Seeing isn’t believing | Master AI Automation in 4 hours Master AI Automation in 4 hours Course About Ayush Modules Sample chapter Toolbox The Microcap Minute Classroom / Module 11: Safety, Privacy & Nonsense Detection / Chapter 3 Seeing isn’t believing Watch first, then read. Same lesson, your pace. What you will learn – Why visual evidence lost its automatic trust – Content-level tells that still catch fakes – Context-level checks that catch the rest The new baseline Assume any photo/video/voice clip could be synthetic. Not paranoia, calibration: generation quality crossed the credibility line, and detection tools lose the arms race regularly. Your defence shifts from “does it look real?” to verification of context and provenance . Content tells (catches lazy fakes) Worth knowing even as they improve: Hands, teeth, earrings/glasses , asymmetries and melting details in images Lighting inconsistency , shadows disagreeing with light sources; reflections not matching rooms Lip-sync drift in video, especially around consonants Audio flatness , cloned voices often lack breath sounds, filler words (“umm”), emotional wobble; listen for too-perfect pacing Background text gibberish , signs, screens, book spines still trip generators Use these to raise suspicion, never to grant innocence. A fake with no visible tells is still a fake. Context checks (catches good fakes) The stronger layer: Provenance first , where did this arrive? Forwarded-with-no-origin is guilty until verified. Trace to earliest poster. Reverse image search (Google Lens/Images), old photos recycled as new events get caught instantly. Cross-source rule , Module 8’s three-source standard: no major outlet carrying it = treat as rumour, however viral. Out-of-band confirmation , for voice notes “from family” requesting money: call the person back on their known number. Every rupee lost to voice-clone scams died skipping this one step. Check for C2PA/content credentials , growing number of cameras/AI tools embed provenance metadata; verify when present. The social obligation You are now a distribution node: forwarding unverified sensational media makes you part of the scam’s delivery network. New personal policy: “Forward only what I’d sign my name under.” Verification takes two minutes; reputations take longer. Try it yourself Build your kit practically: reverse-search three images from your family group’s last week (expect at least one recycled/mislabeled). Then generate an AI image of yourself-adjacent content and run your own tell-checklist against it, learn what generators do well by inspecting closely. Finally draft your family-group PSA message (“voice note asking for money? Call back on their number first”) into learn/deepfake-kit.md . Key takeaways – Default assumption: media could be synthetic; suspicion is calibration, not cynicism. – Tells raise suspicion; provenance + cross-sourcing + callbacks establish truth. – Money/urgency requests via media always verify out-of-band. – You’re a distributor, forward like your name’s attached. Download the exercise sheet (PDF) Module workbook (PDF) ← Prev: Tricking the trickster Next: Trust protocol → Classroom / Module 11: Safety, Privacy & Nonsense Detection / Chapter 3 Seeing isn’t believing What you will learn – Why visual evidence lost its automatic trust – Content-level tells that still catch fakes – Context-level checks that catch the rest The new baseline Assume any photo/video/voice clip could be synthetic. Not paranoia, calibration: generation quality crossed the credibility line, and detection tools lose the arms race regularly. Your defence shifts from “does it look real?” to verification of context and provenance . Content tells (catches lazy fakes) Worth knowing even as they improve: Hands, teeth, earrings/glasses , asymmetries and melting details in images Lighting inconsistency , shadows disagreeing with light sources; reflections not matching rooms Lip-sync drift in video, especially around consonants Audio
Seeing isn’t believing
Written by
in