Twenty-four short utterances — roughly five to six seconds of speech — was enough to push the false alarm rate of two state-of-the-art speaker verification systems as high as 99.11%. Humans do no better: peer-reviewed studies found listeners are "poorly equipped" to spot AI voice clones, labeling them "human" 58–70% of the time. Your help desk cannot hear the difference, and neither can the software you bought to catch it. AI in Its Place is the free weekly newsletter that translates peer-reviewed research into finance, security, and governance moves you can make this quarter. Subscribe and get this week's issue: why voice failed as an authentication factor, and the three controls that cost nothing to enforce.
Written by David Mosher—PhD candidate in Cybersecurity Management and former Microsoft leader—for the SMBs and MSPs navigating real AI adoption.

Subscribe and this week you'll get:

• The 99.11% - how voice-conversion attacks defeated two state-of-the-art speaker verification systems using seconds of audio

• What Listeners Actually Hear— two 2025 studies on why people confidently attribute cloned voices to the real speaker

• The Detection Gap — why purpose-built deepfake detection still leaves residual error against synthesis attacks

• The Three Controls— callback verification, two-person approval, and voice-independent MFA, straight from the NSA/FBI/CISA advisory

• What SMBs Can Do— the detection tooling you don't need to buy until the free controls are enforced

Every issue is grounded in primary research — no vendor hype, no fluff.