When Seeing Is No Longer Believing: World ID Moves to Verify Who Is Really on the Call
Daniel Levi
Posted On October 5, 2026
When AI video company Tavus revealed that its newest model could pass for a real person on a live video call almost half the time, the proof-of-human network World answered within a day. Its response on X was short and pointed: the fix for convincing fakes is not better fake detection, but a way to prove that a real human is on the call. The post directed readers to World ID for Zoom.
This was inevitable. That’s why @world_id for @Zoom exists.
https://t.co/Hhnc1q2FAV
— World (@worldnetwork) October 2, 2026
The timing was deliberate, and the stakes are not hypothetical. In early 2024, a finance employee at engineering firm Arup in Hong Kong wired $25 million after a video meeting with what appeared to be the company’s senior leadership. Every executive on that call was a deepfake. The Tavus results suggest cases like this could soon become routine.
The result World was responding to
Tavus tested its Griffin-Lite model by pairing it with participants recruited through an independent research platform. Each person believed they were joining a one-minute video chat with another participant. In reality, AI generated the face, voice, and responses on the other end live.
Afterward, 26 of the 54 participants said they had spoken with a real person, with an average confidence of 79%. That is 48%, essentially the odds of a coin flip. Tavus’s previous model fooled only 2.4% of callers, so the leap happened in a single generation. More than half of participants never became suspicious, and those who did usually formed doubts within the first 20 seconds.
Speed is a large part of why it works. Griffin produces 720p video in 320-millisecond segments, with an audio-to-video delay of 0.43 seconds on H100 chips, and can clone a voice from roughly ten seconds of sample audio.
Why businesses should pay attention
Deloitte’s Center for Financial Services projects that generative AI could push US fraud losses from $12.3 billion in 2023 to as much as $40 billion by 2027, with a more conservative estimate of around $22 billion. Entrust’s 2025 Identity Fraud Report found that deepfake attempts occurred every five minutes across its verification traffic in 2024.
What World is pointing to
World ID for Zoom checks three images against each other: the signed photo captured when the user first verified at a World Orb, a new selfie taken on their phone, and the live video frame other attendees see. If all three match, a Verified Human badge appears on that person’s video tile. Hosts can require verification before admission or request it mid-call. According to World, Zoom receives no biometric data, only the verification result and the user’s World ID username.
Zoom opened a beta of the feature on 20 May, aimed at financial approvals, medical consultations and executive decisions. Asset manager VanEck is among the early participants named in World’s business announcement. The same rollout extended verification to DocuSign signers, and email security firm Outtake uses World ID to confirm that a real person sent a given message.
The bigger picture
Roughly 18 million people across 160 countries have now verified with an Orb, up from about 450,000 in 2022. That is a small share of the global population, but enough for enterprises to begin building on. The world’s quick response to the Tavus results makes its position clear: for any organization that approves payments or makes decisions over video, trusting what appears on screen is no longer a viable security policy.
