Confronting the CEO of the AI company that impersonated me

Today, I’m talking with Shishir Mehrotra, who is CEO of Superhuman — that’s the company formerly known as Grammarly, which is still its flagship product.

Shishir also used to be the chief product officer at YouTube, and he’s on the board of directors at Spotify. He’s a fascinating guy, and we actually scheduled this interview a month or so ago, thinking we’d talk about AI and what it’s doing to software, platforms, and creativity pretty broadly.

Then things really took a turn. Back in August of last year, Grammarly shipped a feature called Expert Review, which allowed you to get writing suggestions from AI-cloned “experts,” and reporters at The Verge and other outlets discovered that those experts included us. It included me.

No one had ever asked permission to use our names this way, and a lot of reporters were outraged by this — the talented investigative journalist Julia Angwin was so upset she filed a class action lawsuit about it. Superhuman responded to this by first offering up an email-based opt out and then killing the feature entirely. Shishir apologized, and you’ll hear him apologize again.

Throughout all of this, I kept wondering if Shishir was still going to show up and record Decoder, because my questions about decision-making and AI and platforms suddenly seemed a lot harder than before. To his credit, he did, and he stuck it out. This conversation got tense at times, and it’s clear we disagree about how extractive AI feels for people. But I won’t stretch this out any longer.

Okay: Shishir Mehrotra, CEO of Superhuman. Here we go.

This interview has been lightly edited for length and clarity.

Shishir Mehrotra, you’re the CEO of Superhuman. Welcome to Decoder.

I’m happy you’re here. I’m a little surprised you’re here. I think you know what some of the questions are going to be, but I’m really happy you made it. I have a lot of questions about AI, how people feel about AI, and then a feature you launched in Grammarly, which is one of your products, that made people feel a lot of feelings about AI. So we’re going to get into it.

Let’s start at the start. Superhuman owns Grammarly and Coda. You own a bunch of companies. Just quickly describe the structure of Superhuman and all your products.

Superhuman is the AI native productivity suite. We bring AI to wherever people work. Late last year, we changed the name of our corporate entity from Grammarly to Superhuman. We did that because the scope of what we do has broadened quite a bit. And so in addition to Grammarly, which is everyone’s favorite writing assistant, we now have a document space called Coda, and a very popular email client called Mail.

We launched a new product called Superhuman Go. Go is the platform that brings you a network of proactive and personal AI assistance directly to wherever you work. So for people familiar with Grammarly, you can think about Go as taking that core idea and allowing anybody to write agents that work just like Grammarly does. Your sales agent, your support agent, so on, can all help work with you right where you work.

The core idea is that most AI tools require a big change in behavior. We bring AI where you work. Across our products, we see about a million different apps and agents every day. We seamlessly blend AI right into your experience, so you don’t have to think about AI.

That’s what we’ve been doing with Grammarly for years. And now we are opening that up so anyone can build on that with Superhuman Go.

You and I hung out a few weeks ago, and one of the things we talked about was the fact that Grammarly, for most people, is expressed as a keyboard. It shows up on your phone and your documents. You spend a lot of time figuring out how to make sure you work with things like Google Docs.

All of those products are integrating AI in exactly the same way as you’re describing. I think you put AI right next to the insertion point, right next to your cursor. What’s the big differentiation for you?

First off, I think very few of them actually are doing that particularly well. A handful do. But as I mentioned, we see a million unique apps a day. The way to think about Grammarly is it’s your assistant that lives everywhere. You might be in a web app. It could be Gmail, it could be Google Docs, it could be Coda, it could be Notion.

You could be in a desktop app. That could be Apple Notes, that could be Slack, that could be whatever app you’re using. It could be every mobile application. We have, for every one of those applications, figured out the right way to observe what you’re doing, annotate it in a way that is unobtrusive to you and to the application, and to make changes on your behalf. And doing that everywhere is the proposition.

As you jump from tool to tool, there are different types of AI in each one. Most of them actually don’t have that. Like I said, we see a million unique surfaces a day. And the ones that do don’t feel like one integrated experience. That’s why we have about 40 million daily active users and that’s what they use us for.

It feels like the promise there is by looking at all the places you work, your tool will be more intelligent than disparate tools you might encounter in all those places.

Yeah, becoming more intelligent is certainly part of it. For many people, it’s just that one familiar experience that really feels like a virtual human working right next to you.

So is it consistency of experience or is it better and more useful results?

It’s both. The fact that Grammarly is ever present is very important and [it produces] very high-quality grammar results. As we split the product into parts, we said, “We’re going to take the platform layer of Grammarly and we’re going to turn it into a platform.” That’s what we call Go. That’s about allowing other people to create agents and experiences that provide a high-quality experience that we can make ubiquitous for them.

All right. I wanted to understand what you think that the sell of the tools is. I think that’s very important for my next set of questions.

The other thing that I really want to ask is a question I ask everybody, but I think the stakes are a little bit higher here. It’s about decisions. How do you make decisions? What’s your framework?

We have a lot of different thoughts on how to make good decisions. I wrote a piece a long time ago called Eigenquestions, which is about framing not only the right solution, but how do you frame the right question? In terms of rituals we use, the most canonical one is something we do called Dory and Pulse, which is a way to solicit feedback and opinions so that you get rid of groupthink in the decision making process.

But those are probably the two that get mentioned the most if you were to ask teams here at Grammarly or previously at Coda or before that when I worked at YouTube or Google, or so on.

You can see where this is going. Let’s put this into practice. You launched a feature in Grammarly called Expert Review that generated suggestions on how to improve text. It synthesized advice from experts. It used my name among many other names: journalists Casey Newton and Julie Angwin, you can go down the line; bell hooks was in there, which is hilarious in its own way.

You do not have our permission to use our names to do this. You had little check marks next to the name that indicated it was somehow official. People did not like this, I did not like this, and you removed the feature. Tell me about the decision to launch this feature with names you didn’t have permission for and the decision to unlaunch the feature.

I expected we’d talk a bit about this, so I have lots of different thoughts on it.

First off, I’d say I understand and respect how challenging a world it is for experts and idea generators these days. I’ve made a long career out of being a partner to folks like you, to folks like the ones you’ve mentioned. It deeply pained me to feel that we under-delivered for them. And I’d really like to apologize for that. That was not our intention.

On the specific feature you’re talking about, I’m sure we’ll talk more about it, but just to give the high-level view, my view of it is that the feature was not a good feature. It wasn’t good for experts, it wasn’t good for users. It was a fairly buried feature. It had very little usage. You mentioned it last week and talked about it. It took months for anybody to even sort of find it. All that doesn’t really matter. We can do much, much better. I believe we can and we will do better.

We decided to kill it pretty quickly. Notably, we decided to kill it while there was some feedback well before there was a lawsuit and so on. It was just not a good feature. It was misaligned to our strategy. It wasn’t the way we wanted to go after it. We have a much better view on how we think experts should participate in our platform, and I’m a lot more excited about that.

How many people work at Superhuman?

So out of 1,500 people, how many people decided to launch this feature?

It was a small team. It was probably a product manager and a couple engineers.

Inside your decision-making process where you described a way of making sure you solicited the right feedback and then have groupthink, it never came up that using people’s names without permission would make them mad?

Maybe I should step back and talk about what inspired this team and what they were trying to do and what fell short. Let’s start with what they were trying to do. They were heavily influenced both by what we view users to want and what we want experts to want.

Let’s start with users. A lot of people talk about Grammarly as the last mile of AI. They say, “It feels like having your grammar teacher right next to you everywhere you work.” And so many of our users will say things like, “What would it feel like if instead of your grammar teacher, it was all the rest of the people in my life that could be with me as well? I want my head of sales to sit next to me and tell me I’m about to recommend the wrong product. I want my support person to sit next to me and say, ‘I’m about to email this person and you should know they had a big support issue last week and you should acknowledge that before you talk to them.’”

That’s the core ethos of what we’re building. It is taking Grammarly and expanding it so that many of these other experiences come along with you. For some of those people, the people they want feedback from are the people they admire. It’s the experts in the world, it’s the people that they’re trying to look up to and trying to model. They try to do that today with LLMs. They go to ChatGPT and Claude and say, “What would Nilay think about my writing?” That was the inspiration for what the user was trying to do.

On the other side was what the experts were trying to do. As we formed our strategy here, turning Grammarly into a platform, the first people I called when thinking about this were a set of experts. I talked to some prominent YouTubers, I talked to a really prominent book author, and they all told me the same thing. It’s a really hard world for experts out there right now. It’s really hard to drive connection. If you’re a book author, your path to getting to your fans is you just keep publishing more and more books. And they all heard what we were doing and said, “Boy, it’d be really amazing to develop an ongoing connection with my fans. What happens when they put my book down? Can I still be with them and help them along the way?” It feels like the world shifted against them, AI Overviews stealing a bunch of their traffic and so on. This seems like a much better way to go after it.

That was the inspiration behind it. The team and the feature didn’t deliver. It didn’t deliver on either side of it, really. We ended up with an experience that was pretty suboptimal for the user and obviously suboptimal to the expert. The fundamental reason is something you said last week, that it’s really hard to distill what you would do as an editor based on the outcome of your published work. It’s really hard for AI to do that. We need your engagement for that to be a good feature.

So I think they launched something that wasn’t particularly good. Doing that and learning from it is part of the process, but that’s what they thought they were doing.

Sure. How much do you think you should pay me to use my name?

It’s really important to think about attribution and think about impersonation, and so on. As an expert, you have a trade you make on the internet. The idea is that when you put content out there, myself included, you hope people use it. You want to refer to other people’s content. You want people to link to you. You really, really hope they attribute you when they do. When somebody uses your content, should they attribute you? Of course. And to attribute you, you have to use your name.

There’s a different line which is, should people be able to impersonate you? And I think that is a very different standard. And we saw the lawsuit. Respectfully, we believe the claims are without merit. The idea that the feature is impersonation is quite a big stretch. Every mention was very clearly, “This is inspired not only by this person, but also inspired by a specific work from this specific person, with a clear attributed link to get back to them.” It’s far from that test [of impersonation].

If your work is used, should you be attributed? Yes, I think you should. That would be the nice contract. It doesn’t always happen. There are many products that will use your work and not attribute. We thought it was very important to attribute. I think that would be the view.

Let me flip around the other way–

Wait, let me ask you that question again. If you use my likeness, how much should you have to pay me?

We should not be able to impersonate you, period. We did not. If we use your work, if any LLM product or any product at all uses your work, they should attribute it to you and they should link back to you. That’s a human contract we have for how the internet is supposed to work. It’s a really important one. It should be the standard you’re looking for from LLMs too.

It’s a very different question you’re asking here, which I think is a more important one. I’m not really here to defend this feature. I don’t think it’s a good feature. I’m not trying to be close to this line. I think our main goal is to build a platform a lot like YouTube. You should choose to be on our platform. You should be able to choose and build an experience you trust. You should choose your business model. When you choose your business model, you should get paid for your contributions to it. That’s the model we’re working on. That’s really where I want to be.

I hear that you’re saying you’re not here to defend the feature. I just want to put you in the chronology for one second. The feature was launched. It is true. It took a while before we even discovered it, and wrote the story about it. It then blew up. Many other people wrote stories about it.

Your first response to the negative publicity was to offer people an email opt-out where if I didn’t want my name to be used, I could email Superhuman and say, “Please take me out.” Only after the lawsuit did you discontinue the feature.

That’s not true, Nilay. We heard the first complaints from a handful of experts. They said, “I’d like to opt out of the feature,” and we addressed what they asked for. We then sat down and looked hard at the feature, and to be honest, I hadn’t spent any time on it. I came and looked at it and I said, “This is off-strategy for us.”

We announced we were taking it down well before there was a lawsuit. The reason we took it down is it’s all strategy, it’s not what we want to do. That’s not how we want to work with creators. We think we’re building a platform you should want to be on. We think we’re hopefully part of the solution for how you can take your work and make sure it’s present for people everywhere. It wasn’t our goal to be anywhere close to that line. But the feature wasn’t good, so we took it down.

You say it’s off-strategy for you. The feature obviously shipped. What made it on-strategy at the time it shipped?

At the time, the team believed they were doing that. They were looking at users and they were focused on a user need, which is, “I wish an expert could give me feedback at this moment. I wish my salesperson could give me feedback. I wish my support person could give me feedback. I wish my idol could give me feedback. I wish this expert could give me feedback.” In itself, I think that motivation that users have is a really good one, and I think one that I would encourage experts and creators to lean into. It’s a big opportunity.

Why would they lean into it if the value for that is $0?

No, it should be our job to make sure the value is not $0. We want you to–

How much do you think you should pay me?

To be clear, when you do the work to bring an agent, craft it, put it on our platform, then you should get paid for it. Just like how platforms like YouTube work.

Walk me through the economics. If you launch a platform that lets me say, “Okay, Nilay Patel can give you advice inside of Grammarly,” what are the economics of that platform? How much will I get paid to do that?

We’re building this business model now. Our store currently has a payment model for this that has a 70 / 30 revenue split that’s very similar to how a lot of other products do. If you want to go build an agent like that, you can do that today. There are a number of experts that already have. And that’s the core part of our strategy.

If you already had that system, why build another system that used my name for free?

We didn’t have the system at the time. And they are very different features. The team that built Expert Review, they were trying to address this need, they just missed.

How many times did you use my name?

Because it’s a legal case, I really can’t get into details of those types of things, but it was a very small number for basically everybody. The feature had very little usage.

Was there a set group of names? Was it just picking names out of the ether? Was it randomly hallucinating names?

It came right from the popular LLMs. So it’s exactly the same experience you would have if you came to Claude or Gemini or ChatGPT and said, “Can you take this piece of writing, recommend the people who would be most useful to give feedback on it, take their most interesting works and use that to try to give me feedback.”

By the way, that’s a really hard feature to make good for users and it’s going to take work with people like you to actually deliver on that need.

Did you track how many times you were using people’s names?

We’ve certainly logged all the different interactions, yes.

So you do have a record of how many times my name showed up or Casey Newton’s name showed up, or anything like that?

It’s not tagged that way, but we’ll have to produce it obviously for a lawsuit.

Journalist Julia Angwin has filed a class-action lawsuit. There are a lot of ways that could go. You’ve said that claims are without merit. What did your lawyers say to convince you that the claims were without merit?

What did the lawyers say? It’s actually quite clear. It’s a layman’s test, it’s pretty obvious. It’s just not impersonation. When you look at the feature, there’s a disclosure next to every single link at the top and the bottom of the panel, very clearly stating these are inspired by these people. It clearly states we have no relationship with these people, that that’s the future. By the way, I’m not trying to defend it as a good feature. I don’t want to be on this line.

Maybe I could step back for a second and say, this is not the first time I’ve seen a situation like this. I used to run the team at Google — I used to run the YouTube team. When I got to YouTube, we had a big lawsuit from Viacom at the time, a very heavily watched lawsuit that we won. We won on summary judgment actually. We completely crossed the legal bar. But that’s not the standard we held ourselves to.

We looked at that and we said that the law doesn’t require us to do this, but we chose to do a lot more. We launched Content ID as a way to make sure that creators could find content that other people uploaded on their behalf. We launched an open creative program, which, as far as I know, is still the only platform with an open revenue share that’s out there.

I don’t think the legal standard is the right standard to be looking at. I’m not trying to get close to it. It’s fairly clear to me that we didn’t cross below it, but that doesn’t matter. We’re not trying to be close to that standard. We need creators to work. We need their business models to work for our platform to work, and it’s very similar to what happened at YouTube.

I have a lot of thoughts about YouTube. I’m going to ask you about YouTube. I have a lot of thoughts about the Viacom case. A lot of what happened with Google and YouTube is the foundation for the internet and policy on the internet as we know it today. That is changing because of AI. So I do want to ask you about that stuff because I think your history will shed a lot of light on how people feel about AI in particular today.

I just want to stay on this one more turn. You’re saying “impersonation,” but that’s not the claim in the lawsuit. The claim in the lawsuit is the law in New York and California that bars companies from using names and identities of people for commercial purposes without their consent. And so, here you did have a commercial purpose here. You were selling the software and names were appearing as inspired by our names.

I’m not in this lawsuit. I haven’t signed up for the class. The class hasn’t been certified. I promise I haven’t sued you yet. But the bar is very different from straightforward impersonation. It is the use of likeness for commercial purposes. And you’re saying it is without merit, and I haven’t seen you address that specifically anywhere.

I’ll have to leave the legal arguments for the lawsuit and for the court case. I think our view of it is that the set of work that was there was a fairly standard attribution that was well above the bar that any other product would do, what every LLM on the planet is doing and so on. And it didn’t come close to using name and likeness in any way that was beyond attributing the source.

You’ve already said this feature is bad, so I won’t hammer you on this too much, but I’m reading the edit that was generated with my name on it, which is just bad. I would literally never give this edit. It says I should “raise the stakes of a headline by adding emotional or stakes-based words that could underscore why this launch matters right now.” I’ve been an editor for over 15 years. I’ve literally never said anything like that.

You pinned the reason why. The idea that you can uncover your editing style from the end work, I just think it’s not possible. It’s very hard to come back from that end work and say, “What was the editing pass before that?” To do that well, you have to do it. You have to sit down and say, “Here’s how I would edit these things.” And I think you can provide that service and you can get paid for it. And hopefully we’re one of the platforms where you choose to do that.

So, you don’t have an annotated list of whose names are used in the feature, but you have logs of everybody who uses the feature, presuming those logs have the names in it, and you presume you’ll be able to provide that if you get to discovery.

I’m sure we’ll be asked. Yeah.

Do you think you’ll be able to provide that list?

I’m sure we’ll be asked. We’ll see.

Because it strikes me that one way you could get around this lawsuit is by just saying, “Actually, we never used Julia’s name until she went asking for it.” In the same way that OpenAI, when it responds to the New York Times lawsuit says, “This never happened until you prompted us specifically to do the things you said are illegal.” And here you have the same out. You could say, “Actually, until you asked us, we never generated your name.” Has that come up?

There are a lot of things in our defense that I won’t cover, but I think the core of this argument isn’t going to be that. The core of the argument is that what we did is normal attribution of content on the internet.

The reason I’m asking this very specifically is, “Hey, we never actually used your name,” puts you in a different spot than, “Hey, we have different feelings about the value of attribution.” The reason I’m asking this question as harshly as I’m asking it is that I don’t think the defense is whether or not people use the product or whether or not the names ever showed up. I think those are just clear cut, binary on or off. “Your name never showed up, you can’t sue us.” You’re saying the defense is, “Hey, that’s not how attribution should work.”

You used to be the chief product officer at YouTube, and YouTube is defined by creator attribution scandals. Every year, there’s another scandal about react videos. Every year there’s another scandal about the usage of copyright, about whether or not you can make an AI creator out of Marques Brownlee and just run a million videos of him and steal his views. It’s the essence of the YouTube creator ecosystem.

Do you know how YouTube reacted to this feature when we wrote the story? They invited me to an early preview of their AI likeness detection system, because they knew that would be good press for them. If you were still running YouTube, would you have ever allowed a feature like this to go out?

It’s interesting the way you just described it. First off, some of the ones you described, describing react videos as scandals is a very interesting way to describe it. Because I think–

Oh, they’re absolutely scandals.

I understood your definition. They’re also incredibly popular and have led to a whole genre of content being created. Likeness detection, Content ID, they were all fantastic tools for creators. My team built the Content ID tool with the same idea.

If somebody does that to Marques Brownlee and they copy his videos and put them up, then you can use that tool and he can not only go claim them, but he can also go make money on them. That is a tool we built for YouTube, and I think it’s been incredibly popular. We took what looked like a scandal and went well beyond it. To be super clear, it’s not what the law requires.

No, I understand what some of the law requires, but the use of Content ID and the issuing of copyright strikes, which is something I’ve experienced, if you issue a copyright strike as a creator against another creator, that is a nuclear move, that comes with severe social and community consequences.

To be clear, if you use Content ID and you use it for monetization, you’re not issuing strikes.

Right. But I’m saying the YouTube economy writ large is defined and in many ways the products are built around issues of attribution and payment and monetization — where the views flow and where the money flows.

Content ID is a brilliant innovation because it allows people to get some views and the right people to get paid. YouTube doesn’t exist without music. If the music is ever on YouTube, the publishers get paid because Content ID can identify the music and get them paid. I understand that. But that is a system that tracks attribution and delivers monetization.

I’m just saying, I don’t see how YouTube could have ever said, “We’re going to let Marques Brownlee edit your video without paying Marques Brownlee.” It wouldn’t exist in that ecosystem.

No, you just said it. What YouTube did is say, “When it happens, we are going to help you find it,” but you’re not preventing someone from doing it. It’s a very different standard.

But you’re making sure that the people get paid.

You’re making sure after. To be clear, the idea of copyright is very different from a name and likeness claim. If I built a video that said, “Hey, I really like Marques Brownlee, and here’s what I think he would say,” or “let me tell some jokes about Nilay,” it’s a very different standard. The standard for YouTube was about copyright, and that’s a set of regulations that are governed by totally different parts of the law.

In that case, you have a claim, there’s a DMCA statute that allows you to go and enforce your copyright. That’s not actually what we’re talking about here. But the principle of what is similar is that in both cases there’s a law, and the law does not really meet the creative bar. I think the goal of the community, the goal of products like ours, working with people like you, is not to use the law as the test. The goal is to get well beyond that to align our interests, such that your success is our success, and that should be our goal.

Are we required to do it? No. I don’t think that’s a requirement. We choose to do it because it’s the best way to build the right products for our customers.

I used to be a copyright lawyer. I’ll happily admit that I was not the world’s best copyright lawyer. I understand that people don’t understand the difference between copyrights and trademarks and names and likeness. I’m saying that AI is collapsing those differences faster than ever before. There are European countries that are just openly suggesting you should expand copyright law to include likeness.

I should be able to copyright my face, and then that means I can slide in under the existing legal regime instead of hoping that the United States Congress in 2026 can reach a resolution on expanded likeness protections. This is a thing that is being suggested because copyright law is more or less the dominant regulatory framework that exists on the internet.

I look at the big social platforms like YouTube, Instagram, and TikTok, and they have built all these systems to respond to copyright law — specifically copyright, things that can be protected by copyright law, that can be monetized in different ways by copyright law. Our likenesses are not one of them. Our names and faces are not one of them.

This seems like the place where the things you’re allowed to do and the things you should do are going to be ever more divergent. You are the one who’s experienced it the most loudly of late. And I’m curious if you’ve learned anything other than, “There’s what the law says I should do and there’s what I should do and we’re going to find the line down the middle.“

We’ll see if the laws find a ground on that. I do think it’s a catch-22 as a creator. The copyright law has been around for hundreds of years now in its various forms. It started like the way music composition was licensed, it started with Mozart and Bach. It has grown since then. Almost every country in the world has reached a very similar standard.

There’s a very thin line between taking publicly available work and being able to refer to it, and copying it. The idea that defining all references to work as being uses of names and likenesses, it would break the internet, it would break your business. You wouldn’t be able to refer to me. How’d you get on a show last week and talk about me?

Just to be clear — I don’t want to be all inside baseball about making a podcast, but we made you sign an appearance release to come on the show.

To come on the show. But you talked about me before I came on the show. Of course you should be–

We talked about you before you came on the show, but in order to be a real media company and not fly-by-night and then to use clips of your face talking, our lawyers need a release. And if you don’t sign it, they won’t let me use the show, because they need to be protected against you showing up tomorrow and saying, “I didn’t give you permission to use my face.”

No, I understand that. My point is broader than that. You talk about lots of people and that’s part of discourse. That’s part of how we work. Your articles will link to people, you attribute them. I think that’s really important. And if you drew a line that attributing something is like using their name and likeness, then it’s a very hard line to draw.

Again, this wasn’t an attribution. You just made something up and put my name on it. There’s no attribution here. This isn’t anything I ever said. It’s not something I would ever say. I’m not even sure how you would get to the idea that based on my work that I would ever say anything like this. There isn’t an attribution here. There’s no work that exists that would lead you to this outcome with my name attached to it.

I’ll repeat: The feature was, “Here’s a suggestion generated by a specific work from a specific person.” Everything is clearly indicated that it’s a suggestion generated from–

Wait, I’m sorry. You think in my role as editor-in-chief of The Verge and co-host of The Vergecast, I emphasize the importance of crafting compelling headlines that convey urgency?

I already told you it’s a bad feature. That’s not what you’re questioning.

You’re telling me there’s attribution and I’m just wondering what the attribution is.

Just read the rest of it. It says, “Based off of this work from you, we asked–”

No. It just says, “This suggestion is inspired by Nilay Patel’s The Vergecast.” I promise you on The Vergecast, I’ve hosted that show for a long time. I have never said, “What emotional or stakes-based words could underscore why this launch matters right now?” The Vergecast is not a show about editing headlines about smartwatches, first of all.

So I don’t know how you got from A to B and then I don’t know why you think that’s an attribution.

If you were to go and read someone’s work, put it online—you do this on your show all the time—and say, “I read this person’s work and here’s now my conclusion from it,” you should decide whether that is a suggestion generated from attribution or not. I told you I think it’s a bad quality suggestion. I’m not trying to defend it. I don’t think that’s what we want to talk about there. But the question, when you publish work, can humans and AI use it to generate other suggestions, other impressions? They can, and you would like for them to attribute it.

But it’s not work that that person made. Hallucinating a thing that you thought I would make and then saying you’re attributing it to me, doesn’t provide me any benefit. It might actually detract from the benefits I could provide to other people. That’s the disconnect that’s in my brain. I’m not sure why this is an attribution.

If I’m like, “I talked to Shishir and I think here’s what he would say,” that’s very different than saying, “I read all of his work and I’ve asked whatever quick version of Claude or ChatGPT to just make something up and I’m going to put his name on it.” There’s something meaningfully different there. And it doesn’t seem like you’re willing to concede that.

No. I’m not. It’s fairly clear that generating a suggestion based on somebody else’s work… just use the simple task of a human doing it. If you generated a suggestion based on someone else’s work on your show and you said, “I read this person’s work and here’s my impression from that, this is what I think they meant,” you could build a whole show based on that. So you don’t always get it right. You don’t always say things about the people that you’re commenting on that are correct.

Right. But I’m not attributing that idea to them. That idea is clearly mine.

The feature is very clearly stated that this is a suggestion developed by this feature based off of this work.

Let me ask you a different question. I’m curious about this across the whole sweep, from YouTube to now. There’s an NBC News poll that just came out about how people feel about AI. And the answer is bad. People feel badly about AI. AI is polling behind ICE and only slightly above the Democratic Party. This is a tough spot to be in. It’s a -20 perception.

I think the reason for that is because it’s so extractive and the value isn’t there. I would compare this to YouTube, which a lot of people thought was pretty extractive. You fought a pitched copyright battle about YouTube, about whether South Park could be on YouTube without permission, and Viacom was going to sue you. That case was fascinating because the public was decidedly on YouTube’s side.

Oh, that’s an interesting memory of it.

I covered that case. I was in law school studying copyright during the case. The vast majority of people were like, “YouTube is really useful. We love it. And these big Hollywood companies suck.” When Napster was under fire, the public was not on the side of the record labels. They were not on the side of large companies. They were on the side of file sharing. Because the utility was so high regardless of the economic or social cost. I could keep going on and on with this. You can tell people all day long about the labor costs of Uber and they’re still going to use Uber.

There’s a trial right now about whether social media platforms are damaging to teens’ health, whether they’re defectively designed products that hurt kids. That trial is ongoing as we speak. The jury is impaneled right now, and people are still going to use those platforms because they don’t care.

The environmental costs of big, stupid cars — you can tell people all day that trucks will ruin the environment, Americans will still buy trucks. That’s what we’re going to do. AI is only perceived as extractive. It’s less beloved than ICE. That’s crazy to me. Do you understand that the extractive nature of AI is causing a problem for the whole industry? Because you’re sitting in the middle of one of these controversies right now.

I think you’re drawing a pretty broad link for why people are afraid of AI.

I think great consumer products that provide a lot of value overcome their social costs.

Number one, AI has a lot of challenges ahead of it. There’s lots of opportunity. It does meet your other tests. It has created some of the most popular products in history. And there are many people who would have you pry any of those products from their cold, dead hands.

I think that the challenge with AI right now is that it’s challenging people’s sense of the future of their humanity, their ability to work. Those are really the challenges there. The line we’re talking about here, I don’t think that’s actually what you’re reading into that poll.

What would you read into the poll where AI polls below ICE?

People are scared for their jobs.

You think people are just scared for their jobs?

Do you understand that that’s extraction? You’ve taken the sum total of everyone’s work on the internet and now you’re going to use it to replace human beings and their jobs without any economic recompense.

That is certainly one way it could replace people’s jobs. I don’t think that’s the way that most people are worried about how it could replace their jobs. I think they’re wrong about it. I don’t actually think it’s going to replace as many jobs it’s going to create. One of the reasons why is that our model for thinking about AI is about bringing it to people and expanding their work. We like to call it the product that helps you become a superhuman. So I think they’re wrong about it.

But if you’re asking me why it polls so low, it’s because the copywriter feels like, “Maybe I’m not going to need it anymore.” It’s the salesperson who says, or a support person who says, “I wonder if an agent’s going to be able to do my job.” I think the idea that it has something to do with name and likeness is a pretty big stretch.

You’re sitting in the middle of a controversy where a lot of people are mad at you for appropriating their work. If you’re a copywriter at an ad agency — I know a lot of copywriters at agencies — they’re saying, “You took all of my work.” Not you. “The AI companies have ingested all of my work for training and now they’re going to replace me and no one got paid.” Hollywood is basically like, “No one’s paying us for this.” The people who write on Tumblr are saying, “Now OpenAI is going to make a porny fanfic for people. That was our job. Why didn’t you pay us?”

You’re absolutely right. Creators are facing a very hard road right now. I don’t think it’s caused just by this feature or just by the latest advanced AI. They’re facing a hard future for a lot of different reasons. But the poll you’re referring to is of the broad population, and the broad population is not creators. The broad population has jobs that they are afraid may not be available to them. Whether they’re a truck driver, whether they’re a support person, that’s what they’re afraid of.

I’m not diminishing the fact that creators also have an issue with AI. I’m just pointing out that the broad impression of AI, the challenge we have with it, is that the entire industry has done a really bad job of helping people understand why a technology like this can help them and not prevent their job from being taken away. And most people just aren’t creators.

I’m not objecting to what you’re saying about creators. I’m just saying most people aren’t stressed about that because that’s not their job. That’s not what they’re individually afraid of.

No, I understand what you’re saying. I’m just pointing out that almost every major technological shift has been extractive in some way. Google copied all the books in the world without permission, and then we had a Google Books case, and Google had to win that case. And they did. They were able to do it.

Google had to win the Viacom case with YouTube. Google had to win the Google Images case against Perfect 10, which was maybe the least sympathetic plaintiff of all time, because it was a porn company, and Google was doing Google Image thumbnails of softcore porn. It was obvious that Google was going to win that case, but they still had to win that case.

All of this stuff got litigated at pretty intense levels in ways that are precedent still to this day, and it doesn’t feel like we’re spending the time to litigate, “Hey, you can just make a deepfake of my face and use it to sell headphones on Alibaba.” You can just start a company and say, “Well, it’s attribution, so I’m just going to use the names of famous people on my product to say these are the edits.”

There’s a link there that seems very direct to me, maybe just as a creator, but also I would submit to everyone else who says there’s a pretty extractive cost here and the consumer benefits are not nearly as clear.

In some ways I like the YouTube analogy. It’s a good analogy. When I talk to our team about why the legal standard shouldn’t be the minimal standard we try to hit. I will also tell you that what we’re doing here at Superhuman, I don’t expect to be very close to this line. There are other products that are very close to this line. Our core strategy is about building a platform that you can choose to participate in or not. I don’t think it’s going to be a fine line for us. I know in this case, we built a bad feature. It was not received well by either users or experts. I don’t like that. I killed it for that reason, but I don’t expect to be sitting here…

The YouTube analogy: you’re right. The Viacom case had to get litigated for YouTube to exist. And if it had gotten litigated the other way, YouTube wouldn’t exist. Actually, most of the internet wouldn’t exist. And so the idea that it got litigated that way, it was a win for everybody. It was a win for society. I do think it was a win for YouTube. I don’t expect that to be our case here. This is not a line I’m going to be close to.

There are a bunch of copyright cases against the AI companies. I feel like I should disclose that our company, Vox Media, has sued Google over ad tech. It has nothing to do with AI or copyright. I feel like I need to disclose it because I disclose it every time. Vox Media sued Cohere, one of the AI labs, over copyright infringement. The New York Times has sued OpenAI.

There are a million of these copyright cases floating around. There are more every day. One of them could go the other way, and this industry could faceplant. What do you think happens if one of the big AI labs loses a copyright case?

Are you asking me as someone watching the industry or are you asking me in my Superhuman role?

My Superhuman role is straightforward. Whatever the models do is what we’ll use. And so if the models end up needing to restrict that behavior, then that is what it is. We sit on top of the models. I don’t think we’ll be the ones in the middle of those cases. If I look from an industry perspective, I think it’s a really hard case, in both directions. I have real empathy for both sides.

Copyright law is, like you said, what has allowed the internet to work, and not everybody is happy with how the law draws a line. You’re right that YouTube tested that line in a new way with the Viacom case and so on. What OpenAI, Claude, and Gemini are doing will test it in a new way. I hope they find a good line for it. I don’t think that’s where we’re going to be. We’re not going to be the ones in the middle of those lawsuits or those figuring out where that line is.

If the incremental cost of a token skyrockets, because suddenly the AI companies have to pay massive licensing fees to copyright owners downstream, what happens to your business?

I don’t think it really matters to us because it’ll all happen in the models underneath us. It doesn’t matter to us as our own entity. It matters to me as a citizen. I think it’s really important. But I would also remember that for us, the primary agents people are trying to build on Superhuman have nothing to do with this. The expert case is one case.

What people are doing with our product is they’re going and taking their sales methodology and turning it into agents for their salespeople to be able to use. They’re taking their support tools. They’re taking their calendars and making sure that as you’re writing an email and saying, “I can meet tomorrow at 6PM, please make sure that I’m actually free then.” Like I said, this is not a common part of our business.

No, I’m not saying the expert review part. I’m saying you’re describing, “Take all of my sales literature, take my calendar,” that gets loaded in a context for a model that you call, right?

If the incremental cost of a token in that model goes up because the AI companies suddenly have to pay a bunch of copyright licensing fees, what happens to your business?

If I were those companies, the solution I would have isn’t to go distribute that cost across all users. I would charge users a subscription for using that information. That’s the business model they should have.

My personal view of what should happen is I should come to ChatGPT or Gemini or Claude and I should prove that I’m a New York Times subscriber, and then it should give me answers for The New York Times. And The New York Times is going to have to make a choice of, “Do I only want my content to be used for my subscribers or not?” But if I were those companies, that’s what I would promise.

All these cases are different. So I’m going to generalize here and you can attack me for generalizing and that’s fine. But broadly, they split into two lines. There’s one, the thing you’re describing, which is you spit out content that I’ve already made, like Suno can make a Beyonce song that’s copyright infringement on output. Other set of cases where I think much more important–

It’s on input, it’s on training. And saying, “You ingested all my material without permission.” That’s also copyright infringement. If that goes the wrong way for the model companies, their cost structures change in retrospect. You can’t build the systems you’re describing because the model itself–

No, that’s what I was responding to. So output, a copyright law covers it. If you produce something that could be mistaken for the work of another person, then they can file a claim, they can get it taken down; if they choose to leave it up, you can choose to negotiate a revenue share agreement or whatever you might want to do with that. Output is cleared. Input is not cleared, like you said, and the cases haven’t been resolved in a particularly clear way.

The point I was making is if I were them, I wouldn’t take the cost of input and distribute it across all users. I would split the model. If it really went that way, I would say, “Fine, you don’t want your content there. I will build a version of the model that is just for New York Times subscribers and charge them.”

Your particular question was, “Will that cost get passed along to the other users of the LLMs?” That is what’s happening right now. They are paying for that content. It is being passed to us. Does it matter to us? Frankly speaking, the pace of innovation in that category is so high, the profits being generated there are so high, that no, it hasn’t mattered to the upstream users — or to us, to ChatGPT users, Gemini users, and so on. It hasn’t stopped their growth at all. Will it someday? Maybe. I don’t know.

But my point was more that in this world of output, copyright is fairly clear and the law covers it pretty well; input copyright is not that clear. It’s not clear for good reason. If you’re a human and you read a book and then you learn something and then you talk about that thing, what should happen? And that’s a legitimate question that hasn’t been well tested in the courts.

I don’t think the industry is going to take that cost and just pass it along to all users, but we’ll see. If it does, then it does and we’ll have to deal with it. Everybody will.

Most humans cannot infinitely scale to create trillions of dollars of enterprise value by reading one book. That’s the difference. To get that value at that scale, usually lots of people have to buy copies of the book and the economics spread out. The scale is the difference.

I understand that is a very fair argument, that this is not the same as a human reading the book. Obviously that’s the line being taken there. I would postulate that whatever way that case ends up, the correct answer for experts is it’s time for a new business model. And I think the idea is that you’re going to get into exactly the right spot and you’re going to get pennies for every query coming through Gemini. That’s certainly one path.

When I went and talked to people about what we’re doing here at Superhuman, what they told me is, “Actually, I don’t really want to be fishing for pennies whenever my work gets used. I want to build connections with people. I didn’t build content to put it out there and get paid a fraction of every use. I want to go build a product that actually connects with people. I want to do this.” YouTube offers a great way to do that. What we’re doing is Superhuman should offer a great way to do that as well.

Let me ask you about that specifically. I wasn’t at South by Southwest. We have a little baby. I didn’t travel this year, but I watched Instagram. I experienced South by Southwest through the magic of Instagram and TikTok.

You had a suite there at South by Southwest. I looked at some of the videos. The caption on one of the Instagram carousels… I’m just going to read you the caption. This is from the Superhuman suite at South by Southwest. There were a lot of talks there. The summary of the talks was, “AI can’t replace human creativity, empathy, or emotion. It won’t take all of our jobs, but it will reshape how we work. And in the AI era, taste and judgment are more valuable than ever.” Valuable on what metric? Is it dollars?

Valuable on every metric.

Specifically dollars. Dollars are what I pay my mortgage in. Is it dollars?

I’m sorry, I didn’t understand the question.

If my “taste and judgment are more valuable than ever,” but it’s also infinitely replicable and you think I need a new business model or every creator needs a new business model or–

Sorry, you made a big leap from that.

How do I make more dollars? If my “taste and judgment are more valuable than ever,” where do the extra dollars come from?

So just to be clear on the tagline for Superhuman, what we believe is that we can help all our users become superhuman by bringing them tools that allow them to expand their work. The main way we think about people is that Grammarly doesn’t do your work for you. Grammarly helps make you a better writer. And you still publish your essay, you still post your article. It’s our job to turn you into a superhuman. That’s our promise to our users. That’s what the banner’s about. Your question is a very good question.

The banner says “taste and judgment are more valuable than ever.” I’m just asking you to define the value and what value is going up and what value is going down.

If you’re using Grammarly and you’re a student or a salesperson, it is your taste and judgment that is actually what gets valued in the end. We’re here to help make sure you don’t make a mistake. We’re here to help make sure that you present yourself the best possible way. That’s what that banner is about.

We have 40 million users who use our product. The vast majority of them work in professional industries, they’re salespeople, they’re support people, that’s who that’s addressing. And we’re trying to tell them, “Don’t worry about losing your job when you use our products because we’re here to help you scale more. We’re here to help you be a better version of you.” That’s what that banner is about. That’s what our promise is about.

We have a proposition for you, Nilay, as well, which is that you can now become one of those assistants to all those people. Many of them have no idea that they could use your help, but you can build that relationship with them like Grammarly does. People personify Grammarly all the time: “My high school English teacher sitting next to me everywhere I work, that makes me better. It makes my trust and judgment shine through.”

I would like your agent for people for whom you matter. You should be able to build an agent that sits right next to them and you can actually feel like their editor. Now, you have to do some work to make that a good experience. You’re going to have to figure out how to document your editing style in a way that actually produces a good result, not like the one you quoted earlier. But if you can do that, you should be able to build that relationship. You should be able to construct it the way you want, you should control it, and you should be able to make money on it.

Wait, hold on. You understand that you’re saying I have to do that because all of the work I’ve produced in my career to date has been taken without compensation by AI companies.

I didn’t make that statement.

What? You’re saying I need to invent some new business model as an expert and upload an agent of myself to your tool and then advertise it to get a 70 / 30 revenue split from however many people use Grammarly, because my actual body of work has been reduced to zero value. That’s a pretty hard sell.

I’m not here to tell you how to answer every question about what’s changed in the creator economy. One way to look at it is that the path of being a creator has become harder. I assume this podcast is going to end up on YouTube and Spotify and so on. There are paths to becoming a creator that become easier. There were folks that, when YouTube came out, told us all the same things and they said, “We don’t understand. Our business model is screwed over there. Why should we work on YouTube?”

The ones that looked at it that way and saw it as replacement ended up not moving forward to the future. Obviously you did. You run a show on all these platforms and you figured out a way to turn that into a business. You saw that opportunity and you expanded what you could do.

If we look at AI from that perspective and say, “AI is here and it’s reducing the number of people who need to traffic to my current experiences,” that’s one way to look at it. There will be some creators that look at it that way. I would hope we look at it the other way and say, “Some of these platforms are going to give you a way to participate, are going to give you a way to take your expertise and put it in front of people in a way that actually helps them in a different way than you could connect in the past.”

That’s a bright future. I’m not really trying to say you have to or you don’t have to. It’s an expansion opportunity. I’m not really here to defend what some other company is doing with content. What’s happening there is happening there. I’m just saying creators feel that pressure. We recognize it. There’s an opportunity. I had one creator tell me that their traffic in just the last year from Google is down 50 percent. They said that with AI Overviews and so on, traffic is down 50 percent. They sell books.

My reaction to them was, “That really sucks. I understand why that really sucks.” I would also tell them, “If you’re a book author, waiting for people to search your name on Google has got to be the least good way to monetize your expertise. So now let’s talk about how we can take what you do well and get it in front of people in a way that creates value in a different way.”

Maybe we can do it in a way and get it in front of people in a way that creates value in a different way. And maybe we can do it in a way that isn’t tons of incremental work for you and brings you a new type of opportunity. I think platforms like ours are going to give that opportunity to people who choose to take it. Not everybody will.

Can I extend this to you as the CEO of a software company?

This is the same argument I hear about the frontier models, and the AI companies and their relentless expansion into every category. And then what you might call the SaaSpocalypse. Why would I pay your margin on tokens that you’re buying from them when I can just buy their tokens directly and just talk to Claude? Why wouldn’t I just vibe code something that looks like Grammarly and run it instead of paying… what, you’re like $160 a year? This is the thing that’s coming for the software industry writ large. Do you feel that same pressure?

The SaaSpocalypse is not an easy word to say. It’s a little overstated. I’ll give you my view of it. There is a lot of software. The ability to build software is definitely getting much, much easier. I think the reasons why people choose to use software is often because it does a job particularly well and that there’s often a network effect associated with it.

I’ll give you an example and I’ll just focus on customer relationship management (CRM). People look at the SaaSpocalypse, they go and try to judge Salesforce and say, “Why would anybody pay for Salesforce? I could just vibe code my own version of it.” Well, first they say, “Why would anybody have a CRM?” And then if they do need a CRM, why would they pay for Salesforce?

I’ll answer both questions. Why pay for a CRM? When you have groups of humans working together, you need software for them to work together. If I have one salesperson, I can keep all my sales in my head. If I have 10 salespeople, maybe I can do it with a spreadsheet. When I have 100, I need software to keep them together. That software today is called CRM software. When I have 1,000 agents selling on my behalf, I’m going to need a way for them to coordinate with each other. It might be different, but I do think it’s going to be important. Why is it going to be products like Salesforce? I don’t know if it will be Salesforce, but the power of network effects is going to become much higher.

You’re going to say, “These are products for which I’m going to pick the product that is plugged into the ecosystem in different ways.” Why would people rebuild Grammarly? I’m sure they’ll try. My hope is by that point, we are the platform for all the best agents that work right where you work and you [don’t] have to go replicate all of them. I’m sure there will be people that will, but I think most people won’t. That’s an important bet for how the software industry moves on. The need for software is only going to increase. The importance of network effects will only increase.

You don’t think that OpenAI, or Anthropic, or Google will say, “Well, Grammarly is pretty useful. We can build a tool that looks just like it in seconds and ship it and kill their product. They’re just buying our tokens anyway. We can just kill them pretty easily.”

The ability to build that tool has existed for a long time. So if that were true, our business wouldn’t be growing. We wouldn’t have 40 million people using it every day. The idea is getting easier and easier. Yeah, we can’t stand still. If we stand still and don’t continue to innovate, if we don’t build that network effect, if we don’t continue to add value for people, we’ll get caught. That’s always true.

I just want to end on a big thing. Again, you used to run these platforms. You’re on the board at Spotify. I know you think about the economy here and how work gets produced and who gets paid as deeply as anyone. I look at the shape of the media landscape right now, the information landscape that you might call the internet. And I say, “Boy, everything is slowly turning into QVC.” Making this stuff is getting devalued every single day. Being the person who makes the stuff is getting harder and harder. It’s something you’ve repeated several times now over the past hour.

At the end of it all, the creators all have to pivot to selling something. The Paul brothers have to sell you bottled water. Mr. Beast has to sell you energy bars. We’ve devalued the work so much that unlike any other industry in the world, the internet industries, the information ecosystem pivots from bits to atoms. That’s pretty rare in the history of business.

Most businesses pivot from atoms to bits. The margins of bits are historically much better than the margins of atoms except on YouTube, except every major artist has to be on tour forever because the money from selling music itself is so low. AI is bringing that at scale. You can feel the pressure. This whole conversation has been about that pressure.

Maybe the legal doctrines don’t line up exactly and maybe I’m making too many generalizations and I hear the criticisms that you’ve parried me with, but that’s what I feel. All of these platforms, at the end, are becoming about someone trying to sell you something else. AI is just accelerating that. I’m just wondering where you think the endpoint is.

It’s an interesting characterization. There are multiple business models out there. What you described as bits to atoms, I think is one way to look at it. I’m sure some creators feel like the ad revenue from YouTube is not enough. It’s because there’s an opportunity, right? Why would you not take an opportunity? I think “have to” is one way to describe it. “Get to” is a different way to describe it. The other thing I’d say is I don’t really think it’s quite accurate to say bits versus atoms. It’s much more advertising versus subscriptions versus purchases. And I don’t think the spread on that is really about the bit and atom piece. It’s about the connection piece.

There are a set of platforms that are built off eyeballs. What I built at YouTube was primarily built off eyeballs. Over all of history, the amount of advertising spend has always been some percentage of GDP. It’s covered between 2% and 4% of GDP forever. That gets divided up amongst all these eyeballs and that is one business model. Yes, the number of creators fighting for that has dramatically fragmented over the last couple decades on every platform. What can come from that is smaller. There’s also the ability to sell products. The ability to sell products is as old as time, and in the middle of that is the ability to build connections. Those products tend to do a lot of work with subscriptions.

It’s interesting when we think about some of my favorite creators, many of them subscribe to the 1,000 fans theory: that if you can get 1,000 people to pay you 100 bucks a year, you all of a sudden have a $100,000 business. There’s a whole class of people who have decided, “I can either go somewhere I get a little bit of money every time somebody happens to blink and look at me. Or I can get them all the way down the funnel to buy my hamburger or my water bottle. Or in the middle, I can build a deep enough connection with a person that they’re willing to pay me a substantial amount of money on an ongoing basis and I don’t need a lot of them. If I can do that, then I can build a real business out of it.”

There are some fantastic creators who have done a really good job of that. Many of the ones I’m sure you know. What I’d like to do and what we’re trying to do with Superhuman and our agent platform is enable people to build that level of connection. A lot of them are doing newsletters. It’s very meaningful to say, “I got a newsletter. It’s 100 bucks a year. Here’s how you can do it. 1,000 people gets me to 100 grand. 10,000 people gets me to a million bucks a year.” That feels like a meaningful connection.

In our case, I’m saying AI is going to allow us to do more than show up in your inbox. It’s going to allow you to show up with a red pen and a blue pen right next to the person and say, “I can help you in the thing you’re doing, at least the part of it that we’re working on.” And I’m willing to gamble that, can you go get 1,000 people to say “that’s worth 100 bucks a year to me”? I think you’ll be able to.

Wait, I’m just going to ask you this as directly as I can. Do you think that feature will be good?

It’ll be as good as the work that the creator puts into it. Are all newsletters good? No, most newsletters suck. There’s no guarantee that the newsletter platform can make them good. Is every YouTube video good? No, mostly they’re quite terrible. But does it allow–

I don’t know what your tool looks like to build an agent inside your platform, but I haven’t seen an LLM that can replicate my writing, let alone my editing. And you’re dependent on the capabilities of models themselves. So I’m asking you kind of a general way, but you know how your tool is built, can you actually make a tool that can do that well?

I think so. I would say that we did a pretty good job with Grammarly, that we replicated a grammar teacher pretty well. Can we do that with a broader spectrum of things? I believe so. We have some good evidence of it already with some of the agents working on our platform. Can we build a good one for you or can you build a good one for you? I don’t know. I’d love to work with you on it.

What does that tool look like? What does “build a good tool that lets me edit” look like?

It’s what you said earlier, you have to write down that viewpoint of like, what is your editing like?

No, I mean, literally describe the interface that your tool provides me to do that.

Oh, the big part of the interface is a prompt box in what we call triggers. You’re going to say, “Here’s my instruction.” Think of it like you’re going to publish your manual and here’s your trigger. Here’s a set of things that say, when you see this, do this. And here’s my manual, here’s how I think about things. And when you see this, do this. You gave the example of feedback on a headline. You didn’t like the feedback you gave on the headline. It’s reasonable. I wonder if you could write down what feedback you would give on a headline?

Let me suggest a different way to think about it. Pretend for a moment you were trying to train someone else. You’re saying, “Hey, I’m going to hire an employee and I’m going to scale myself and I’m going to teach them to be like me.” How would you teach them? You’d probably sit down with them and you’d write some things down. And then the second thing you’d do is you’d watch them do it and then you’d correct them.

The other piece we have to do is we have to say, you need to get feedback and you need to be able to come through and say, “That was a shitty suggestion. Don’t do that again.” And so that’s what that interface has to feel like. You give a set of instructions, you give a set of triggers, and then you get feedback. And you say, “This worked, this didn’t work.” You’re going to come back and you’re going to look at it and say, “Yeah, that clearly didn’t work.” Maybe it didn’t work for the user, they ignored your suggestions. Maybe it didn’t work for what you think was good work. You looked at the output and said that wasn’t particularly good work and you’re going to train it.

The idea of being able to train a custom agent for each person, for each product, is really interesting and compelling. I don’t think it’s going to be easy to do for everybody, but the people who do it well will be like the prominent YouTube creators of today. You’re going to make a very deep connection with a broad set of people in a way that you’re never going to capture with ad dollars or with selling water bottles.

Do you have an example of one of these that you think works well today?

I think Grammarly is the most obvious one. Most of the other really good ones—

Grammarly is like grammar, right? It’s rules-based and a very specific one. Grammar has rules, it has a logic. It’s squishy on the margin, but there’s good grammar and there’s bad grammar and you can pretty clearly detect the two.

It’s actually interesting. Grammarly is a stack of models. The base level model is actually spelling. Spelling is the very core definitional thing. Grammar has pretty good rules. Spelling has really clear rules. Grammar has pretty good rules.

But actually the reason why people use Grammarly is we go well beyond that. So we do advice on tone, we do advice on style. We do, “Hey, this is making you sound harsh.” These are all things you get when you pay for Grammarly. That’s the type of suggestions they get from us and they seem to like them — 40 million people use it every day. There’s a wide set of partners that have jumped onto the platform and built agents as well. Many of them are closer to tools.

So one launched a couple of weeks ago from Gamma that helps you build a really good slide deck. They did a lot of work to take “what did you write?” to “how do I turn into a slide deck?” We’ve seen a lot of them being built inside of companies. The sales example I gave, which is a very common one, is, “Hey, if I’m a head of sales, I have a sales methodology. You should always ask these three questions. You should always pitch our product in these ways.” They write those down, they turn it into an agent and say, “Make sure this is in front of people while they’re working.” And I think some of them are doing great.

Those are enterprise uses and I actually understand the sales use case a lot. You need the salespeople to all say the same thing all the time. I understand they don’t do that all the time. We have salespeople.

Actually, can a creative one work?

I’m asking because I don’t think taste is rules-based. Our producers are in the background here just in a puddle, because part of their job every week is to try to write like me. They get a lot of feedback from me directly on that. I’m literally editing the documents so I can read the intros and outros and I’m changing the questions. And it’s really hard even when it’s just three people who have spent years working together to try to get to an output that works. And they’re really good.

Yeah. It’s totally fair. My guess is the types of experts that will first prevail here won’t be the ones you’re describing. Those that make something creative, sound unique, make it sound better, are probably not the ones that’ll work first. But I do think there’s a set of experts and creators that will work great. Maybe I’ll pick the ones that are right next to Grammarly.

There’s a set of teachers for whom this is going to work really well. They’re going to say, “Hey, in addition to making sure your grammar is good, it looks like you’re writing something about history. I can probably help you cover history more clearly.” It’s not quite as clear as grammar facts, but it’s pretty close. “This is what happened in this period. You should know these different elements of it.” Teachers will be a great example of that.

What are LLMs really good at? They’re really good at averaging what everybody says. So can they do something really unique like you do? No, probably not. Can they take some part of your suggestion and turn it into something useful enough that you can get 1,000 people to pay 100 bucks a month? I bet you can come up with something because the bar isn’t high.

I know we’ve flipped the conversation around a little bit. If we’re talking about you and your business opportunity, you don’t really need to replicate yourself the way you would be in person. You just need to create enough benefit that 1,000 people pay you 100 bucks a year. That’s what you need to do. Is there some part of your methodology that you think is so good that people would do that? I bet there is.

I’m going to have to think about that quite a lot. Thank you so much for coming on, for answering the questions, for being game to answer the questions. I appreciate it.

I have a lot of other questions. We’re going to have to have you back sometime soon to expand the full scope. What’s next for Grammarly? Tell the audience what they should look for.

We’re very busy building out Superhuman Go. We have a big set of launches coming in the next couple months, so keep an eye out for that.

All right. Shishir, thank you so much for being on Decoder.

_{Questions or comments about this episode? Hit us up at [email protected]. We really do read every email!}