1. The documented fact
In its 2025 Internet Crime Report, the FBI's Internet Crime Complaint Center (IC3) tracked artificial intelligence as a distinct fraud category for the first time — 22,364 complaints, nearly $893 million in losses, led by AI-assisted investment fraud at $632 million. Employment scams using deepfake video job interviews cost a further $13 million; "distress" scams impersonating a family member in an emergency, increasingly using cloned voices, added over $5 million more.
2. Two government agencies say it plainly
United Kingdom
"This will highly likely increase the volume and impact of cyber intrusions through evolution and enhancement of existing [tactics]" — the National Cyber Security Centre's official assessment of AI's effect on the cyber threat landscape through 2027.
NCSC (National Cyber Security Centre), official report. Accessed 20/07/2026.Canada
"AI technologies are almost certainly lowering the barriers to entry and enhancing the quality, scale, and precision of malicious cyber threat activity" — the Canadian Centre for Cyber Security's National Cyber Threat Assessment 2025-2026.
Canadian Centre for Cyber Security, official report. Accessed 20/07/2026.3. The probability
Our calculation — cheaper convincing, not smarter targeting
The principle: a scam's success has always depended on two things — how many people it reaches, and how convincing it is to each of them. AI doesn't change the first variable much; it collapses the cost and skill needed for the second.
What the scammer exploits: a cloned voice needs only a few seconds of real audio, publicly available on social media or a voicemail greeting. A deepfake video call requires no special access, just off-the-shelf tools. The plausibility that used to require real effort per target is now close to automatic.
Limits: government agencies explicitly frame this as an amplifier of existing tactics, not a wave of fully autonomous AI-run scams — we reproduce that nuance rather than overstating the threat, consistent with our anti-hype doctrine.
4. The mechanism
This scenario amplifies nearly every mechanism in the Grammar of Manipulation at once, but especially Credible Coincidence (a voice or face that seems to prove identity) and Fabricated Bond (a deepfake video call with someone who appears to be a real relationship).
5. The defense
🎙️A familiar voice is no longer proof
Voice cloning needs only a few seconds of audio. Any urgent request by phone, regardless of how the voice sounds, deserves independent verification.
📹Video calls aren't immune either
Live deepfake video on a call is now demonstrated, not theoretical — the same verification discipline applies as to any urgent, unexpected request.
🛡️Verify through a channel the caller doesn't control
Call back on a number you already had, or contact the person through a completely separate channel — never one suggested during the call itself.
The takeaway
Scammers don't need to know you. The statistics already do.
The scam has become bespoke. The protection had to become bespoke too.
Frequently asked questions
Does AI let scammers target people more precisely?
Not primarily. Government cyber agencies agree AI's main effect so far is making existing scam tactics more convincing at scale, not enabling more precise targeting.
How much has AI-enabled fraud cost so far?
The FBI tracked $893 million in AI-related fraud losses in the US in 2025 — the first year it measured this as a distinct category.
Can I still trust a video call from someone I know?
Treat any urgent financial request on a call or video call with the same verification discipline as an unexpected text — confirm through a second channel before acting.
Rigor note. The UK (NCSC) and Canada (Canadian Centre for Cyber Security) quotes are confirmed by direct citation from the agencies' own official reports. The US (FBI/IC3) and Australia (ASD/ACSC) figures are corroborated across multiple independent outlets but their primary government pages were not re-fetched by direct access in this pass — marked 🟡 accordingly, per our methodology. This page adapts — not translates — the French original (L'IA industrialise la vraisemblance, ANSSI/CERT-FR data), with every figure replaced by market-specific US/UK/Canada/Australia sources.
What data leaks changed
The Targeting Scale — how leaked data and AI-assisted plausibility combine to move a scam from blind to bespoke.
Learn more
Egidio — The Threat Laboratory, "AI industrializes plausibility: the demonstration", egidio.app/en/laboratoire/mathematics-of-manipulation/ai-industrializes-plausibility/. Licensed CC BY 4.0.
Free to reuse, including commercially, with attribution. Reuse terms.