Tavus and face-to-face AI: a video call is no longer proof of a person

Tavus says its new face-to-face AI model convinced 48% of the people who spoke to it that they were talking to a human.

Griffin is billed as the first Human Interaction Model from Tavus. It interprets and responds to expressions, tone, gestures and timing in real time, and its demo passed five million views within hours of launch. Superhuman reports the 48% figure as the lab's own claim. It is not an independent test, and I have not seen the study design. But the direction is clear enough: a video call is no longer proof of a person.

That matters well beyond novelty. Large parts of commerce and law still treat seeing someone's face as a form of verification. Remote identity checks, telehealth, job interviews, sales calls and the "quick video chat" that confirms a payment instruction all lean on it to some degree. If a meaningful share of people cannot tell a model from a human, the assumption has failed, whether or not anyone has updated the process.

The detail worth pausing on is the benchmark itself. A company marketing a product is reporting how often its own product fooled people, and that is a different thing from a controlled test.

I expect the response to be less about better human detection and more about disclosure obligations, cryptographic proof of identity and, for high-stakes decisions, deliberately old-fashioned channels. Call back on a known number. Meet in person. Confirm a payment change through a second channel that the caller did not supply.

The commercial opportunity is real for customer service and training. So is the fraud opportunity, and fraud tends to move faster than policy.

Which of your critical processes still assumes that a face on a screen is a person?

Source: Superhuman AI, "Tavus reveals its first real-time human interaction model" (2 October 2026), https://lnkd.in/gRErX4wU

#AI #Trust #Cybersecurity