One Step Closer to Deepfake Town

There has been talk of convincing deepfakes that can react in real time arriving soon. Well, the day has come.

In the video above you can see the founder of the AI company Tavus, Hassaan Raza, talking to a red-haired colleague about how they are going to launch their latest product, Griffin.

Plot twist? The colleague is not a real person.

She is an AI.

"You can still tell, blah blah blah"

And yes, you can still clearly tell it is an AI if you look closely. There is something odd about the lips, the eyes, the voice and the general uncanny valley vibe she gives off.

The demo video is also carefully edited and promoted so that the model comes across as human as possible. If you got test access to Griffin today, you would realise fairly quickly that it is an AI. The claim that their model can fool 48% of people also comes from the company itself and has not been fact-checked by outsiders.

But with all the caveats out of the way.

What the fuck???

Tavus has managed to bake in a thousand small, subtle details we associate with real people. Look, for instance, at how she discreetly rocks her chair back and forth while she "thinks". Or how her background consists of a half-drawn curtain letting daylight in onto a realistic desk. Notice the pimples on her nose too.

They have also made big strides in giving her "personality". In the video she "jokes" that Raza has been up too late, she takes a friendly "breath" before giving a thumbs up and "smiles" when she gets a compliment. All of this has, until now, been extremely human. Not any more.

Compare this with the models we have had until now, and you will clearly see the difference.

Technology that should not be built

Technically, Griffin is very impressive, and the people who built it are no doubt very smart and have put a lot of time and effort into getting everything right. The problem is just that the technology tears the world down. Because what can the model actually be used for?

Uhhhhhmmm, let me think, ehh, huh, no. Of course it will be fraud. What did you expect?

Tavus – which incidentally calls itself The Human Computing Company – claims that Griffin is meant to be a tool that humanises contact with your AI assistants. Founder Raza writes that "talking to a machine should feel as natural as talking to a friend or colleague." They envision their model as a human way for people to talk to their health assistant or a teacher.

That is not going to happen. Because who will benefit most from a technology that tricks people into thinking they are talking to a real human? Scammers will be rubbing their hands. Or intelligence agencies. Or just lonely people.

The market is big.

Memories of days gone by

I still remember the first time I realised AI videos were going to become as good as real ones. It was when my brother showed me this video, "The Great Catspy", in April 2023.

My brother (who works in film) was deeply shaken by the video. At the time I struggled to understand why he thought it was so bad. You could clearly see it was AI, the characters were completely incoherent, and it was nothing more than a fun cat video.

What I did not understand, but he did and explained to me, was that The Great Catspy was only the first thing we were going to see. Already in 2023 you could – based on the trends and data available – foresee that AI models were going to become far more powerful.

It is not going to stop here

One point I really have to hammer home when I talk to people about AI is this:

"The AI you see today is the worst AI is ever going to be. From here on, the technology only gets more powerful."

It is easy to think the models have peaked exactly where we can see them today, but that is a serious fallacy.

Given the current pace, we are only going to get even more convincing AI videos. Griffin will become as widely available as image generation is today.

That means we are going to drown in AI videos across more and more channels. According to Pew Research, 35% of all written content on English-language websites has been written by AI since ChatGPT came out in 2022. We should expect the same for video and images.

Strong is the one who does not bury their head in the sand, but prepares for an ever less human world.