Artificial intelligence did not invent the scam, it perfected it. In the last three years, deepfake video, cloned voice, and large language models have scaled social engineering, turning old tricks into fast, localized confidence attacks that sound like family, look like leaders, and read like trusted colleagues. This article synthesizes 2022–2025 evidence on AI-enabled fraud across homes, schools, and enterprises, with emphasis on regions where policy and practice lag. We present a human-first threat taxonomy, representative cases, and simple defender rituals that work under stress: a callback culture, two-to-say-yes for money movement, and sandboxed document previews with content provenance. We outline file-rail risks (PDF/SVG/image preview paths) and provide a pragmatic controls stack for consumer and enterprise contexts. Finally, we propose a fear-less public education model (“Calm, Check, Confirm”), a minimal policy kit for AI-voice, synthetic sexual content, and election deepfakes, and a measurement plan any school or SME can run. The goal is practical: replace panic with protocol and make the perfect scam boring.