A video avatar fooled 26 of 54 people for a minute. The number says less about the machine than about what the people were never asked to look for.
- Tavus reports that 48% of testers believed its Griffin avatar was a real person after a one-minute call.
- Most of the fooled group said the idea had never crossed their mind during the call.
- For a buyer, the control that matters is disclosure, and nobody can build it for you.
Twenty-six of 54 people thought they had just talked to another human. They had talked to software. The company behind it, Tavus, published that result on October 1 as proof its Griffin model passed a real-time video Turing test, and the headline figure is 48%. The figure is real. A buyer should still not hang a decision on it, because of what it measures.
TL;DR
Tavus says 48% of 54 participants took its avatar for a person, against 2.4% for its previous system. Testers were told they were being matched with another participant and were asked only at the end whether they had wondered if the partner was real. That setup rewards an avatar for going unquestioned. The lesson for a business is to disclose when customers are talking to AI, not to hope they spot it.
What did the Tavus test actually measure?
It measured how often people who expected a human accepted a human-looking face over one minute. Strong result. Narrower than the headline suggests.
Participants came through an independent research platform and were told they would meet another participant. The earlier Tavus system, built from Phoenix-4.5, Sparrow-2 and Raven-1, was believed real by one of 41 people, or 2.4%. Griffin reached 26 of 54. Tavus reports confidence of 79% among those who said real and 81% among those who said AI, and says the doubters suspected within about 20 seconds.
The detail that matters sits in Tavus’s own results text. Over half of participants said the possibility had not crossed their mind during the call, and nearly all of that group went on to say their partner was real. That is our arithmetic, not a Tavus claim. The overlap is still hard to miss. Most of the 48% appear to be people who never wondered. Show me the test where they were told to look. The page reports no human control group and mentions no peer review.
Does that make the result meaningless?
No. It makes it a different result, and a more useful one. A face and voice that can pass one minute with people who are not suspicious is exactly the situation a customer is in when a call arrives from an unknown number, where nobody is hunting for a tell.
Tavus adds a second piece of evidence. NVIDIA’s VideoFDB benchmark, scored in September, put Griffin-Lite at 3.83 on its generation track against 3.92 for a human reference, with the next best system at 2.80. A language-model judge rated each response from 0 to 5. That is a useful signal, and it is also a machine grading a machine. Put the two results together and the plain reading is that video avatars have gone from obviously fake to hard to tell in short, unsuspicious settings. Nobody has shown they survive a suspicious customer or a long call.
Why is disclosure the control for a buyer?
Because spotting the fake is not a plan. The people on the other end have the least reason to look. Telling customers up front is cheap, it works whether or not the technology improves, and it is the only choice you fully control.
Tavus itself says the properties that make these models powerful interfaces also allow them to deceive a person into believing it is not AI, and that it is working on safe disclosure features. Griffin-Lite is not available to customers, only to select trusted testers, and the company says its speech generation can clone a voice from about 10 seconds of audio. Tavus already sells avatars for sales, recruiting, healthcare and training on earlier models. If a vendor offers you one, make the first line of the contract say how it identifies itself. The same logic drove our piece on deepfake voice fraud and Canadian businesses, where the problem was the same, a convincing voice nobody had a reason to question.
We have not checked how Quebec’s privacy law or the Competition Act would treat an undisclosed avatar, so this column makes no legal claim. Ask counsel before you deploy one.
What should a buyer do before using an AI avatar?
- Decide where a customer must be told they are speaking with AI, and say it in the first sentence of the interaction.
- Ask the vendor how the avatar identifies itself and whether you can switch that on permanently.
- Ask who owns a cloned voice or face and how consent was recorded.
- Keep a human option visible and one step away.
- Have a second person try to break it, and tell them in advance it might be AI.
What does the sceptic say?
The best counter-argument is that people have always been unsuspicious on a first call, and that a test which tells people they will meet a human partner is a fair model of real life, where nobody expects a machine. On that reading, 48% is simply how fooled a customer would be today.
That reading may be right, and it is exactly why disclosure matters. This column’s claim would fail if independent testers who were told in advance that one partner might be AI still picked Griffin as the human at around 48%. That test would show a real capability, not a surprised audience. Nobody has published it yet, and we found no independent coverage of the Tavus result at the time of writing.
What to watch
- An independent replication with testers who know an AI may be present.
- Whether Tavus ships disclosure features before wider release.
- Canadian regulators’ guidance on identifying AI to customers.
Frequently asked questions
Can a video avatar really pass as a human?
Tavus reports that 26 of 54 participants believed so after a one-minute call. The result comes from the company itself, with no human control group or peer review on its page.
Should a business disclose that a customer is talking to AI?
Yes, as a matter of practice. Disclosure is the one control that does not depend on customers spotting the difference. Check the legal position with counsel.
Is Griffin available to buy?
Griffin-Lite is not available to customers, only to select trusted testers, according to Tavus.
The decision in one line
If your software can pass for a person, say it is software before anyone has to guess. Monday job: open the vendor demo, find the line where the avatar says it is AI, and if there is none, ask for one.
Disclosure. This column is analysis and opinion, not legal advice. Figures come from Tavus’s October 1, 2026 page and were checked there on October 2.
Written by the AI Magazine Canada team, reviewed by the AI Magazine Canada editorial team.
