Reporter outlines how he used Lyrebird's free AI-powered software to make a realistic copy of his voice, says the digital recreation fooled his mom
Charlie Warzel / BuzzFeed News :
Context & Ripple Effects
In 2018, BuzzFeed's Charlie Warzel showed how little it took to fake a voice: Lyrebird's free tool produced a copy convincing enough to fool his own mother. At the time it read as a curiosity — a journalist stress-testing a novelty app.
The years since turned the demo into a template. Reporters went on to use cloned voices against real security systems — one bypassed Lloyds Bank's voice verification outright — while criminals moved from experiments to wire fraud, including the $243K impersonation of a German company's boss. The commercial tools only got better and more concentrated, from ElevenLabs to OpenAI's partner-gated Voice Engine.
First-order effects
- Warzel's experiment demonstrated that voice cloning required no money or expertise — free consumer software was enough to produce a replica that defeated the person who knows his voice best.
- Lyrebird got proof of its product's power, along with the first wave of attention to what unrestricted access to that power means.
Second-order effects
- Voice-as-password systems became the obvious target: subsequent reporting showed clones defeating bank voice biometrics at Lloyds and in a WSJ columnist's test of Synthesia and ElevenLabs, forcing financial institutions to treat caller voiceprint as an attack surface rather than an asset.
- The same capability migrated into fraud operations, with commercially available voice software used to impersonate executives and trigger transfers.
Third-order effects
- If the pattern holds, voice authentication survives only by adding liveness checks and secondary factors, while the burden of preventing misuse shifts to the cloning vendors themselves — visible in OpenAI limiting Voice Engine to roughly 100 partners and in the ElevenLabs founders' acknowledged unpreparedness for their technology's political fallout.
- Consent, watermarking, and access controls for synthetic voices move from academic debate to product requirements as cloning becomes a mainstream capability.
The trend: The human voice is being repositioned from a trusted biometric credential to a freely replicable asset, with vendor access controls — not detection after the fact — becoming the main line of defense.