A video call of just one minute was enough for nearly half of the participants in a recent test to be unable to distinguish whether they were speaking to a person or a machine. The culprit is Griffin, the new artificial intelligence model developed by the startup Tavus, which has challenged the human ability to detect artificial interlocutors in real time.
This data is particularly relevant at a time when video calls have become a daily tool for work, education, and personal relationships. If an AI can pass for human in this context, the implications range from customer service to digital security.
A leap that multiplies previous deception by twenty
According to data released by Tavus, 48% of participants in its tests believed they were talking to a real person during a one-minute video call with Griffin. This figure contrasts sharply with the performance of previous systems, which barely managed to confuse 2.4% of users, according to information reported by The Decoder.
This leap is no accident. Tavus has designed Griffin to process audio and video almost instantaneously, allowing the AI to react while the user is still speaking, rather than waiting for them to finish their turn, as was the case with previous models.
How Griffin achieves such naturalness
The technical key lies in processing speed. According to data cited by Progressiverobot, Tavus's video generator takes an average of 0.43 seconds from the moment it receives audio until the corresponding effect appears on screen.
The system generates video in 320-millisecond chunks and transmits speech in packets of just 10 milliseconds, a granularity that explains why the interaction is so fluid. Griffin doesn't just listen: it also observes what the user shares on screen and nods during the conversation, incorporating that visual information into its responses.
A score close to human in NVIDIA tests
Tavus claims that Griffin obtained a score of 3.83 in an audiovisual interaction test developed by NVIDIA, a benchmark used to measure this type of performance. The result is close to the 3.92 achieved by human participants in the same test.
That score clearly exceeds the 2.80 obtained by the next best system published to date, according to figures reported by Doomers. The company has presented these results as evidence that Griffin has passed what it calls the "video Turing test," a test aimed at checking whether a machine can behave indistinguishably from a human.
With these results, Tavus has dubbed Griffin the first "Human Interaction Model" (HIM), a category the company itself announced in a post on the social network X.
Promising applications and clear risks
The possibilities opened up by a technology like Griffin are vast. Among the uses mentioned by The Decoder is camera-based technical support, where an assistant could see the user's problem in real time and respond naturally.
Applications such as virtual tutors for study or digital companions designed to combat loneliness, especially among the elderly, are also being outlined.
However, the same capability that allows for these beneficial applications poses considerable risks. A tool capable of deceiving nearly half of its interlocutors could facilitate scams or disinformation campaigns that are difficult to detect, according to several analyses of the launch.
Why it is not yet available
For the moment, the full version of Griffin is not available to the general public. Tavus is keeping it as a "research preview," limited to a small group of selected testers.
As indicated by Doomers and Ground, the company is working on resolving security concerns before considering a wider release, aware of the potential for misuse of such convincing technology.
What this breakthrough means for users
The emergence of Griffin anticipates a scenario in which digital interactions—at work, in customer service, or even in the personal sphere—could become increasingly difficult to distinguish between human and artificial.
This does not necessarily imply a problem, as it can translate into more efficient and personalized services. But it also requires a higher level of attention from users when verifying who—or what—they are actually interacting with in a video call.
Frequently Asked Questions
What is Griffin and who developed it?
Griffin is a conversational artificial intelligence model developed by the startup Tavus, designed to hold video calls with a level of naturalness that the company itself describes as a "Human Interaction Model."
How many people believed Griffin was human?
In the tests conducted, 48% of participants believed they were talking to a real person during a one-minute video call with Griffin, compared to the 2.4% that previous systems managed to fool.
How was Griffin’s performance measured?
Tavus used an NVIDIA audiovisual interaction test, in which Griffin scored 3.83 points, close to the 3.92 reached by humans and well above the 2.80 of the next best published system.
Can I use Griffin currently?
No. The full version of Griffin operates as a research preview limited to selected testers, while Tavus resolves security issues before a wider release.
What risks does this technology pose?
Being capable of deceiving nearly half of the interlocutors in a video call, there is concern that it could be used for scams or to spread disinformation more credibly than with previous tools.
What could Griffin be used for in the future?
Potential uses mentioned include camera-based technical support, virtual tutors for studying, and digital companions aimed at combating loneliness, especially in the elderly.
