For most of the internet era, the bargain for experts and creators was relatively straightforward: publish your knowledge, build an audience and hope that distribution eventually leads people back to you.

AI has fundamentally changed that equation.

Today, an expert’s work can become part of an AI-generated answer without necessarily identifying the source, compensating the creator or accurately reflecting that person’s point of view.

Onix wants to build a different model.

On this episode of The Media Copilot podcast, Pete Pachal speaks with Onix co-founders David Bennahum and Nicholas Nadeau about their vision for expert-owned AI: specialized models built around an individual expert’s knowledge, voice, judgment and perspective rather than a massive model trained to know a little about everything.

Bennahum and Nadeau argue that expertise is more than factual information. It includes the judgment and perspective an expert develops over a lifetime, including knowledge that may never have appeared online. Onix combines publicly available work with what the company calls “dark data,” including private notes, manuscripts and other material that can capture how an expert actually thinks.

“𝗘𝘅𝗽𝗲𝗿𝘁𝗶𝘀𝗲 𝗶𝘀 𝗳𝗮𝗰𝘁𝘂𝗮𝗹 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗽𝗹𝘂𝘀 𝗮 𝗽𝗼𝗶𝗻𝘁 𝗼𝗳 𝘃𝗶𝗲𝘄, 𝗮 𝗽𝗲𝗿𝘀𝗽𝗲𝗰𝘁𝗶𝘃𝗲 𝗮𝗻𝗱 𝗷𝘂𝗱𝗴𝗺𝗲𝗻𝘁.”

— Nicholas Nadeau, Ph.D., P.Eng., Co-Founder & CTO, Onix

The conversation goes well beyond the technology.

Pete digs into authorship and intellectual property in the AI era, data sovereignty and privacy, the economics of expert knowledge, small language models versus frontier models, hallucinations and whether creators can build sustainable businesses around AI versions of their own expertise.

Bennahum describes the opportunity as creating an “authorized version” of an expert in AI, rather than relying on what he calls the “bootleg” version assembled by general-purpose models from information found across the internet.

“𝗧𝗮𝗸𝗲 𝗯𝗮𝗰𝗸 𝗰𝗼𝗻𝘁𝗿𝗼𝗹 𝗼𝗳 𝘆𝗼𝘂𝗿 𝗴𝗲𝗻𝗶𝘂𝘀.”

— David S. Bennahum, Co-Founder & CEO, Onix

The implications could extend far beyond Onix’s initial focus on health and wellness. The company envisions eventually expanding into areas including personal finance, self-improvement, relationships and lifestyle expertise, with a model that could ultimately allow more creators to build, privately share and monetize their own Onix.

Pete also asks the larger question hanging over the creator economy:

If AI is going to generate enormous value from human knowledge, who should own that value?

For Onix, success would mean that the economic benefits of AI are distributed among the experts whose knowledge makes these systems valuable, rather than concentrated among a handful of technology companies.

It’s a conversation about AI, but ultimately it’s about authorship, ownership and what human expertise is worth in a world where knowledge can suddenly scale.

In this episode

Pete, David and Nicholas discuss:

• Why Onix is betting on small, domain-specific AI models
• The difference between factual knowledge and true expert judgment
• Why AI’s current approach to authorship and intellectual property may be unsustainable
• Giving experts ownership and control over their AI models
• The role of private or unpublished “dark data” in creating higher-fidelity AI
• Why privacy and data sovereignty are central to the Onix model
• Whether specialized models can reduce hallucinations
• Turning an expert’s knowledge into a subscription business
• The potential for journalists, creators and other professionals to build their own AI products
• What a healthier “author economy” could look like five years from now

Onix is currently in early access. Media Copilot listeners can visit Onix.life and use the access code COPILOT to try the platform.

🔗 About the 👤 Guests

David S. Bennahum | Co-Founder & CEO, Onix. | https://www.linkedin.com/in/dbennahum 

Nicholas Nadeau, Ph.D., P.Eng. |  Co-Founder & CTO, Onix.| https://ca.linkedin.com/in/engnadeau 

About the show: To explore more conversations like this and see what’s new, visit the Media Copilot website at mediacopilot.ai. You’ll find new episodes, expanded resources, and tools designed for journalists, communicators, and media leaders navigating the fast-changing world of AI. It’s the home base for everything Media Copilot and it’s just getting started.

Enjoyed this episode?

Subscribe to The Media Copilot on Substack, Apple Podcasts, Spotify, or your favorite app. On YouTube? Tap the Like button and Subscribe to the YouTube channel. For more AI tools and resources built for media professionals, visit mediacopilot.ai.

Produced by Pete Pachal and Executive Producer Michele Musso
Edited by the Musso Media Team 

Music: “Favorite” by Alexander Nakarada, licensed under CC BY 4.0

All rights reserved. © AnyWho Media 2026

TRANSCRIPT 

Guest interview with David S. Bennahum and Nicholas Nadeau, Ph.D., P.Eng. | Onix

[00:40] Pete Pachal

Hi, welcome to the Media Copilot. It’s a podcast about how AI is changing media, news, and communication. I’m your host, Pete Paschel. I covered tech for a long time as a journalist, and now I talk with the media leaders, the builders, and the creators, all trying to answer the question: how will we get information in the future? And how will that transform journalism and the business of media?

For most of the internet’s history, experts had a fairly simple deal. You published your knowledge. Platforms help distribute it, and hopefully some of the audience found its way back to you. AI has changed that arrangement. An expert’s work can now become raw material for an answer that may or may not identify the source, compensate them, or even accurately represent the work in the first place.

A company called Onix is proposing a different model. It builds a dedicated AI around one expert’s body of work, including material that may never have appeared online. Users can subscribe to that system, ask questions, and receive personalized guidance based on that expert’s knowledge and their stated point of view.

My guests this week are the co-founders of Onix, David Benham and Nicholas Nadeau. David started his career as one of the original writers at Wired and has spent a lot of time working at the intersection of media and technology. Nicholas is a PhD in human robot interaction, previously served as the CTO of Humanoid Robotics Company One X. And has worked on the data infrastructure behind major AI systems.

What I find particularly interesting about Onix is what its model could mean beyond the health and wellness category that they’re starting with. Could a journalist, an analyst, a creator, or even a comms expert turn their knowledge into an AI product they can control? And then who owns that system when the expertise was might have been developed inside a company? And maybe most importantly, When an AI speaks with a person’s name and authority, who is responsible for what it says?

I’m excited to get into all that. but first, quick note. If you’re listening on Apple or Spotify, please leave a five star review and maybe a nice comment. and if you’re watching on YouTube, please like the video and subscribe to the channel. Those things really do help more people find the show.

[02:53] Pete Pachal

All right, let’s get started. David and Nick, welcome to the media copilot.

[02:58] Nicholas Nadeau

Thank you so much for having us.

[02:58] David S. Bennahum

Thanks for having us, Pete.

[03:00] Pete Pachal

Awesome. So let’s get started on Onix a little bit. so David, for for someone just hearing about Onix for the first time, what exactly is it? And as I sort of mentioned in the intro there, like what is the subscriber paying to receive? Give us a little more texture on that.

[03:18] David S. Bennahum

Sure. So think of it as a platform where people with deep knowledge, expertise they’ve built over a lifetime, can take that knowledge, pair it to their own small language model that we provide them with. That model is then really fit to their tone, their voice, the depth of their expertise. And then the consumer side can summon those models, essentially summon that expertise to help solve some of life’s most challenging problems. And in our case, we’ve started with health and wellness because we think that’s an area where unique expert opinion and experience matters, we think it’s an area where trust and privacy are really paramount and that’s another key piece of this equation. But ultimately think of it as a marketplace where individuals that have acquired a lifetime of knowledge can make that knowledge available for subscription to anyone for whom that could make a big difference in their lives.

[04:13] Pete Pachal

Nice. I like I like the word summon. It sort of implies almost a sort of a magical quality to this that this sort of expertise on demand. Nick, why don’t you tell me a little bit about the the model you have? You have like a one expert, one system model. walk me through sort of what happens between an expert sort of handing you their work, you know, and a and a functioning Onix version of them appearing in the app. yeah.

[04:16] David S. Bennahum

Yeah.

[04:35] Nicholas Nadeau

Yeah. so, yeah, so it’s actually, it’s not that difficult. It’s quite easy in a lot of ways. A lot of the experts we work with that are already published, they have books, have podcasts, have Substacks, blog posts, you name it. already have a lot of internet presence. Now, this is a lot of the work that has been stolen and scraped by big AI in a lot of ways. But we’ve published some technical white papers basically showing that if you build these models with the intent and the actual intention to create high fidelity to that expert and their point of view, even just off that first layer of public data, we can already outperform. And this is where you get into the domain focus, you know. big god models. They are trained on all of the internet from the history of Japan to the French language to how to code. That’s not our focus at all. We’re domain constrained to the specific expertise of that expert. And then you bring a multimodal architecture. We have nutritionist experts, have sports medicine experts, have longevity functional health. Each one doesn’t need to know all of health and wellness, but you bring them in, as David mentioned, summon them in to your health and wellness journey. And they all take part in this approach. And so the architecture starts to look a little more like coding agents in a lot of ways. We’re working together on a workspace together. I come in as a user with, know, for using a terminology world model of me and my humanity with these experts and expertise working together towards some end goal, whether it be navigating perimenopause or fertility or looking at even like I myself using it for my marathon training, bringing in a nutritionist and a sports medicine recovery expert. to progress on that journey. And so we might start with the public data, and then we could go deeper working with our experts to unlock some of that dark data, the manuscripts and memos and notes that have never gone onto the internet, but that really show the nuance in the way they think, their judgment, and their persona in a lot of ways. And then we work with them, human-in-the-loop through an active learning system, where we can bring in the things that have never been written down, their real nuance.

[06:20] Pete Pachal

Well let’s

[06:45] Nicholas Nadeau

point of views. And so the core belief for us is that expertise is factual knowledge plus a point of view, a perspective and judgment. And together, that’s what true expertise is, not just, you know, statistically average.

[07:00] Pete Pachal

Yeah, I I think what you’re saying like is the sort of a more articulate way of talking about what I think is probably an overused term in AI, which is like personalization or hyper personalization is kind of the the cool term. And yeah, I I the seeing sort of the well, I would like you to articulate for me for like like the advantage that this has over big generic systems, because I think that’s what most people are trained to do now, right? Like it’s like

[07:25] Nicholas Nadeau

Yep.

[07:29] Pete Pachal

just ask ChatGPT, just ask Claude for your advice on what supplements to take or what have you. And for a lot of people, like maybe that’s enough. But the idea that, like, you know, you want to go deeper, but you also, you know, don’t necessarily have the budget or the time to go to a a a human and make make appointments and calls and all this stuff, like it feels like this is fulfilling that that gap, but in a much

[07:52] Nicholas Nadeau

Yes. Yeah. David, do want to take this one? Yeah.

[07:58] Pete Pachal

sort of more efficient way.

[08:01] David S. Bennahum

Yeah, there’s a well, I mean, just there’s a a huge latent demand for expert opinion. And that latent demand is really unfulfilled because people with it only have so much time in the day, and we don’t live near them, and they’re even hard to see physically. And when you think about all the domains of human knowledge that are out there, which are just basically limitless, what does it mean to unblock that expert in being able to give that knowledge to really anybody? And the the key is Would the fidelity or the accuracy of the model be such that the expert would feel really good that this would maintain the integrity of their own brand and their own commitment to their knowledge, right? This is a key piece. So this doesn’t work if experts felt it was misrepresenting them or falling short. But if they do, wow, the implications are tremendous because the volume of latent demand around service knowledge. Literally in the trillions of dollars. I mean, it’s actually huge. It’s one of the biggest drivers of inflation in the world today is actually the increase in prices for services. And that’s because there’s essentially a kind of wage inflation that occurs there. That’s a function of the scarcity. It doesn’t have any efficiency in it. You know, a world-class doctor can only see so many hundreds of people a year. End of story. And that’s not going to change. It was the same in 1850, it’s the same in 2026.

[09:05] Pete Pachal

Mm.

[09:28] David S. Bennahum

What does it mean when all of a sudden that latent demand can go from a few hundred people a year to a few million people a year? Again, presuming the doctor or the expert just felt that that AI faithfully represented their wisdom. If the answer is yes, then holy moly, this is one of the biggest contributions to human flourishing we can think of, because it suddenly democratizes access to all this wisdom that otherwise would have not been available.

[09:55] Pete Pachal

Well it also seems like you’re coming at this like at a at a at a good time, like with everything you were talking about, Nick, with regard to smaller models, right? And they’re like the the biggest thing that everyone’s talking about right now in AI is like how much it costs to d just operate some of these sort of frontier models and how do how do we prevent our token costs from getting out of control. And so getting like a more efficient model in there and also the edge computing, right? There’s that that’s a part of that equation.

[10:10] Nicholas Nadeau

Yeah. Yeah.

[10:24] Pete Pachal

And it seems like your your system like takes care, like it is just sort of inherently coming at this time when, hey, you can actually not only run more efficiently, but get better results by sort of narrowing the focus. This whole small model approach, which I think was not but not theoretical, but I think the the sort of mass application of it was probably a little a blurry, but now you’re sort of giving it a little more a little more concrete in terms of application.

[10:38] Nicholas Nadeau

Yeah. Yeah. Yeah, exactly. We’re really at a perfect point right now where like the crest of the wave, where the models are becoming smaller and more efficient and the phones and like the supercomputers we carry in our pockets every day are becoming more more powerful. And so if you have like an iPhone 17 Pro, you can already start running these models in your pocket. And so our end goal from a data sovereignty and privacy and trust point of view is let’s run everything local to you in your data vault and your safe place basically where your wearable data lives, where your Apple HealthKit lives, where your deepest, darkest secrets live, giving you that space to explore your wellness journey. Because honestly, we don’t need that data. And that’s the entire premise of Onix at the end of the day is we’re at a point where you don’t And you no longer need to trade your privacy and your trust for utility. That was the Web 2.0 era where you wanted Facebook, you wanted to hear, I’m great, it’s free, but give me all your data. That’s no longer the case and doesn’t need to be.

[11:46] Pete Pachal

Mm-hmm. Yeah, let’s talk about that. Let’s talk about the ownership, the data sovereignty aspect of this. So I I I understand sort of like your model is that the experts own their model, right? And they can they can leave anytime. And the all of the conversations that any user might have with the the AI version or the wh what is the term? Is it is it an Onix? What do you what do call it? Yeah, so for their Onix.

[12:14] David S. Bennahum

Yes. Anonymous.

[12:15] Nicholas Nadeau

Yeah, analytics.

[12:17] Pete Pachal

are encrypted. And so so tell me a little bit about like like how you see data privacy and also the data sovereignty of the experts and what safeguards specifically that you have in place to ensure that well that the data doesn’t get used in ways people might not might not like.

[12:38] David S. Bennahum

Well, maybe maybe we just begin with a basic North Star. And let’s say instead of expert, let’s say authorship, right? That you as a human being, Pete, create knowledge. You do that for for a living. And many people do that in different domains. And just think of that as authorship. It it’s fundamentally part of the success of our civilization has been giving people the property right to own their intellectual work.

[12:52] Nicholas Nadeau

That’s one.

[13:06] David S. Bennahum

Within certain guardrails, and then from there to allow other people to build on that in ways that are called fair use, et cetera. But ultimately there’s a regime that protects people’s intellectual property. And guess what? It’s been part of why we’ve thrived as a society since at least the eighteenth century. Now, what is it about AI that raises its hands and says, geez, you know, authorship doesn’t matter?

[13:19] Nicholas Nadeau

So,

[13:32] David S. Bennahum

Property around owning knowledge doesn’t matter. We somehow have this ability to go and take it all without compensation. Well, that’s something they just said. And that partly they got away with it because it was first mover advantage, you know, in 2017 when the first general purpose transformer models really started to work. A lot of people took a lot of stuff from the internet and nobody really noticed until 2022 when GPT, ChatGPT came out. And then all of a sudden I was like, ooh, look at this. It’s so cool. Then people wondered how did they get all this data? And it was a little bit like the cat.

[13:48] Nicholas Nadeau

So.

[14:01] David S. Bennahum

Was out of the bag. Okay. Having said that, authorship remains paramount. And the reality is the frontiers of human knowledge, what you do today, Pete, what other people are doing, those things deserve to be protected. And people won’t have much incentive to continue pioneering the frontiers of knowledge. If all that happens is their knowledge gets extracted without permission and used by somebody else for their profit. It’s it’s just completely laughable. And

[14:16] Nicholas Nadeau

So I’m just going to take a moment to our Dr. David

[14:27] David S. Bennahum

One of the many reasons there’s a huge backlash against AI is in part this perception is just taking everyone’s stuff willy-nilly and doesn’t seem to care. That is coming to an end. Now, when you enforce authorship and what I would say is ownership adjacent to it, right? Your ownership of what you’ve created, Pete, then you create a system in AI that allows for the frontiers of knowledge to flourish and even more interestingly, allows for orchestration to occur. Where Pete’s knowledge can potentially connect with, say, Malcolm Gladwell’s knowledge, not that the two things get synthesized in the same model, not at all. Presuming Malcolm had his own, shall we say, Onix as well. Those two things could be summoned by a human being to discuss something around, say, innovation and company formation, where they wanted Pete’s point of view and Malcolm’s point of view in the discussion. Currently completely impossible with a frontier model today. That wouldn’t really mean anything to do what I just said, because their knowledge has been blended together, right? pureed and munged together into the same model. So authorship is at a point now where it’s going to become ascendant in the domain of AI for the reasons Nick also pointed out, which is small models are are having their moment. They can be really fit tested to a person’s knowledge. And from there they can be orchestrated, as we say, or summoned collectively together to solve problems in ways that a blended model can’t actually have the discernment to solve. Because it doesn’t have a differentiation of judgment. It just keeps looking for the average between, say, Pete and Malcolm Gladwell, hypothetically. Like what is the average between Pete and Malcolm? Well, on some level, who the heck cares what the average is? That’s not the point. The point is, what’s the debate, maybe, between Pete and Malcolm? that’s much more interesting. How do you get that debate on a frontier model? you can’t. That’s interesting. So when enforcing authorship, enforcing ownership of that authorship back to you, leads to a better form of AI that can solve problems in ways a frontier model.

[16:01] Pete Pachal

Hmm.

[16:25] David S. Bennahum

cannot, that is a rare moment where actually the moral, the moral architecture happens to be the more effective architecture in terms of performance. When that happens, we all ought to seize it vigorously. And that is where we are today, in a very exciting moment where I think a lot of the assumptions around AI are about to be challenged. And guess what? Always happens in our industry. This is nothing new, guys. It’s happened to, you know, from inception Different topic. We could talk about it, but there’s nothing new about big tech incumbents getting blindsided by a change and how things can be done.

[17:02] Pete Pachal

So the the whole aspect of creators owning their work and you know the the the everything you just sort of touched on there where where you know there’s been what a lot of people see as sort of this grand theft of of knowledge on the internet. you may have heard there’s a bit there it’s sort of led, I think, in some ways to a bit of a backlash against AI, particularly around like IP owners. And so I’m curious as you talk to people to come into your network, how much of a process of, I guess, education or convincing them that, you know, you’re you’re the good guys in this, you know, they like as opposed to the people who’ve been strip mining the internet for data. and you know, pope I’m both curious about the general sort of cultural

[17:46] Nicholas Nadeau

So.

[17:47] David S. Bennahum

Yeah.

[17:59] Pete Pachal

stuff you’re seeing, but also like what is the the thing in the value prop that that they can look at the most that that is like this is different.

[18:11] David S. Bennahum

Well, maybe let’s begin with an assumption, which is if you’re somebody who’s created a large body of knowledge proprietary to you, right? Your knowledge, and a fair amount of that has been put on the internet, right, which is pretty typical for a person on the frontier of knowledge, you’re pretty well educated in being upset about AI. Right? Because one of the things you did is you ego tested yourself on on Anthropic, on Claude, and on GPT, and one of the things you probably discovered pretty quickly was it took all your stuff and you’re just like outraged.

[18:21] Nicholas Nadeau

Okay.

[18:40] David S. Bennahum

You’re not e even necessarily a technically proficient person. It’s not even related to tech savvy. But maybe you’re flattered initially. Let’s these are the stages of grief, right? And the very, very first moment may in fact pe be flattery, like when you googled yourself for the very, very first time and were like, I exist, I’m on the internet. fine. I exist, I’m in GPT, fine. And then after a little while, you’re like, What the frick is going on, man? How what

[18:41] Nicholas Nadeau

Sorry.

[18:43] Pete Pachal

So some people might have been flattered, but anyway, go on.

[18:50] Nicholas Nadeau

So.

[19:07] David S. Bennahum

This isn’t right. And then there’s a sense of resignation that kicks in. They’re so big, I can’t do anything about it. Life is not fair. Or maybe I’ll try to sue them and they’re so big, they don’t care. Life is not fair. at which point you become very open to taking that headwind, so-called headwind, and making it into a tailwind, which is like, how do I get out of this box? And so for us at Onix, one of the preconditions we see with every expert we talk to is they’ve been through this journey of sadness and rage. And then when they talk to us, they’re pretty educated about their sadness and rage, and they ask us pretty smart questions right away. And we’re prepared for them. And so really one of the big messages here is take back control of your genius. what does that mean? It means take all the stuff that’s been put online that they stole without your permission, the bootleg version, put it into the proprietary model we’ll give you. And then since it’s the authorized version, not the bootleg. You have an incentive to put in there the stuff that they couldn’t have stolen, because you didn’t put it on the internet, but it’s germane to your depth of wisdom. So it’s the stuff Nick had mentioned, your personal notes, maybe even audio recordings of your thinking that you make exclusively for your AI to train it further. And when you do that, the fidelity gap between what I would literally call the bootleg, which is what you get on GPT, and the authorized, which is what you get on Onix, is sufficiently large that then you can look to your audience and say, Do you value me enough to pay for me?

[20:34] Pete Pachal

Hmm.

[20:34] David S. Bennahum

If the answer is no, then use the bootleg on GPT. If the answer is yes, because what you’re working on is so challenging, and you actually do respect my expert opinion, then the ten dollars, twenty dollars, thirty dollars a month to use my wisdom is a no-brainer. And welcome to the authorized version. And the good news, Pete, is we saw this happen in music, right? 2001, Napster steals everything, LimeWire has it, music business says it’s over, we’ll never make money again as a recording artist.

[20:35] Nicholas Nadeau

Okay.

[21:04] David S. Bennahum

An Apple computer said, What about this thing called iTunes? What if we create a way for music to be elegantly downloaded and synchronized, pay ninety nine cents a song, high quality streams, ease of search, ease of curation. Optional. You can still go on The Pirate Bay, steal the music. And the reality is music piracy is no longer an issue today, even though it’s just as available as it was in 2001. We think the same phenomenon is going to happen with AI.

[21:28] Pete Pachal

Mm.

[21:31] David S. Bennahum

Sure, it’s really easy to go on Claude and ask for the bootleg version of so and so, but when the authorized version is out there and is demonstrably superior to the bootleg, welcome to the new world of AI.

[21:31] Pete Pachal

Right.

[21:42] Pete Pachal

Sweet. So I’d love to talk a little bit about the first area you want you went into, but also like where it could go further. Because obviously we already talked about health and wellness. That’s where you’re starting. but you know, when I first talked to you guys, I was thought about the model that you have and just thinking about it sort of writ large, scalable, so that people with domain expertise, people with a body of work, might want to just simply create one of these and you know, obviously thinking selfishly of my own profession, journalism and media creators, but of course that’s what I cover. give me some thoughts on like how scalable this is, because I also know part of your value prop, at least currently, is that you don’t you’re not exactly flooding the zone with like experts and have you know, tons of people in every field, it’s it’s it’s more curated than that. So as you expand, how are you going to balance that curation with regard to like making it scale? And also like areas like creators, like, like independent journalists, like you know, even even comms in independent comms people, like that sort of thing. Like, how are you thinking about that?

[23:01] David S. Bennahum

Yeah. But we thought about it initially as we had to demonstrate quality over quantity. Like we have to demonstrate that the quality of the AI, meaning the experts you’re summoning, can outperform the frontier. And at the same time, we felt a dense network of affiliated experts with high quality would be something to really demonstrate fully for all the reasons around orchestration of knowledge. So when you just take that thesis.

[23:08] Nicholas Nadeau

So, thank

[23:31] David S. Bennahum

You’re gonna sacrifice the other verticals and you’re gonna potentially just go to one. And the one we went to with went to was health and wellness. And again, we did that because when we really looked at the domain opportunities, health and wellness scored highest on where expert opinion really matters. Like I wanna have the perspective of this cancer doctor, not that cancer doctor, or this nutritionist because I’m vegan, or that nutritionist because I’m paleocarnivore, right? These are just Fundamentally reasonable discernments around different types of expertise, what we might call expert opinion. Really matters. Number two, privacy really matters. The health and wellness domain tends to be a very intimate, personal one. It’s different than say, what should I watch on Netflix tonight? That’s a different domain. It’s also private, by the way, but it’s not the same level of privacy as I’m having maybe IVF treatments and going through fertility.

[24:07] Nicholas Nadeau

So, you.

[24:25] David S. Bennahum

And I want to talk about that with an expert AI, that’s a different level of privacy than maybe which sitcom should we watch tonight, right? And then number three, trust being paramount. And so if you don’t trust that the expert has a high quality AI that really represents their knowledge, and if you don’t trust this conversation is going to be private and my data is private, then the whole thing falls apart. So health and wellness is a really high trust domain. So if you look at those boxes. goes right to the top. And then there’s other verticals adjacent and beyond that we believe we can go to over time. So things like personal finance is a high trust area too. Not quite the level we think is health and wellness, close. There’s things like relationship and self-improvement. So that’s all the domain of maybe I want to be an entrepreneur or I’m struggling with something in my marriage or give me advice on something pertaining to how to be the best, most optimized version of me. It’s very close to wellness, right? You can see that adjacency. And then we can go further from there. There’s lifestyle expertise. But again, that’s a little further out. So I’m thinking of traveling somewhere, and I want a really great expert to advise me on where to go in Sri Lanka, because I’m gonna do Sri Lanka. Okay.

[25:38] Pete Pachal

Yeah.

[25:40] David S. Bennahum

Again, an interesting area. And so you could just go on here, at which point I think there’s a second part to your question, which is to what degree do you kind of open this up so that people can just become creators and come on the platform? And the answer is yes. And think of it happening in three phases. Phase one, where we are today, invitation only. We work together to build your Onix. Once you, meaning the expert, and we are aligned, it’s fit for production, it goes live. Again, health and wellness being that initial domain. Phase two. You have the ability to create an Onix, share it privately. Right? Privately, which means it’s not in our discovery environment, but it’s available. That is happening soon. Like we’ve built the technology. We actually have some people testing it now. We think that’s very, very interesting. It will teach us a lot around the kinds of things people want to do. Phase three is take that private Onix and submit it to us for listing in the marketplace with a price.

[26:23] Nicholas Nadeau

So, thank

[26:24] Pete Pachal

Mm.

[26:39] David S. Bennahum

‘Cause you want discovery, you want revenue. Then we’re gonna have to judge it in some way. And that’s where we’re gonna have to start getting into like with any app store and a little bit of a discernment, right? Why this, why not that? That would be phase three of the process.

[26:53] Nicholas Nadeau

Yep. Yeah.

[26:53] Pete Pachal

And can every every creator or or person on your platform set their own price? Could be sort of whatever they want.

[27:00] David S. Bennahum

Yes. There’s a minimum price, obviously, because there’s a certain cost of goods sold. Yeah, there’s inference, there’s compute. So we think below about there’s no free well, there’s no free tier if you just want to pay for it directly. This is a whole other conversation. But let’s just say for now in the current framework, around ten dollars a month would be the logical minimum. There’s no ceiling. You know, if someone said, Listen, you know, I have a very limited edition Onix, I think it’s worth several thousand dollars a year.

[27:03] Pete Pachal

Sure. So there’s no there’s no free tier.

[27:13] Pete Pachal

Right.

[27:28] David S. Bennahum

And that’s on you. And if you feel like you can make that work, by golly, we’ll be delighted to be your partner and you should totally do it. Not on us to say.

[27:37] Pete Pachal

Nice. And as I understand it, you have this sort of seventy, thirty cut, right, in terms of just this how that’s how you guys make money.

[27:43] Nicholas Nadeau

Yeah.

[27:44] David S. Bennahum

Correct. So it’s essentially like an app store commission, right? That we take thirty percent of the revenue. If you don’t make any money, we don’t make money. So we are we often say to the experts, we’re business partners. You’re we’re not a you’re not like we’re not like a vendor and you’re like a client. You know, we don’t charge you to come on the platform. There’s no setup fee. There’s no charge to come on. What we’re really doing here is saying let’s work together to build a million flourishing businesses on this platform. So having said that, if you don’t succeed, we don’t succeed. We have an incentive to promote and market you. And you obviously have an incentive to promote and market yourself. Similar to Substack in that way. but ultimately if you don’t if you don’t make money, we don’t make money.

[28:24] Pete Pachal

I like I liked how you described almost I don’t know if this is how a stand actually I’ve that’s my question. Is it kind of a standard? ‘Cause you said like, is what’s coming out of your Onix better than what you get out of like a generic AI as kind of a bar, right? Like it’s kinda like, Well, that’s that’s I think it’s a good sort of measurement. I’m just curious, like how do you measure it? Like it’s got sounds a little more qualitative than quantitative, or is that how you think about it or

[28:47] Nicholas Nadeau

Yep.

[28:49] David S. Bennahum

Well, I’m gonna let I think Nick should actually talk about how we do the evaluations because there is an element of of quantitative science here, which is in really important. And yeah, there’s a qualitative element too, which is does the expert ultimately feel like their Onix is awesome? That’s a qualitative call. But I’ll let Nick talk talk about the quantitative piece maybe a bit.

[28:55] Nicholas Nadeau

So I’d actually start with the qualitative. It’s something when the true expert, the person who knows themself and their expertise the most, they battle test their Onixes. The fact that we have 25, 30 experts live public on the platform, they have

[28:57] Pete Pachal

Nice.

[29:24] Nicholas Nadeau

battle tests of these Onixes with them, their teams around them, the people they know best, even sometimes our clients and patients have battle tests of these Onixes before they go in public. And they say, wow, this is exactly me and my expertise, but now at scale, at mass personalization. On the quantitative side, you know, this is where evals come in. So evaluations, what you typically see in across all the different benchmarks, you know, we work with a lot of the common evaluation frameworks and whatnot. We also started publishing, we said we have two papers out on how we do our evaluations. And there’s really four pillars. There’s accuracy, groundedness, persona, fidelity, and judgment. And so this, we come together and call it the expert fidelity score in a lot of ways, where is it actually representative of the expert? Is it grounded in their work? Is it accurate to their work? Is it reflecting their persona, how they approach a problem, their style, their tone? you know, a happy-go-lucky doctor versus somebody who’s a little more straightforward and blunt, and then also their judgment. And so this is where I get into not only just, you know, one-off analysis, but what is the multi-turn approach and nuances? And so we had one of our second paper that we published, one of our awesome experts, Dr. Dave Rabin, psychiatrist and neuroscientist, he takes a very different perspective than the statistically average view on neuroscience. he looks at things like, and so it came up that he doesn’t subscribe to what is the psychology 101 polyvagal theory. He actually explicitly rejects that theory as part of his practice, but nowhere has it been written down that he doesn’t subscribe to that. And so if you’re asked, know, ChatGPT or Claude or whatever, say, pretend you’re Dr. Dave Rabin and tell me your thoughts on this, it will reach out towards what is the Wikipedia definition of stress response in the body, polyvagal theory, not knowing that he personally rejects that. But working with us, working human-in-the-loop to build the model together, these are the types of dark data that come into making his model not only a higher fidelity, but his. Not some other big AI companies. This is his model that is verified, validated, stamp of approval based on his work and his approach and his judgment and nuance. And that’s what expertise is.

[31:43] Pete Pachal

And speaking of like measuring things and quality, one of the things we haven’t really touched on too much is hallucinations, even though, you know, I know sort of by implication you have smaller models and there’s probably a lower rate. But I mean, it’s all generative, so like it can’t be zero. So how do you warn against that, think about that, and and measure it and and sort of mitigate it as a, you know, because all AI systems have to deal with this at some point?

[32:09] Nicholas Nadeau

Yeah, so as you hit it out, generative AI as a whole is a probabilistic process. There is literally no way to prevent to a 0 % chance hallucinations. But as you also mentioned, small language models, domain constrained, really isolate to anthropomorphize a bit. These models think in their domain and their domain only. And when you start to cross domains, it’s outside of their trained body of knowledge. So they have nothing to really reach towards. Whereas if you take a, you know, a God model, a giant model, it’s been trained on everything under the sun from 4chan conversations to Wikipedia to, like I said before, like the history of Japan and what else. And so it’s always going to try to reach for things to fulfill the requests of the user, because not only are the things trained on, you know, all the internet, but they’re trained to be pleasing bots that always have an answer for you to maintain engagement and keep the conversation going. And so think about the incentive structure underneath this is that You have a system where engagement is the metric. Well, everything feeds that metric to keep on answering you. We have a system where wellness and empowerment and the expert themselves, the human version, the in-real-life version is the metric that we’re optimizing towards. And it’s a very different line of thought, which leads to even how you build the system from the ground up. What is the architecture going into this from a business model, from a technical perspective that goes into that. And then. You know, like all other tech companies, have guardrails, have evaluations, we are nonstop spending our time comparing, contrasting ourselves to what is frontier, how do we do it? Have we done, have we thought of this, how we thought of that? Corner case is getting feedback from users, feedback from experts. And that’s why it’s a bit of a community approach with experts in the loop that we’re always getting this feedback, tweaking, fine tuning, and making sure that’s best representative of them and best representative of what the community needs.

[34:03] Pete Pachal

corner cases where two edge cases meet. I don’t often I don’t often hear about that term, but yes, I’ll I take a point. listen.

[34:06] Nicholas Nadeau

Hahaha! Yeah, well, hang out with my developers and you’ll hear it a lot more.

[34:11] David S. Bennahum

Tonight. May I add one thing, Pete, which is the AIs will on on our side will actually say I don’t I don’t know. And it’s awesome.

[34:21] Nicholas Nadeau

Don’t know. But that’s what’s fun is you talk to a nutritionist model and you start talking about sports medicine recovery. And so I’m like, well, I don’t know that, but I know somebody who does. Let me summon Dr. Jordan Metzl into this conversation. that’s like one, it reaches for its, that’s to be discussed of how we.

[34:22] Pete Pachal

Nice. Very smart.

[34:37] Pete Pachal

Very nice. Just just another twenty nine ninety nine. First answer’s free. Awesome, guys.

[34:41] David S. Bennahum

Yeah, but that’s about right. Yes, let’s support let’s support authors, you know, economically.

[34:49] Pete Pachal

We’re almost out of time. I just wanted to hit you with sort of a variant on my normal final question. so this is for both of you, but like if you look out, say five years from now, and what what would be the outcome that would convince you that your view, your basically your point your your I don’t know if it’s a model or what you would call it, but basically what you’ve put forward in the realm of expert owned AI, the the author. economy you might call it. How how will you know it’s healthier? Or what is the outcome that would make you think it something had gone wrong? go ahead.

[35:27] David S. Bennahum

Yeah. So I think one of the big signals of health is that the value being created by AI has distributed outward to more and more people economically, where you can look to many, many people having flourishing economic experiences thanks to AI as experts, that would be a real signal of success. And a signal if something went really, really wrong is at the end of the day, a handful of companies controlled.

[35:28] Nicholas Nadeau

So, have a lot of things to do. I think we lot of to do. I think to do. I we have of things I lot to I we have a lot things to lot to

[35:57] David S. Bennahum

all this knowledge and reaped all the economic benefits to the detriment of everybody who wasn’t in that inner circle. It’s a very simple arithmetic that’ll point to a source of truth.

[36:11] Nicholas Nadeau

I’m going to go with, I’m going to extend David’s a bit where in, I think you said like five years, a few years, we’re no longer talking about IP theft or data scraping because there is a better model, just like David’s analogy to the iTunes model. We don’t talk about Napster LimeWire, even though they technically still exist in a lot of ways. The other side is like the socioeconomic access to expertise. If we achieve our mission, expertise is fairly democratized where, you know, what does it look like when the whole world has access to a nutritionist or nutritional support? What does it look like when people have this expertise in their pocket to help them flourish? What kind of, you know, what kind of society looks like now when it goes wrong is, you know, what happens if it only supports one cohort and doesn’t get democratized, know, that’s the worst case.

[37:05] Pete Pachal

Awesome. Great stuff. We’ll just go for it.

[37:05] David S. Bennahum

May I add one small thing, Pete, if I could. I would like to let you know that we have an access code for all the copilot listeners. We’re currently in early access. If you want to try Onix, you can go to Onix.life, which is our website, Onix.life install for iPhone. And then when asked for an access code, just type in copilot. And once you type in copilot, you’ll go in and you’ll enjoy experience the product.

[37:07] Nicholas Nadeau

Thanks.

[37:23] Nicholas Nadeau

Stuff.

[37:34] Pete Pachal

Good stuff. Thank you guys. I really appreciate you coming by, giving me your the four one one on the author economy and what you’re doing to augment it.

[37:44] Nicholas Nadeau

Thank you, Pete. Thank you for having us.

[37:45] David S. Bennahum

Thanks, Pete, for having us. Yeah, great to see you.

The post Who owns your expertise in the age of AI? appeared first on The Media Copilot.

By The Copilot

What if the future of AI in media isn’t replacing storytellers, but giving them better tools to tell better stories?

This episode is sponsored by Descript.com 

In this episode of The Media Copilot, host Pete Pachal sits down with Laura Burkhauser, CEO of Descript, to explore how artificial intelligence is reshaping media production without sacrificing creativity, editorial judgment, or human storytelling.

Laura explains why Descript’s vision has never been about replacing editors, but about eliminating the repetitive, time consuming tasks that slow production. From text based video editing and AI assisted collaboration to intelligent clipping and personalized workflows, she shares how AI is becoming a creative partner rather than an automated replacement.

The conversation also examines why memory, personalization, and collaboration may become AI’s greatest strengths, how media organizations are adapting to an increasingly creator driven landscape, and why the next generation of production tools will focus less on software features and more on helping teams create great content.

Finally, Laura offers a refreshing perspective on the future of AI, arguing that the industry has spent too much time selling fear instead of showing creators how these tools can unlock more original storytelling, more creative freedom, and entirely new ways of producing media.

The conversation explores:

• Why AI should replace repetitive labor, not human creativity
• How Descript evolved from a text based editor into an AI powered production platform
• Why collaboration remains the most valuable feature in modern media workflows
• How AI memory is creating more personalized creative assistants
• Why editorial judgment, taste, and storytelling still belong to humans
• How newsrooms and media companies are producing more content with smaller teams
• Why the future of media production is conversational, collaborative, and AI assisted
• How creators can use AI to spend less time on tools and more time telling better stories

Why this matters

Artificial intelligence is transforming every stage of media production, but the biggest opportunity isn’t replacing creative professionals. It’s removing the friction that keeps them from creating. As newsrooms, marketing teams, and independent creators face growing pressure to produce more content across more platforms, AI has the potential to become an intelligent collaborator that accelerates production while preserving human judgment, creativity, and authenticity. This conversation offers valuable insight for journalists, marketers, content creators, media executives, and anyone navigating the future of AI powered storytelling.

🔗 About the 👤 Guest

Laura Burkhauser is the CEO of Descript, the AI powered audio and video editing platform used by creators, media organizations, podcasters, and enterprise teams around the world. Under her leadership, Descript continues to pioneer AI assisted media production, helping storytellers simplify complex workflows while keeping human creativity at the center of the creative process.

LinkedIn: https://www.linkedin.com/in/burkhauser/
Descript: https://www.descript.com

This episode is brought to you by Descript.

Descript makes editing podcasts and video as easy as editing a document. Record, transcribe, clean up audio, remove filler words, create clips, and more with powerful AI tools built right into your workflow.

Spend less time editing and more time creating with Descript.

Visit Descript to learn more.

About the show: To explore more conversations like this and see what’s new, visit the Media Copilot website at mediacopilot.ai. You’ll find new episodes, expanded resources, and tools designed for journalists, communicators, and media leaders navigating the fast-changing world of AI. It’s the home base for everything Media Copilot and it’s just getting started.

Enjoyed this episode?

Subscribe to The Media Copilot on Substack, Apple Podcasts, Spotify, or your favorite app. On YouTube? Tap the Like button and Subscribe to the YouTube channel. For more AI tools and resources built for media professionals, visit mediacopilot.ai.

Produced by Pete Pachal and Executive Producer Michele Musso
Edited by the Musso Media Team 

Music: “Favorite” by Alexander Nakarada, licensed under CC BY 4.0

All rights reserved. © AnyWho Media 2026

THE MEDIA COPILOT TRANSCRIPT

THE MEDIA COPILOT

AI Won’t Replace Creators

Featuring Laura Burkhauser, CEO of Descript

[00:22] Pete Pachal:

Hi, welcome to this sponsored episode of The Media Copilot, presented by Descript. I’m your host, Pete Pachal. I covered tech for a long time as a journalist, and now I have deep conversations with the media people, the builders, and the creators who are all trying to answer the question: How will we get information in the future? And how will that change the jobs and the industries whose business is information, especially media?

My guest this week is Laura Burkhauser, CEO of Descript.

Every media team right now is trying to answer a set of questions that are straightforward, but the answers are increasingly difficult. The main one is: How much AI should we bring into production?

That question alone leads to many more questions. Which parts of the workflow are safe to automate? How do we use these tools without flattening voice, lowering quality, or losing editorial control? And how do we avoid stacking one vendor on top of another just to record, transcribe, edit, clip, caption, summarize, and publish a single piece of media?

That last part really matters. A lot of newsrooms, comms teams, and branded content teams aren’t suffering from too few tools. They’re suffering from too many. One tool for recording, another for transcription, another for editing. The list goes on.

The promise of AI is supposed to be that you can go faster, have better quality, and have more leverage. But in practice, it can also create more complexity, especially if your workflows aren’t entirely thought out.

That’s why I’m excited to talk to Laura today. Descript is one of the companies trying to rethink what media production looks like when AI is actually built into the workflow rather than just bolted on afterward.

So this isn’t a conversation just about one product. It’s a conversation about how audio and video production are changing, how bigger teams should think about workflows, and how the interface for making media may become less about individual tools and more about what teams want to create.

[02:30] Pete Pachal:

We’re going to talk about where Descript fits into that shift, what newsrooms and media organizations should know if they haven’t used it or looked at it in a while, and what responsible AI looks like when the output isn’t just that one single piece of content.

Laura Burkhauser, welcome to The Media Copilot.

[02:46] Laura Burkhauser:

Hi, Pete. Thanks so much for having me.

[02:49] Pete Pachal:

Sure, it’s my pleasure. Let’s get into it. Let’s get into Descript.

As I understand it, it started as a very specific idea: editing audio and video via text. Now it feels more like a full production system. How would you describe the company’s evolution, and what has changed the most since generative AI became central to the product or integrated into the product?

[03:12] Laura Burkhauser:

Yeah. So Descript has been AI native since we were founded several years ago, and that’s because the concept of editing video like text is itself AI. Editing video like a document is itself AI.

The way that Descript works fundamentally, or our core feature several years ago, was that you record something like this. It’s going to be a 40-minute-long video podcast.

In the old world, you would go into a timeline and do a lot of cutting and splicing, a lot of work on audio, and cutting out all of my filler words. Well, many of them, maybe not all of them.

If you talk to sound engineers, they’ll tell you, “I know what that looks like on a timeline. I could draw it for you.”

[03:55] Pete Pachal:

And manually doing it too, literally putting little spaces and stuff.

[04:08] Laura Burkhauser:

Exactly. So Descript said, “Well, what if you could edit things like you edit a document?”

That doesn’t just mean deleting filler words or deleting things like retakes, but also copying and pasting and moving things around. That was kind of the initial idea.

Then Descript acquired a company called Lyrebird, which did AI audio generation, and brought that in as a research team to bring more AI into the product.

If you think about the things you do with a document, you can cut, you can copy and paste, and you can also write. We were one of the first products to provide text-to-voice, where you can write new paragraphs. Or, in my voice, if I use the wrong name or you pronounce my name wrong, you can actually regenerate it without having to re-record.

Those are some of the original features that we went to market with, and we immediately achieved product-market fit, especially in media.

Some of our longest-running customers are actually customers in some of the biggest, most noteworthy newsrooms that you know. That’s because, in addition to editing video like a document, we also had collaboration from the very beginning.

It’s like editing video the way you would edit a Google Doc, right? Someone in your newsroom may know the hell out of the story, but you have someone else who’s really a legal expert and understands the things that you can and can’t say. You have an editor coming in and checking the work. You have a copy editor coming in.

They all bring a different amount of expertise into this rough cut, eventually fine cut, and eventually polished step that you have toward making video.

Collapsing all of that collaboration together into a document that everyone can see, and allowing people to do things like paper edits, tremendously shortened the amount of time it took to get from recording to publishing.

That’s the foundation upon which Descript was built and why I think, to this day, we are in most newsrooms. In some, we’re wall-to-wall. Every single person on the production team has a Descript subscription.

Now, you asked what has changed because technology has changed so much since that initial promise. The biggest thing is LLMs.

[06:34] Laura Burkhauser:

What’s great is, if you’re able to edit video like a document and then LLMs come along, what are LLMs incredibly good at doing? Editing text.

Well, we understood all video as text. So having an LLM come in and be able to edit text really well kind of gives you multimodal editing for free.

[07:01] Laura Burkhauser:

Then adding captioning to all of our videos, so you know what’s happening on the screen every few frames, gives you visual understanding.

A dog walks by, and we’re able to say, “dog walks by,” and store that as text. That gives you visual understanding in a real multimodal way before you can build a multimodal model, which takes hundreds of millions of dollars and hundreds of researchers.

Instead, we were able to get that to you early, a couple of years ago now, which was way ahead of the times.

Then last year, we built an agentic co-editor, Underlord. Because nobody needs an AI overlord, but couldn’t you use an AI Underlord?

Underlord can come in, take direction from you, and take the first pass at editing.

[07:45] Pete Pachal:

Hope it doesn’t ever vie for a promotion.

But that all sounds really fascinating. The bets that you placed early ended up really paying off in the AI era because of how language models work.

Are there specific features that editorial teams have really zeroed in on and that explain why it ended up being popular for them?

I imagine it probably varies a bit by the type of newsroom, whether you’re more broadcast versus publishing, but give me some thoughts on what really resonates with that crowd.

[08:20] Laura Burkhauser:

Originally, text-based editing itself was huge, and collaboration, which isn’t an AI feature, but it’s just a great feature. Being able to all edit at the same time, leave comments, do approvals, things like that.

But then, as video repurposing became more important, turning one piece of media into the 17,000 pieces of media that you need to feed every single content maw across all of the publishers, using us for clips has been really exciting for people.

We have a button that creates clips, but clips is actually the feature that forced us to build Underlord.

When we released clips, we had a bunch of different parameters you could set. You could say, “I want this to be vertical. I want the clips to be about 30 seconds. I want you to use this layout that I’ve already created with my brand stuff and look for clips about this subject.”

Those were the parameters you could add in.

Immediately, people had a bazillion other parameters they wanted to add. They were like, “I want to create clips for every single one of my channels at the same time. When you do it, I want you to actually create a social post about each of the clips at the same time. I want you to only choose clips where my guest is speaking. Don’t include me in the clips that you’re adding.”

And I’m like, well, this is so much direction that at this point we’re not talking about buttons anymore. You need a little Underlord to yell at.

That kind of led us to create this idea of Underlord, which I think people really like.

Now Underlord has memory. So you can work with Underlord once and say, “Look, when I say go out and create clips, what I actually mean are all of these types of clips, along with this definition of a publishing packet that I want you to create as part of that.”

Now every time I tell you, “Go make clips,” I want you to do those 50 different jobs that we got right once.

That totally changes your workflow for some people.

You’re right, it doesn’t work for everyone. But you can really have Underlord do the first pass at everything. Then you go in and act as a reviewer.

I think for a long time, and probably honestly forever, when we’re talking about things where voice is really important, you go in and change and shift things and add in the layer that makes it you.

But Underlord can do the perfunctory things that you might hire an intern to do.

[10:44] Pete Pachal:

Nice. So, yeah, clipping is obviously huge these days, and it sounds like that’s sort of the core of maybe your value proposition to a lot of these media companies.

But you tell me. When you boil down why someone would use your tool versus something else or a stack of things, is it that multimedia capability? That you can make more from a single piece of content more easily, both in terms of speed and maybe in terms of the intelligence behind how it does it?

[11:21] Laura Burkhauser:

It really depends on the customer.

In actual newsrooms, if you ask people what they love about Descript, it’s not one-shotting editing jobs with generative AI. It is a fully functioning collaborative editor that uses AI to speed up the parts of your job that you don’t like.

My vision for newsrooms isn’t that you should just go from recording to one-shotting everything.

A lot lives in the editorial decisions that go into the rough cut. If you think about a long-form audio podcast that you love, the obvious one might be something like This American Life, but there are a million now.

There’s so much audio that’s not just a straight show like this. You and I are doing a chat show.

[12:18] Laura Burkhauser:

Editing this is going to be pretty easy for your editor, and a lot of the magic comes from the interview that you and I are doing.

I don’t know how complex you get in your editing job, but my guess is this is a pretty quick editing job. Speed is what I would use to sell this to you. Speed and repurposing all the clips is what I would focus on if I were selling to your editor.

If I’m selling to NPR and they’re doing a lot of really high-quality, prestige podcasts or video, there I’m talking about the depth of the editor. I’m talking about having your whole team in there.

Yes, we have AI features that make things way easier to do and save you hours and make you able to create a daily show instead of a weekly or monthly show.

But every story I’m telling is about humans doing a lot of work in the editor. And I think that’s because, to create really high-quality stuff, you just need humans doing work and making decisions.

That’s just something I stand by.

[13:13] Pete Pachal:

Right. So you have that inherently collaborative nature of the product, and now you have Underlord, which is kind of a collaborator too.

[13:20] Laura Burkhauser:

That’s exactly what I think about it as.

Underlord is another editor in Descript with you that is able to do some editing and has its own area of expertise, which is not the same area of expertise that you might have, that Steve from Legal might have, or that your copy editor might have.

Thinking about Underlord not as someone that you’re delegating work to in all cases, I think for some people it can be that, but in newsrooms, Underlord is more like a collaborator.

[13:52] Pete Pachal:

I’m glad you brought up memory. I’d like to chat a little bit about today’s models and how the features and capabilities overall are really accelerating.

I’d like to know, not just what helps Descript, but broadly what you’re excited about as things are progressing.

Memory, of course, has been around for a while, but it’s been rapidly evolving. That just seems like a key way things will evolve.

We think about memory as personalization for you, but in the case of a tool, it’s personalization to the workflow and someone’s rhythm. The way one person uses the tool might be totally different from someone else’s.

Talk a little bit about memory and how the new models are evolving. What excites you the most?

[14:52] Laura Burkhauser:

I think memory is really exciting, and I’ll tell you a little bit about how it shows up in creative software like Descript.

One of the fundamental criticisms of AI is that AI is, at its heart, a derivative piece of technology.

The way that it works is it does a good job of predicting what might be a good thing to do. And the way that it predicts is it trains on a huge amount of data.

That is a terrible way to create, right? What if everything you created was fundamentally derivative?

To some extent, I guess you could say you go through life and take in all of this data and have these experiences and work on a voice. And out of that comes, we hope, as creators, something that’s interesting enough that it stands out.

[15:50] Laura Burkhauser:

AI on its own will never do that because it’s fundamentally derivative.

However, what memory can do over time is begin to pick up on some of the nuances of the way you work, the context that you’re in, the workflow, and how you make things.

It stops being this regression to the mean of the most boring way to edit a podcast, tell a story, write a document, or paint a picture.

It can understand Laura’s voice, Laura’s priorities, the beats that Laura loves to hit in her stories, the importance of authenticity and humor in the way that she comes through. That might make you make different editing choices than you would make for Pete and his storytelling.

When Underlord can get a sense of your voice and your taste and your perspective, I think you start to get more excited about collaborating with it, the same way you would with a real editor whose job is to help your voice and your vision shine.

That’s why you see the same creative teams work together over and over again.

Great, your collaborator gets it now. You get the vision, you get the context, you get the kind of thing I’m trying to make. Let’s go make four more of these because you get it now.

That’s the excitement of memory for me in creativity.

[17:18] Pete Pachal:

Nice. It sounds like, in terms of it being a collaborator, it’s not even necessarily a collaborator just for people who are professional editors.

The whole text-based editing concept is a bit of a mental shift, but it’s making video editing capability, and just broader multimedia capabilities, possible for people who aren’t professional editors.

[17:44] Laura Burkhauser:

Exactly.

[17:47] Pete Pachal:

How do you think about that in terms of the ongoing discussion in society around AI? Is it here to help us? Is it here to replace us?

Does it help make teams more lightweight? Does it support them? Does it move certain lightweight tasks onto nonprofessionals? How do you think about all of that?

[18:15] Laura Burkhauser:

I think we need to absolutely throw out this concept of AI replacing humans, wholesale. I don’t want to hear about it anymore.

I’ve been saying this for a while. I think now it’s finally catching some steam, so I feel less sheepish saying it in San Francisco.

I do think AI will replace some labor, which is very different from saying AI will replace humans.

The way that I think about it, the way that I’ve worked with my team on it, and the way that I would encourage creative people to work on it, because you will come up with different answers than what me and my team came up with, is to list out the labor that you do to get to the product that you make.

Do you like doing all of that labor? My guess is you don’t.

Let’s separate it into a couple of buckets.

In the bucket of labor you don’t enjoy doing, how do you think about delegating this to AI?

You can’t always do it. I too would love to delegate putting away my clothes to AI, but unfortunately I can’t. That’s a human job, probably forever.

But how can I use AI to replace my labor with its labor here?

Then, for the stuff that I enjoy doing, how can I use AI to enhance my ability to do this labor?

[19:41] Laura Burkhauser:

I think there’s an enhancing bucket and a replacing bucket that probably follows patterns along different functions, but isn’t the same for every single person in the same function.

AI can help on both sides of the equation, but we have a choice.

We have a choice as individuals. We have a choice as organizations. We have a lot of choices for how we use AI.

I encourage everyone to use AI because I think it is foolish to say, “Hey, we have this thing that could make you so much more productive, and you’re just not going to use it on principle.”

To me, that feels like, “I don’t own a computer. I prefer a typewriter.”

Sure, as an affectation, that’s hilarious, but come on. No one actually thinks it’s a better way to work.

I think AI is going to be part of every non-affectatious life.

[20:39] Laura Burkhauser:

But I think how we use it is really going to depend on the function and on the person, to some extent.

And I don’t think it’s going to replace a single human.

[20:46] Pete Pachal:

Right. It’s more like you say, labor, tasks, whatever the word is.

To my knowledge, and I don’t have the data in front of me, the data is kind of supporting that.

I would be curious, though, about what you’re seeing, whether it’s data or anecdotes or whatever, with regard to the organizations you work with.

As they use your tool over time, do you find they’re doing more things that they weren’t able to do before? Do the teams start to get reorganized a bit?

What happens over time as people get acclimated to both the tool and working with the AI features?

[21:36] Laura Burkhauser:

Yeah, that’s a great question.

First, the way that I would talk about Descript’s customers is that we work with media teams and marketing teams.

We work with media teams all the way from a solo podcast to a national or international newsroom.

We work with marketing teams all the way from a solopreneur who is creating content to market their coaching business to some of the biggest enterprises in the world.

We work with storytellers in media and marketing. Those are our main kinds of customers.

What we find that they’re able to tell us is that, because of Descript, people who normally cannot take part at all in the media editing or storytelling part of the equation are now able to help.

But really, the story is that we can do video where we couldn’t do video before.

We hear from places saying, “We stopped doing video as a team because it was just too much. It was too much work. We were doing all text and maybe some audio, but no video.”

Now, we can do video as a content type.

Or we moved from doing video once a month to being able to do a daily show.

If we’re talking about an actual newsroom, doing a daily show is insane. The fact that you can do multiple daily shows and just have one or two editors working on that is unthinkable in the podcasting world of a decade ago.

And that’s because of Descript.

We can do it more often. We can do video where we couldn’t. We can do video more often.

[23:32] Laura Burkhauser:

And we can do video across many more channels.

We can do long-form video and all of our short-form clips. We can do video that is tailored for all of these different channels where we need to go out and meet the customer because they aren’t necessarily coming to our site anymore.

What I actually see is a lot of media companies moving toward a world where, yes, they have their own site and their own channel, but they also have a network of creators that you may never know is part of that media conglomerate.

They work with all of these podcasters or creators who are monetizing the way creators monetize, through things like sponsorships and YouTube ads, and the media company gets some portion of that.

There are these new economic models being built on a creator economy that depends on being able to create video much faster and across many more channels.

[24:32] Pete Pachal:

That sounds really interesting because it dovetails with a lot of media strategy right now. It’s taking advantage of this creator economy, the systems, and the ways that creators execute.

It’s an interesting merger. Obviously, like you say, creators are doing it, but even the bigger companies want to take that approach.

[24:57] Laura Burkhauser:

They absolutely are.

That’s critical to more traditional media companies surviving in the age of new media.

You want to keep the brand recognition that you have. You want to keep the valuable asset that you have. But you need to start going out and meeting people where they are, across all of the channels where they’re consuming news, which is not necessarily your site.

[25:19] Pete Pachal:

Of course, bigger companies, when you get to the enterprise level, have a set of concerns around liability and security and all these other things.

Has that been a challenge at all, trying to address all those things they want to do, like SSO or whatever?

[25:36] Laura Burkhauser:

Every startup loves going enterprise. It’s everyone’s absolute favorite thing to do.

As a product manager, it is always hard for me to say, “Yeah, I don’t want to go build agent memory because I would so much rather build SCIM,” which is a permissioning solution for companies that need to let some teams have access to some things and some teams not have access to other things.

If you had asked me two years ago, I would have been pulling my hair out about that problem.

I think we’ve mostly crossed the chasm there and are able to offer mature enterprises all of the wonderful features that they would love.

We’re also very careful about things like data retention. An enterprise company is obviously very private about their data.

Making sure that we’re able to meet customers where they need to be met when it comes to things like data privacy, permissioning, and all of those other wonderful enterprise features is important.

[26:49] Pete Pachal:

If you don’t mind getting into the weeds a bit, I’d like to ask about some of the features that ended up being surprisingly important.

As I edit things and do my amateur-hour stuff on podcasts or whatever, depending on your passion for the material, you can end up being kind of a perfectionist about things like chapters or show notes.

Particularly with AI features, the promise is always that it gets you a certain percentage of the way there and then the human takes it from there.

Has there been something where, surprisingly, people really needed it to be 95% there where you thought 70% or 80% would be enough, and then you worked to perfect that?

There’s a lot of iteration on some of these creator features, I imagine, depending on the feedback you get from customers. Is there anything that stands out or was surprising?

[27:56] Laura Burkhauser:

That’s a good question.

I’m always humbled by customers. It’s one of the things that I love about this job.

You have an idea of what people will use a feature for and how it’s going to land, and then the data always humbles you. The interviews always humble you.

I’ve been surprised by many features.

For me, maybe something that was surprising was when I was an amateur podcaster. I actually found Descript through podcasting.

I started a podcast with my friend, and we used Descript to edit it. I was like, “My God, I think I have to work at this company. This is the coolest thing I’ve ever used.”

[28:42] Pete Pachal:

You didn’t tell me you were a professional at this.

[28:47] Laura Burkhauser:

I am definitely not.

But one of the features that I loved was Remove Filler Words because, if you can’t tell, unless Pete has kindly removed all of my filler words, I use a few filler words.

[28:58] Pete Pachal:

I like to keep it real.

[29:16] Laura Burkhauser:

Exactly.

So I was removing all my filler words, and then as soon as I joined Descript, I found out that there is a huge camp of podcasters who are pretty against removing filler words and actually pretty against a lot of the automated editing, just-one-click, don’t-think-about-it kinds of buttons that we created, like removing pauses.

They’re like, “Look, when you get further into the editing game, you will realize the value of a pause and the value of a filler word, and of retaining the integrity of the sentence the way that the person created it.”

You don’t always do that. You never want to embarrass your guest. Do you hear me, Pete? You never want to embarrass your guest.

[29:50] Pete Pachal:

Wouldn’t dream of it.

[30:10] Laura Burkhauser:

You do want to get rid of some of them, but it takes judgment.

The more you talk to people who have been in the editing game for a long time and who are really working on S-tier kinds of content, the things that you love and have a ton of admiration for, the more you realize how much judgment goes into even something like that.

Which “ums” do you remove?

Automation can be really helpful for a lot of kinds of content. But that’s where I don’t lean a ton on automation when I’m talking to an experienced editor because it just doesn’t ring super true to them.

[30:34] Pete Pachal:

It reminds me of when Ben Affleck had that thought a couple of years ago about AI not having taste.

I wholeheartedly agree with everything you said there about integrity. Sometimes you want it out, sometimes you don’t, and that could even be in the same piece of media.

By way of transitioning to looking forward, do you think we’ll ever get there with AI?

Could it look at a conversation and think, “A pause right there would be great,” and then leave something in and apply a certain amount of taste to it?

This is getting pretty advanced, but who knows? New models come out every week.

[31:17] Laura Burkhauser:

I think yes. And that’s what we’re aspiring to build.

But again, I think that taste is not generic.

When we talk about taste, we talk about it as either you have it or you don’t. I think that’s true up to a certain bar, and I think we can get AI up to that bar.

But then the question is: Is it interesting?

[31:47] Laura Burkhauser:

Is it authentic? What is the taste? Taste to what end?

That’s where it’s all about who’s driving the car.

I do think it will get better. I think there are also going to be limits.

When it gets that good, will it get that good in a way that works with the business? Will it get that good in a way that is cost-efficient and works across all tools?

It’s not just intelligence that is the bottleneck here, but certainly the level of intelligence is a bottleneck.

I think taste is eventually a solvable problem, but it won’t fundamentally change the fact that taste plus an excellent human will still outperform generic taste every time.

[32:39] Pete Pachal:

Yeah, that makes total sense.

I’d love to chat a little bit about the general trends in AI these days. There’s been a bit of a backlash here and there.

I know that’s a generalization, but people are getting booed at graduation speeches. Whether it’s generational, I don’t even quite know how to think about the various dimensions of this, but there’s no doubt there’s a phenomenon out there.

Thinking about the people you talk to and creatives, who have always had a segment that has been skeptical of AI, how do you hear all that criticism?

What do you think is fair? What isn’t?

And what do people, particularly at AI companies, get wrong about how they talk about AI, particularly to this demographic of creatives, writers, editors, et cetera?

[33:34] Laura Burkhauser:

I think the way people talk about AI in San Francisco is absolutely insane and so alienating and inaccurate.

People are always inaccurate, but it feels self-defeatingly inaccurate in a way that, if you have any kind of connection to reality, if you have relatives who don’t live here that you respect and talk to, you’re just like, “What? I’m sorry?”

Why is the leader of every lab walking around saying, “Someone should stop me. My God, I’m building this really dangerous thing that’s going to ruin democracy and kill everyone. Someone should stop me,” and then is surprised that people are like, “I’m super freaked out. You should stop building this. This is a bad thing.”

It’s like, you’re giving us the talking points, dude.

Stop telling us that you’re building bad things. You’re freaking everyone out.

I just don’t understand it. I kind of understand it. You want to feel like, “Wow, everyone should give me more money because I’m building this thing that’s so amazing that it might kill us all.”

I still think that’s questionable. Please stop doing it.

You’re not watching this podcast. You’re very busy. But maybe someone can show this person a clip.

Stop doing it. Stop saying that.

[34:50] Pete Pachal:

I think their comms teams are listening to this podcast, so hopefully that’ll get through.

[34:55] Laura Burkhauser:

Yeah. Stop doing that.

I got started in consulting, and we used to have to take trainings on change management.

Whenever there was going to be a big change, they’d say, “You’ve got to identify the burning platform and you’ve got to identify the beach.”

The burning platform is the reason why you absolutely have to change. We cannot stay the same. It’s not going to work.

The beach is the party that you get to have when you finally jump off the burning platform. This is the thing you get people excited about.

No one is selling the beach right now.

No one is telling us why we’re excited about AI.

To the extent that they are, they’re using language that makes capital owners excited, like “more productivity.”

But to someone doing the work, when they hear “more productivity,” they hear, “Wow, sounds like I’m doing a lot more work in the future.”

Or you have people talking so far into the future that it doesn’t feel credible and isn’t credible, like, “There won’t be any work in the future.”

That’s scary, too far away, and probably not ever true in anyone listening to this lifetime.

I just think we’ve gotten it totally wrong about getting people excited about this stuff.

[36:18] Pete Pachal:

So then, zeroing in on the creative set, your customers, how do you think about that in your messaging?

How do you talk about AI with customers or even just people in general?

[36:32] Laura Burkhauser:

In the creative world, it’s even worse because you have this Twitter feed of some of the lamest stuff you’ve ever seen with AI-generated video.

Then you have a bunch of people in the tech world being like, “My God, Hollywood’s dead.”

Anyone who actually makes video is looking at this dog surfing and is just like, “I’m super unimpressed by this dog surfing video. Am I insane?”

Why is Hollywood dead because you made a pretty mid video of a dog surfing?

This is bad.

[37:05] Pete Pachal:

Yeah. It’s a standalone thing that, on its own, might be interesting and curious.

Show me the movie where that’s somehow integrated and it’s flawless, and maybe we’ll talk.

[37:17] Laura Burkhauser:

Yeah, and maybe we will.

I frankly think that we will start to see generative AI technology in video and in movies in Hollywood. We probably already are.

I think we’ll see this technology come in, and it’s going to be boring.

It’s going to be the way that Pixar changed the way we all do animation, with a pretty boring, to most people, behind-the-scenes change in how animation happens.

No one thinks about that. They’re just like, “Wow, Monsters, Inc. was such a good movie. I really liked that story. It made me cry.”

That’s what gets people excited.

It’s not the technology that people are using. It’s actually making really cool, admirable, exciting art.

When I look at the box office right now and see a movie doing super well that was made for under a million dollars and has already made something like $250 million at the box office, this is exciting to me.

This is a beach.

What if the beach for AI and creativity is that you are empowering people to make really exciting, original storytelling for under a million dollars and get distribution?

More stories get told. More original stories get told.

We’re not going to see just a bunch of $500 million superhero movies in the theater. We’re going to get to see original storytelling and be able to take more bets on early-stage filmmakers.

I’m excited about that beach. These are some of the stories you can tell.

At Descript, because we don’t do a lot of this generative, “there’s no real human in this ad” stuff, we do recorded media.

[39:01] Laura Burkhauser:

A lot of how I talk to people is that I say: Struggle with your art, not with your tools.

We’re not trying to replace the art here. We’re trying to help make it easier for you to make it.

You’re still going to need to struggle with that storytelling, but not with the tools that it takes to build that story.

[39:20] Pete Pachal:

I usually like to end these conversations on things you’re hopeful about and maybe concerned about as we go forward, but I feel like you’ve answered a lot of that in what you just said.

I’d be curious to jump off your last point about making things more accessible because I feel like that zeroes in on a trend in toolmaking.

You see this in a lot of different creative apps. Basically, it feels like everything is slowly moving away from discrete buttons, menus, and hard-to-find features.

If not overtly a chatbot, it’s becoming more conversational. You tell the tool what you want, and it figures out what it needs to do.

How far does this idea go?

Putting aside Descript for a moment, what does interacting with software look like in three to five years?

[40:14] Laura Burkhauser:

That’s a great question and one we talk and debate about all the time at Descript.

I don’t think buttons are going away, but I do think chat is still on the ascent.

I’m a big believer in voice. The way that I talk to Underlord is that I dictate to it.

As you can tell, I’m a bit of a rambler. It’s hard for me to get it into five words. That’s the difficulty with writing for me.

[40:44] Laura Burkhauser:

It’s sitting down and writing what I think succinctly.

So I turn on the microphone and just kind of ramble to Underlord and let it know what I want it to do with the video, and it does it.

[40:54] Pete Pachal:

I feel like the AI question of the moment is, “What do you use to dictate?” as opposed to, “What’s your phone, iPhone or Android?” from the previous era.

That’s the pro move.

[41:02] Laura Burkhauser:

Totally. I use Wispr Flow. I love it.

But I don’t think buttons are going away.

Buttons and dials are important for people who need control and want to get in there and set things the way that they want them.

I’m excited about UI on the fly, where the computer is smart about generating the right buttons and dials at a moment’s notice.

[41:32] Laura Burkhauser:

If it’s easier to turn a dial and I’m talking to you and I’m like, “Hey, I need more confetti in this shot,” maybe it generates more confetti in the shot and also gives me a confetti knob.

It’s like, “Okay, I put in more confetti, but if it’s not right, just turn the knob or dial to the right amount of confetti.”

You would never need a permanent confetti knob, but because it’s so trivial to generate software in the future, it can actually just generate that knob for you at a moment’s notice to give you the control.

Buttons and knobs are really useful for high-level control, which you’re always going to need for precision editing.

[42:05] Pete Pachal:

Nice. Yeah, I think some people might need a confetti knob. A lot of people edit wedding videos, so that could be a thing.

Laura, this has been fantastic. I want to say thanks so much for dropping by The Media Copilot and sharing your thoughts.

There’s a lot to absorb, but this has been great. I really appreciate you coming on.

[42:40] Laura Burkhauser:

Pete, thanks so much for having me. A really great conversation, and I look forward to chatting again soon.

[42:46] Pete Pachal:

For sure.

The post AI won’t replace Creators. It will replace the work they hate. appeared first on The Media Copilot.