A WSJ columnist used Synthesia, ElevenLabs, and ChatGPT to create an AI clone that could trick her bank's voice biometric system, make a TikTok, and more
This is a perfect example in the difference between proper #ai technology … Jay Meier : Voice biometrics has never really been shown to be terribly effective. I know banks have used it and claimed it's great. … Tweets: Paul Armstrong / @paul__armstrong : The bank voice cloning is a massively serious issue that financial institutions need to find a response to. https://twitter.com/... Matthew Yglesias / @mattyglesias : We definitely don't want to slow down this tech and risk missing out on all the benefits https://twitter.com/... Steve Skojec / @steveskojec : Places like India and China, where regulations go to die, will receive the massive economic benefits of any collective western litigation freak out over AI. And all the dangers will still come to pass anyway. https://twitter.com/... Kenny Wassus / @kgw : spent a lot of time w @JoannaStern trying to get her replaced by AI. fun project to produce, fun edit, and i was able to sneak in lots of ai/robot pop-culture deep cuts https://www.youtube.com/... Christina Binkley / @binkleyonstyle : Uh oh. Chase Bank couldn't tell the difference between Joanna and her AI -generated voice, and put AI right through to customer service. https://twitter.com/... Glenn Gabe / @glenngabe : Well... that's not good. Super-interesting experiment from @JoannaStern -> https://www.wsj.com/... https://twitter.com/... Dare Obasanjo / @carnage4life : Voice biometrics is a technology that's gotten obsoleted before it could even catch on. Smart technology that only responds to your voice is now insecure given AI voice cloning now exists. Crazy to think both technologies ussd to be science fiction. https://www.wsj.com/... Joanna Stern / @joannastern : My AI voice clone fooled my family, bank and Evan Spiegel. And my video clone + a script from ChatGPT made a pretty decent TikTok. We're all going to be just fine. My latest video and column on the crazy—and creepy—world of voice and video AI: https://www.wsj.com/... Jeremy Olshan / @jolshan : When an AI version @JoannaStern showed up on our team's weekly Google hangout, she interrupted people with pointless non sequiturs. In other words, the AI was a humorless imitation of too many video attendees, and nothing like the genius of real Joanna. https://www.wsj.com/... Derek T. Muller / @derektmuller : We're a few years away from “Do Androids Dream of Electric Sheep” for all online contact. https://twitter.com/... Sam Gregory / @samgregory : Couldn't agree more with quote in great @joannastern explainer: “you can't make everyone an AI detective” @witnessorg consultations say don't devolve responsibility to every individual, and that its getting easier to “carpet-bomb the internet with it” https://www.wsj.com/...
Context & Ripple Effects
Joanna Stern's experiment is the sharpest public stress-test yet of a security model banks have quietly relied on for years: that your voice is who you are. Using off-the-shelf consumer tools — Synthesia, ElevenLabs, and ChatGPT — her clone defeated Chase Bank's voice biometric system outright, which analyst Jay Meier notes was never shown to be very effective in the first place.
The experiment also lands on an arc the industry has been building for a while: Hour One has been paying people to license their likenesses for AI-voiced characters since 2021, and OpenAI later packaged voice cloning into a product itself with Voice Engine's 15-second voice replication, gated to roughly ten partners including HeyGen. Stern's piece shows what happens when the same capability needs no gatekeeper at all.
First-order effects
- Chase Bank's voice biometric system failed its core test: an AI-generated copy of Stern's voice routed through to customer service indistinguishably from the real one.
- Stern demonstrated the full clone stack is now consumer-accessible — ElevenLabs for voice, Synthesia for video, ChatGPT for scripting — meaning the attack requires no specialized skill.
Second-order effects
- Banks using voice biometrics face pressure to respond, as commentator Paul Armstrong put it: financial institutions need a countermeasure to voice cloning, likely pushing them toward knowledge factors or liveness checks layered on top of voiceprint matching.
- Toolmakers like ElevenLabs and Synthesia come under scrutiny over verification of whose voice customers are allowed to clone, since their products double as fraud instruments.
Third-order effects
- If voice can no longer serve as proof of identity, banks may have to restructure customer authentication away from biometric-only flows — a structural retreat from a channel they had marketed as both convenient and secure.
- The same cloning capability that breaks bank security feeds the commercial likeness market: as Hour One's paid-likeness model and OpenAI's partner-gated Voice Engine show, consented use of a person's voice is becoming a licensed asset while unconsented use remains an open exploit.
The trend: Consumer-grade voice and video cloning is outpacing biometric identity verification, pushing banks toward layered authentication while voice becomes a contested commercial asset.