises 68 flags, and only 19 of them are fakes.
In Australia it is a crime to share sexual deepfakes of someone without their consent, and several states also make it a crime to create them. Sexual images of anyone under 18, real or AI-made, count as child abuse material.
Counterexample challenge. Keep fakes rare (2% of the feed) and find detector settings where most flagged videos really are fake. What would a detector need to be like? Do you think real detectors are that good?
Correct calls on at least four of the six cases, naming the check that decided each one; a completed frequency table for one detector setting, with the chance that a flagged video is fake; one sentence explaining why the absence of glitches is not evidence that something is real.
Every case, post, account and person is invented, and no real deepfake is shown. The first animation simplifies image generation: real generators work on a compressed form of the picture. Facts quoted: early deepfakes rarely blinked (Li, Chang and Lyu, 2018) and newer ones adapted within about a year (Lyu, 2019); people judged real or AI-made faces correctly 48.2% of the time (Nightingale and Farid, PNAS, 2022); in Meta’s 2020 Deepfake Detection Challenge the winning detector scored 82.6% on public test videos but 65.2% on unseen ones; research voice models have copied a voice from about three seconds of audio (Microsoft’s VALL-E, 2023). Legal statements are general; laws on creating deepfakes differ between states.
Never ask students to find, view, save or share real deepfakes, and never show sexual or abusive material. Deepfake abuse is happening in Australian schools: in 2024, explicit deepfakes of about 50 girls in Years 9–12 at one Victorian school were made from their social media photos, and eSafety reports that under-18s’ reports of digitally altered intimate images more than doubled in the 18 months to June 2025, with four in five targeting girls. If a student discloses an incident, follow your school’s child-protection and mandatory-reporting procedures and eSafety’s guide for schools on responding to image-based abuse involving AI deepfakes: report to police first, then eSafety; put the young person’s wellbeing first; and do not collect or store the images.