I
Ishan Pandey
Guest
Forty-eight per cent of the people who spent a minute on a live video call with Tavus's new Griffin-Lite model came away believing they had been talking to another person, which is roughly the rate at which they would have guessed correctly by tossing a coin. The previous Tavus system convinced one caller out of 41, so the jump from 2.4% to 48% happened inside a single model generation. World, the proof-of-human network built by Tools for Humanity, responded to the result on X within a day by pointing to World ID for Zoom while the timing reflects a broader argument the company has been making all year about what the internet needs once machines can pass for people.
Tavus's test was simple by design, since participants recruited through an independent research platform were told they would meet another participant to talk about what they were looking forward to this year, while the partner was in fact an AI generating its face, voice and replies in real time. Only after the call were they asked whether their partner had been real, at which point 26 of the 54 people who met Griffin-Lite said yes with an average confidence of 79%. Tavus also reports that more than half of the participants never grew suspicious during the call, while those who did usually formed the suspicion within the first 20 seconds.
Chart 1. Text models crossed the 50% line first while live video jumped from 2.4% to 48% between two Tavus systems, although the text and video studies use different designs and should not be read as a single ranking. Sources: UC San Diego and Tavus.
The result belongs in the same line as earlier research on text conversations with large language models. A UC San Diego study by Cameron Jones and Benjamin Bergen found that GPT-4.5 with a persona prompt was judged human 73% of the time in five-minute three-party text chats, while LLaMa-3.1-405B reached 56% and the 1960s chatbot ELIZA managed 23%. Video had been the harder channel because faces, timing and gaze carry so much social information, which is why the Griffin result reads as a milestone for the field.
Tavus also reports results on NVIDIA's VideoFDB benchmark for full-duplex video conversation, where a model must speak, listen and react at the same time as a person would. On the generation track, which scores fluency, matched emotion and nonverbal behaviour, Griffin-Lite scored 3.83 out of 5 against 2.80 for the next-best system and 3.92 for the human reference, which closes about 92% of the distance between the previous state of the art and a person. On the perception track, which scores understanding, visual grounding and conversational flow, it scored 3.73 against 3.44 for the strongest baseline and 4.20 for humans.
Griffin-Lite sits close to the human reference on VideoFDB generation and further from it on perception, which suggests that looking and sounding human has advanced faster than understanding a conversation like a human. Source: Tavus, reporting NVIDIA VideoFDB results
The engineering numbers help explain why a conversation with Griffin feels natural to the person on the other side. Griffin generates 720p video in 320-millisecond chunks with an audio-to-video latency of 0.43 seconds on H100 hardware, which Tavus says is half that of the next fastest method, while its voice cloning needs about ten seconds of reference audio.
The video result arrived in the middle of a busy month for agents that act on a person's behalf. Meta launched Muse on 8 September as a personal agent that sends email, books travel, completes forms and makes purchases. Three weeks later, on 29 September, OpenAI launched dots, always-on agents with their own cloud computers that connect to more than 4,000 apps. The invite-only Instinct takes a lighter route by working through text messages without any dedicated app.
Four launches and decisions between 8 September and 1 October 2026, followed by World's post on 2 October, put human-like AI on both sides of the screen, with World's proof-of-human groundwork laid in April and May. Sources: Enterprise DNA, Forbes, MacRumors, Tavus, World and Biometric Update
The friction between agents and the websites they visit showed up first in online commerce. Amazon blocked Muse from shopping on its site, saying the agent had not identified itself as an AI agent while browsing as Amazon's conditions of use require, while TechSpot reported that Amazon intends to apply the same position to shopping agents from Google and OpenAI. In a blog post on 2 October, 24 days after Muse launched, World framed the underlying question every website will now ask of an agent, which is whether a real person stands behind it.
These agents arrive on a web where automated traffic already exceeds the traffic generated by people. The 2025 Imperva Bad Bot Report found that automated traffic reached 51% of all web traffic in 2024 and overtook human traffic, with malicious bots alone accounting for 37% against 32% a year earlier according to the 2024 edition. Imperva also counted a 40% rise in account-takeover attacks in 2024, which matters because an agent holding a person's credentials looks very similar to an attacker holding them unless the site has some way to tell the two apart.
Human traffic slipped below half of all web traffic in 2024 as the bad-bot share rose five points to 37%, while account-takeover attacks grew 40%. Sources: Imperva 2025 and Imperva 2024
The financial case for verification is easiest to see in the forecasts that banks and consultancies publish on fraud. Deloitte's Center for Financial Services expects generative AI to lift US fraud losses from $12.3 billion in 2023 to as much as $40 billion by 2027, a compound annual growth rate of 32%, with a conservative scenario of about $22 billion. The best-known single case remains the Arup employee in Hong Kong who transferred $25 million after a video call on which the company's executives were deepfakes, while Entrust's 2025 Identity Fraud Report recorded a deepfake attempt every five minutes across its verification traffic in 2024.
World's approach does not try to spot a fake by analysing pixels, which is the kind of contest that models like Griffin are steadily winning, but instead asks the person on the call to prove that they are a verified human. World ID for Zoom performs a three-way cryptographic match between the signed image captured at the person's original Orb verification, a fresh selfie taken on their phone and the live frame that other participants see, after which a Verified Human badge appears on the participant's tile. Hosts can require the check in a waiting room or request it on demand during a meeting, while World says that no biometric data is shared with Zoom and that only the verification status and the user's World ID username leave the device.
Zoom opened a beta for the Deep Face feature on 20 May aimed at financial approvals, healthcare consultations and executive decisions, with VanEck among the early participants named in World's business announcement. The same release extended proof of human to Docusign signers, while Outtake uses World ID to sign outgoing email so that recipients can confirm a verified person pressed send.
The agent half of the problem uses the same World ID credential in a different way. World's blog post notes that the number of people on the planet is finite at roughly eight billion while the number of AI agents is unlimited, so rate limits, limited product drops and purchase allocations only stay fair if each agent can show that a unique person authorised it. World ID lets a verified person delegate a proof to their agent through zero-knowledge proofs, which tells a website that a unique human stands behind the request without revealing who that person is, so World describes the result as private, unique and neutral across agent makers.
Orb-verified humans grew from about 450,000 in March 2022 to nearly 18 million across 160 countries by April 2026, while the integrations listed at right show where that credential is now being accepted. Sources: World, Computerworld and Wikipedia citing company disclosures
Developers can add this through AgentKit, which is in beta, while several infrastructure companies are already building on it. Okta offers a Human Principal service in early-access beta, Vercel is adding human-in-the-loop approval to its Workflow SDK, Exa gives verified humans 100 free API requests a month and Browserbase expects verified agents to face fewer website blocks.
The deeper point in World's response is that the deepfake on a video call and the agent at a checkout are two versions of the same question, because in both cases a counterparty needs to know whether a real and unique person is present or accountable. Nearly 18 million people have verified at an Orb so far, which is small against a planet of eight billion yet large enough for enterprises to start building on, while the April release of World ID 4.0 added the key rotation, recovery and session management that production deployments require.
Griffin's 48% suggests that the window in which people could rely on their own eyes to judge a video call is closing faster than most organisations expected. If proof of human becomes as routine as a padlock icon in a browser, the companies that move first on Zoom calls, agent traffic and signed documents will have the easiest time telling their customers and colleagues apart from software that only looks like them.
Don’t forget to like and share the story!
Vested Interest Disclosure: HackerNoon has reviewed the report for quality, but the claims herein belong to the author. #DYOR.
A Video Turing Test Passed at Coin-Flip Odds
Tavus's test was simple by design, since participants recruited through an independent research platform were told they would meet another participant to talk about what they were looking forward to this year, while the partner was in fact an AI generating its face, voice and replies in real time. Only after the call were they asked whether their partner had been real, at which point 26 of the 54 people who met Griffin-Lite said yes with an average confidence of 79%. Tavus also reports that more than half of the participants never grew suspicious during the call, while those who did usually formed the suspicion within the first 20 seconds.
The result belongs in the same line as earlier research on text conversations with large language models. A UC San Diego study by Cameron Jones and Benjamin Bergen found that GPT-4.5 with a persona prompt was judged human 73% of the time in five-minute three-party text chats, while LLaMa-3.1-405B reached 56% and the 1960s chatbot ELIZA managed 23%. Video had been the harder channel because faces, timing and gaze carry so much social information, which is why the Griffin result reads as a milestone for the field.
How Close the Machines Now Sit to a Human Baseline
Tavus also reports results on NVIDIA's VideoFDB benchmark for full-duplex video conversation, where a model must speak, listen and react at the same time as a person would. On the generation track, which scores fluency, matched emotion and nonverbal behaviour, Griffin-Lite scored 3.83 out of 5 against 2.80 for the next-best system and 3.92 for the human reference, which closes about 92% of the distance between the previous state of the art and a person. On the perception track, which scores understanding, visual grounding and conversational flow, it scored 3.73 against 3.44 for the strongest baseline and 4.20 for humans.
Griffin-Lite sits close to the human reference on VideoFDB generation and further from it on perception, which suggests that looking and sounding human has advanced faster than understanding a conversation like a human. Source: Tavus, reporting NVIDIA VideoFDB results
The engineering numbers help explain why a conversation with Griffin feels natural to the person on the other side. Griffin generates 720p video in 320-millisecond chunks with an audio-to-video latency of 0.43 seconds on H100 hardware, which Tavus says is half that of the next fastest method, while its voice cloning needs about ten seconds of reference audio.
The Same Month Put AI Agents in Consumer Hands
The video result arrived in the middle of a busy month for agents that act on a person's behalf. Meta launched Muse on 8 September as a personal agent that sends email, books travel, completes forms and makes purchases. Three weeks later, on 29 September, OpenAI launched dots, always-on agents with their own cloud computers that connect to more than 4,000 apps. The invite-only Instinct takes a lighter route by working through text messages without any dedicated app.
The friction between agents and the websites they visit showed up first in online commerce. Amazon blocked Muse from shopping on its site, saying the agent had not identified itself as an AI agent while browsing as Amazon's conditions of use require, while TechSpot reported that Amazon intends to apply the same position to shopping agents from Google and OpenAI. In a blog post on 2 October, 24 days after Muse launched, World framed the underlying question every website will now ask of an agent, which is whether a real person stands behind it.
Humans Are Already Outnumbered Online
These agents arrive on a web where automated traffic already exceeds the traffic generated by people. The 2025 Imperva Bad Bot Report found that automated traffic reached 51% of all web traffic in 2024 and overtook human traffic, with malicious bots alone accounting for 37% against 32% a year earlier according to the 2024 edition. Imperva also counted a 40% rise in account-takeover attacks in 2024, which matters because an agent holding a person's credentials looks very similar to an attacker holding them unless the site has some way to tell the two apart.
What Being Fooled Costs
The financial case for verification is easiest to see in the forecasts that banks and consultancies publish on fraud. Deloitte's Center for Financial Services expects generative AI to lift US fraud losses from $12.3 billion in 2023 to as much as $40 billion by 2027, a compound annual growth rate of 32%, with a conservative scenario of about $22 billion. The best-known single case remains the Arup employee in Hong Kong who transferred $25 million after a video call on which the company's executives were deepfakes, while Entrust's 2025 Identity Fraud Report recorded a deepfake attempt every five minutes across its verification traffic in 2024.
How World ID Proves a Person Is on the Call
World's approach does not try to spot a fake by analysing pixels, which is the kind of contest that models like Griffin are steadily winning, but instead asks the person on the call to prove that they are a verified human. World ID for Zoom performs a three-way cryptographic match between the signed image captured at the person's original Orb verification, a fresh selfie taken on their phone and the live frame that other participants see, after which a Verified Human badge appears on the participant's tile. Hosts can require the check in a waiting room or request it on demand during a meeting, while World says that no biometric data is shared with Zoom and that only the verification status and the user's World ID username leave the device.
Zoom opened a beta for the Deep Face feature on 20 May aimed at financial approvals, healthcare consultations and executive decisions, with VanEck among the early participants named in World's business announcement. The same release extended proof of human to Docusign signers, while Outtake uses World ID to sign outgoing email so that recipients can confirm a verified person pressed send.
Giving Agents a Human Behind Them
The agent half of the problem uses the same World ID credential in a different way. World's blog post notes that the number of people on the planet is finite at roughly eight billion while the number of AI agents is unlimited, so rate limits, limited product drops and purchase allocations only stay fair if each agent can show that a unique person authorised it. World ID lets a verified person delegate a proof to their agent through zero-knowledge proofs, which tells a website that a unique human stands behind the request without revealing who that person is, so World describes the result as private, unique and neutral across agent makers.
Orb-verified humans grew from about 450,000 in March 2022 to nearly 18 million across 160 countries by April 2026, while the integrations listed at right show where that credential is now being accepted. Sources: World, Computerworld and Wikipedia citing company disclosures
Developers can add this through AgentKit, which is in beta, while several infrastructure companies are already building on it. Okta offers a Human Principal service in early-access beta, Vercel is adding human-in-the-loop approval to its Workflow SDK, Exa gives verified humans 100 free API requests a month and Browserbase expects verified agents to face fewer website blocks.
Why One Credential for Both Problems Matters
The deeper point in World's response is that the deepfake on a video call and the agent at a checkout are two versions of the same question, because in both cases a counterparty needs to know whether a real and unique person is present or accountable. Nearly 18 million people have verified at an Orb so far, which is small against a planet of eight billion yet large enough for enterprises to start building on, while the April release of World ID 4.0 added the key rotation, recovery and session management that production deployments require.
Griffin's 48% suggests that the window in which people could rely on their own eyes to judge a video call is closing faster than most organisations expected. If proof of human becomes as routine as a padlock icon in a browser, the companies that move first on Zoom calls, agent traffic and signed documents will have the easiest time telling their customers and colleagues apart from software that only looks like them.
Don’t forget to like and share the story!
Vested Interest Disclosure: HackerNoon has reviewed the report for quality, but the claims herein belong to the author. #DYOR.