The live call can now be faked.

The best independent evidence on live, two-way AI video calls comes from a team at UC Berkeley led by Hany Farid. In a preprint published in June 2026, they ran 200 conversations between paid participants and AI video agents built on two commercial platforms, Tavus and HeyGen, with an ethics-approved deception: participants were not told in advance.¹ Afterwards they were asked how long it took to realise they had been talking to an AI.

The answer was quickly. 80.5% said within 10 seconds. Only 1% never realised. The researchers also found that people did better than the audio and video deepfake detectors they tested; one text-based detector performed well.

Tavus's test of Griffin-Lite used a similar idea: participants were told they would be matched with another participant for a one-minute call "to discuss what they were looking forward to this year". Their partner was the AI. Only at the end of a survey were they asked whether it had crossed their mind that their partner might not be human.²

26 of 54 participants believed their partner was real. On Tavus's previous system, the same kind of test produced 2.4% (1 of 41). Tavus says "over half of participants said the possibility had not crossed their mind during the call", and those who did suspect "tended to suspect within the first 20 seconds".

An independent signal points the same way. NVIDIA's VideoFDB benchmark, which scores how naturally video agents handle the back-and-forth of conversation, rated Griffin-Lite close to its human reference on generating natural responses.^3,4 Two caveats apply: the scoring is done by language models rather than people, and it rates outputs the developer submits, though Tavus says NVIDIA ran the scoring independently.

The fake you were trained to spot is already gone from voice and text. The same test is now arriving for live video.

A careful reader should hold two thoughts at once.

It is a weak Turing test. Tavus ran the test itself, on 54 people, with one persona, for one minute, and has released no data.²There was no comparison group talking to real people, so we do not know how often a real human would have been judged real under the same conditions. The model tested was the lighter Griffin-Lite, not the full model. With a sample this small, the true figure could plausibly sit anywhere from about 35% to 61%. RuntimeWire summed it up: "a small company-run study, not a broad measure".⁵

It is a realistic fraud test.In Alan Turing's original game, the judge knows one of two parties is a machine and must pick which. In Tavus's design, nobody was told a machine was involved. That makes it a poor test of machine intelligence, and a good model of a payment fraud call: victims are not expecting a machine either. The jump from 2.4% to about half, under the same conditions, is far larger than the uncertainty in either number.

Griffin has the strongest published results, but it is not alone, and it is not the cheapest. Several companies now sell real-time conversational avatars by the minute, with response times they quote in fractions of a second. An open-source tool that swaps a face onto a live webcam feed from a single photo, Deep-Live-Cam, had almost 97,000 stars on GitHub when we checked.¹⁰The prices below are vendors' own figures and are not directly comparable, but the direction is clear.

Criminals are already using real-time video. The FBI warned in July 2026 that scammers "generate videos for real time video chats with alleged company executives, law enforcement, or other authority figures".¹⁵Eleven governments warned the same month that North Korean IT workers in video meetings show "video feeds that appear to be manipulated or artificially generated".¹⁶ And in March 2025 a finance director in Singapore joined a Zoom call in which the chief executive and other executives were deepfakes, and sent US$499,000.¹⁷

The obvious response is to tell staff to watch harder. The 2025 and 2026 research suggests that backfires.

Siwei Lyu, who leads the Media Forensic Lab at the University at Buffalo, wrote in December 2025 that current models produce faces "without the flicker, warping or structural distortions around the eyes and jawline that once served as reliable forensic evidence of deepfakes", and that on low-resolution video calls "their realism is now high enough to reliably fool nonexpert viewers".⁶Even a heartbeat, once a promising sign of a real face, is no longer a dependable test: a peer-reviewed study in April 2025 found that its own high-quality deepfakes "exhibit valid heart rates".⁸

A September 2026 study from Ben-Gurion University measured what happens when you ask people to look.⁷ Participants caught only about two thirds of real-time deepfake videos. When told to look for deepfakes, they called about 41% of genuine videos fake. For about half of the identities tested, the fake was rated more trustworthy than the real person.

That second number is the one programmes overlook. A culture of "be suspicious of everyone on video" means one genuine colleague in every few gets treated as an impostor. That cost can make people reluctant to challenge at all. Verification has to be routine and impersonal, or it will not happen.

Much of the advice circulating today was written for older tools. Here is where it stands on the 2025 and 2026 evidence.

Meeting platforms have started to respond. In April 2026, Zoom announced a beta integration with World ID that shows a "Verified Human" badge when the face on the video tile matches a person enrolled through World's iris-scanning device.²⁰ It confirms that a face matches an enrolled person; it does not confirm who that person is in your organisation or what they are authorised to approve.²¹ In May 2026, iProov launched a meeting product that checks whether video comes from a physical camera rather than software.²²In Microsoft's documentation for Teams, we found reporting of suspicious calls and controls over external domains, but no native deepfake detection.²³

Each of these signals proves something narrower than it appears to. An "External" label proves which organisation an account belongs to. A camera check proves the video is not injected by software. A biometric badge proves a face matches an enrolled human. None proves that this is your chief financial officer, or that your chief financial officer wants this payment. That still needs a business process.