These 10 Key Ideas Go to 11!

This article originally appeared on LinkedIn

I started the 10 Key Ideas on Deepfakes exactly one year ago today, and concluded the series in February this year. I, along with several colleagues, tried to provide a concise and comprehensive view of the key aspects towards building effective deepfake-resilient technologies in the world of voice authentication.

And while the series focused on voice authentication, several of the key ideas had to do with the approach to developing AI-based technology in the first place.

About halfway through publishing the 10 Key Ideas, one colleague asked what I would do once we'd completed all 10; do I just stop posting? "Hey, I'll be happy if we make it to 10!" I'd replied at the time.

So now, one year later, what is there left to explore? For the astute reader, they would have noticed that in "Key Idea 3 - voice authentication is the first line of defense against deepfakes," I posited how any AI-based technologies used as deepfake defense must continue to evolve. OpenAI and other generative AI vendors don't stop innovating and improving their models year over year, and so those of us involved in the business of catching AI abuses must continue to evolve our methods and technologies. But to put it explicitly -

Key Idea 11 - you must keep up with the evolving quality of generative AI by continually improving deepfake detection and voice authentication technologies

We continue to see new generations of generative AI technologies - this year's notable releases were Veo3 and Sora 2 (which made us forget about Veo3 and every other model). We also continue to get lots of opinions on voice authentication from those outside the field, who don't have any conflict of interest at all in their claims <insert sarcasm emoji>.

Future issues of this newsletter will continue to cover topics related to deepfakes, and I will take liberties to expand to adjacent topics in the world of generative AI. And, whenever necessary, I will review and update our original list of 10 Key Ideas on Deepfakes. As before, I will strive to package some technical concepts into a more broadly palatable serving, and will dig into my experiences as a researcher, technologist, and most importantly real-world practictioner in voice authentication and deepfake detection.

Before we conclude, let's take a quick look at the original list and see if anything stands out as being out of date, or plain wrong.

The 10 Key Ideas on Deepfakes est. 2024

1. Is even more true today than it was last year

2. This will need more digging into - on social media, deepfakes are everywhere. But have voice deepfakes become weaponized at scale in call centers? Have video deepfakes? Perhaps a topic for another article

3. This remains true, but only with the caveat of Key Idea 11, introduced above

4. Still valid

5. More important today than ever before - we still see a gap in published material on deepfake detection models and what would be needed to align with reality. Definitely a topic for another article

6. I think still valid, but we can probably formalize this a bit more nowadays. Will likely revisit in a future article

7. Still valid

8. Still valid

9. This idea remains more forward-looking, so let's see how things pan out in the future - will we indeed interact with AI agents by voice? If so, this idea will still be valid. And it can likely be extended to a world where AI agents have interactions on your behalf, perhaps using an avatar of you (like a digital clone of the real you), and interact with other AI agents - shouldn't we equip these agents with a way to tell what's real and what's not?

10. Of course this remains true, but still a question on how it would be enforced, if at all. Big tech has taken the attitude that they will move fast and break things, leaving it to lawmakers to figure out things like copyright and privacy, and deal with class action suits when the time comes. I don't think I will cover this with much depth as it is out of my wheelhouse to say the least. But still it is related to Key Idea 6 - about building AI technologies responsibly