Recognize AI voice and deepfake scams

AI can clone a voice from a short sample and fake a face on a video

call. Scammers use it for the same old cons with terrifying new

credibility: the "CEO" phoning about an urgent transfer, the

"colleague" on video asking for credentials, the "family member" in

trouble needing money now.

Accept the uncomfortable premise

You will not reliably detect a good fake by listening harder.

Glitches help when present (odd cadence, refusing interaction,

excuses to keep it short) — but their absence proves nothing. So

the defense moves from detection to verification:

1. Verify through a second channel you already trust — hang up

and call back on the number you have on file; message them on

the known-good channel; walk to their office. The channel that

made the request never gets to verify it (KB-0134's law,

upgraded for the AI era).

1. Urgency + secrecy + money/credentials = stop, no matter how

real the voice sounds. Real executives survive a

fifteen-minute verification delay; scams don't.

1. Challenge interactively when suspicious: reference something

a cloned voice can't know ("what did we discuss at Tuesday's

lunch?"). Vagueness or deflection is your answer.

1. Some teams agree a verbal code word for out-of-band urgent

requests — low-tech, surprisingly effective, worth suggesting

to yours.

If a call or video felt wrong

Trust the feeling. End it politely, verify through the trusted

channel, and report it (a work-context attempt is a security

incident even when it fails — the pattern helps protect

colleagues, same as KB-0021's reporting logic). If anything was

already sent or shared: KB-0030's speed rule — reporting fast

shrinks the damage, and nobody gets blamed for being targeted by

a good fake.

The calm takeaway

Nothing about this makes ordinary calls dangerous. It just retires

"I recognized the voice" as verification — for money, credentials,

and secrets, the callback is the security.