00:11:44 · Mati Staniszewski
[00:11:44] S3: You could also argue that with recent to release, they are taking much more risk than we would ever So as we think about ourselves as like a leading company in the voice space, so important that you combine a lot of releases with the safeguards. So for us, though, what this means is from the moment we release, you can always trace back the content back to who generated that content. You can trace and understand whether it's AI or not AI, which I think will be an important part in the space, that you understand AI versus not AI content. But then second, we do invest lot in the moderation of different voices and different content itself, so you can catch that before it's created or flag it for further review. And then the last thing is we, especially over last year, we collaborated a lot in how we can use a lot of those things that we find ourselves with other players to hopefully bring that as a standard to the space. So one of those examples is you need to be able to detect AI content, but not only for ElevenLabs, but for all audio companies. And now the biggest wave that we see is a lot of open source companies, a lot of other commercial models from China that are coming as a competition, and you cannot detect that. So we are collaborating with University of Oxford, Berkeley, Reality Defender to create effectively classifier models to detect it. But maybe last thing, which was a good example, recently, a few months ago, we worked with a charity in UK where there was a use effectively, the voice agents, AI voice agents to, as the call was coming in, based on IP, if it was likely a scammer, the voice agent would take over with the purpose of wasting the scammer's time, and it worked beautifully. And it was a nice way of using AI voices exactly against the scam and fraud that was happening.
00:13:16 · S0, Taya
[00:13:16] S0: What moves our world forward? Energy. History has been shaped by it, revolutions won with it, and progress dependent on it. For one hundred ninety years, Schneider Electric has been a leading force in this story, creating the technologies that harness energy and transform the competitive landscape. And today, in the dawning era of energy and industrial intelligence, Schneider is helping customers build more efficient, resilient, and sustainable operations across every sector, from data centers to homes, from buildings to industry. Schneider Electric, your energy technology partner. Discover more at se.com. [00:13:53] S2: I wanna ask about the deepfakes because it feels like a problem that we were talking about a lot a year ago, and then I think a lot of people have just kinda felt like we're in a world where we can't tell what is the difference between an AI voice and a real voice. I mean, what's your response to that? Like, how do you think in the next few years, like, can you talk to me about some of your efforts to combat that?
00:14:12 · Mati Staniszewski
[00:14:12] S3: So there are two parts. One is what will definitely happen is we think there will be more AI generated content than human content. You might have seen some of that recently on AI music, where that's closely becoming over 50%. So I think that should be a default that everybody is operating on. AI generated content will be out there, with more of human content, and by extension, we'll need to expect and know that most of the content that's out there can be AI generated. Now the tricky thing is, what will start happening is, you'll have, of course, fully AI generated content, but also semi edited AI content, and then a little bit AI edited content. And then it's like very hard to tell. We worked. So Audible today bans AI audiobooks from Bing out there. And we have this interesting tricky dilemma with some of the book authors that use it, that sometimes you would use it for a sentence. Is that AI generated? Or is it fine? Or a 100% AI generated, probably not fine. So it's like, where is that balance? And then as you start editing, removing sounds, removing noise, AI is of course such a now buzzwordy term, it actually is used in all parts of that production process. So it becomes super tricky. So that's the first thing. I would make it clear, and I think it's the expectation that we are going with that most of the content out there will be AI generated. And then the second thing, what's the actual solution? If we like leapfrog in how we think about it, and how we are collaborating with the governments and institutes working on that is kind of a three step approach. The first layer would be a detecting for human content. So instead of thinking about AI generated content, you actually detect whether this is a human site. There's a separate problem of human authenticity piece, but in general, we encode the devices that this is device belonging to me. If I'm calling you, your device decodes, okay, this is call coming from Mati, and then you know that this is likely a trustworthy source. All fine. Second level is opted in watermarked AI content. So all of us here will have their own voice agent that will do tasks on our behalf. We'll call a restaurant to book it. We'll reschedule doctor's appointment, and you will need the voice agent. But you want it to be authenticated voice agent and the other side to know that this is authenticated voice agent.
00:16:12 · Mati Staniszewski, Taya
And here, that watermarking aspect of effectively bringing that information that I give the permission for the voice agent to act on my behalf will be important. And then the third layer, by default, everything else is AI and fake. And then you can add additional models like check or not check, but that's where we should be going. And we operate on on all three levels, mostly on a second and third one today. The first one is super tricky about the collaboration of the entire space, but the second that we are exploring and and already working on watermarking, and then on like more additional metadata that you can carry across. [00:16:43] S2: Specifically to your products, do you see people using ElevenLabs tools to orchestrate adversarial attacks? [00:16:51] S3: No, we are fighting all the time against those. So it's both the system of moderation and flagging that, but then also just the restrictions we have before you even use the product are fighting out a lot of the bad users. Of course, if the new use cases appear, we try to extend this. Think the tricky thing here is if you think about the misuse for the world, is that technology is already out there in the open without any restrictions at ElevenLabs booths. So as you're thinking about open source, you can do exactly some of the nefarious cases, and there is no way to have traceability, to have moderation, to have the ability to detect whether this was AI or not, or who generated that. So it's a tricky space, and here too, we are collaborating with others to find this.
00:16:51 · Mati Staniszewski, Taya
[00:16:51] S3: No, we are fighting all the time against those. So it's both the system of moderation and flagging that, but then also just the restrictions we have before you even use the product are fighting out a lot of the bad users. Of course, if the new use cases appear, we try to extend this. Think the tricky thing here is if you think about the misuse for the world, is that technology is already out there in the open without any restrictions at ElevenLabs booths. So as you're thinking about open source, you can do exactly some of the nefarious cases, and there is no way to have traceability, to have moderation, to have the ability to detect whether this was AI or not, or who generated that. So it's a tricky space, and here too, we are collaborating with others to find this. [00:17:30] S2: I'm curious with all this, like, talk about, like, you know, you said there's gonna be more AI generated content out there than human. What do you think the value is of a human generated voice and a human generated piece of music in the AI age?
00:17:45 · Mati Staniszewski
[00:17:45] S3: Look, there's so many things. One is the nuance and the slight... And and I think it applies actually both in AI and a human space. Nuance of dialect, the accent, the voice. So I'll I'll speak about this a little bit. But then there's also the second piece. As you think about music, if you think about creating an audiobook, if you're creating a movie, at the end of the day, it's about the story. It's about some of the emotions that people felt in a given moment. Something that connects you to the other side that actually makes it human. I'm sure you follow a lot of a lot of people in this space, and it's a lot more frequently than the just the act that was created that makes it so valuable. And I think this will remain true as long as all of us are are here, where you do attach to the other person, you follow that story and and an emotional part. And then on the voice side, so first of all, even in AI space, today we've created a voice marketplace where everybody can create their voice and share the voice, and when the voice is being used by other users, you earn money as a result. The voice works across all 70 languages. Our top voice is a Spanish voice that sounds incredible in English. So most of the users are English speakers. So if any of you want to register on that platform, we need more voices. But there's just so many different ones that effectively show how dialect, accent, different languages is important. And still, we are just scratching the surface. We have 10,000 voices shared this way, paid $10,000,000 back to the community. And I think there's like just a wave of other special, unique voices across across the universe. So that's one. And then two, ability to collect on the on the deeper level.