Ethan Mollick on Staying Human at Work | Dreamforce 2026

Thank you for having me. This is great.

I really appreciate all the pre-applause. It really raises No, I don't. No, it's just scaring me, guys. I have to I have to meet it after the fact now. I feel like a beggy for applause. This is going very badly. Let's keep going.

So I mean of the rules in my book I these four rules of the first book the one that has become most obsolete most quickly is probably the idea that you have to be the human in the loop at least in the way that we were viewing it before like if you think about co-intelligence is the chatbot era right you ask an AI to do something and give you an answer you get an answer back back and forth and like the truth is that as of about you know eight months ago nine months ago we entered really true agent era agents were mentioned all the time before that but now we have the actual agent thing people have been talking about that I've been writing about for a while which is asking AI to do something that does it and I have all sorts of weird projects that are you know they I can now delegate write a unique paper in economics and I get a paper an hour later I can say make a video of me singing um you know in a 1980s rock band about agents and I get that like I don't know what it's doing I don't know how it's delegating stuff it's making lots of choices lots of judgment calls and so I'm in the loop the way a manager is in the loop right I am assigning a task and so that intern model of AI which is like it's going to make mistakes it's okay that non-threatening sort of angle is sort of vanishing and in some ways that's great much more powerful and another way it's kind of unnerving because I can ask for things and they get done I don't actually know how they're getting done and is an agent swarm attacking everybody right now to get information I don't actually know what's happening so it's a big change

All right, so one of the big concerns I hear a lot about is like do we stop thinking? Um and the answer is like uh probably about some things. Yes. Right. Like what technology does is deskill us. Um my grandfather um was uh was an engineer who built the um fire suppression system at Cape Canaveral and he did it with a slide rule and matrix math that I have no idea how to use a slide rule. I can't do matrix math in my head. I can't do matrix math at all. Don't tell anyone because it's essential to AI. [laughter] Um but um it is um you know but we are in a we're like we desilled on that. On the other hand, there's things we don't want to lose skills on. And I think AI is absolutely going to take away some of your knowledge and ability to do things. What we have to do is be pretty thoughtful about what we want to give up. Um, and that's hard because, you know, we've seen this in education first, right? The first thing everyone does is ed like learning is hard. Um, and it feels annoying. Why do I have to solve the same problem 15 different times? I'm just going to let the AI do that. Then we don't gain the basic skill. So, we're not very good at deciding what we should be skill on. That's part of what companies and um and and schools and things need to help us think about. But, I mean, we're absolutely going to lose some skills, but that's not the same thing as damaging your brain, right? like we get to decide what we're outsourcing and what we're not and people who are deliberate about that are going to end up in a very different place than people who are not deliberate.

So okay and I will say there at one point I'm disagreeing with you. I think human judgment is overrated. Uh not not because it's not good. human is important and valuable but we I worry a lot about shining lines in AI discussion which like it was creativity before it was judgment like there is no if you have a four hour a four day long running task the exercising judgment all the time on that task constantly making decision constantly exercising creativity you cannot do tests without judgment and creativity it may not be your judgment and creativity that's an important distinction but to say it can't do it is a problem right um and I don't think I disagree with you on that so you know but I think people take comfort in stuff that I wish I could give more comfort like oh this is irreplaceably human that doesn't mean humans aren't irreplaceable but like I think we have to make decisions about what we want to give up and what we don't on the bigger issue of training like you know I worry about this a lot because I teach generalists at Wharton and right now everything is kind of screwed up in education we'll figure it out I mean education is always screwed up we're going to figure out the right path through AI as tutors are really powerful we'll we'll we'll find the right balance of testing people in classrooms or not we'll be fine I'm worried about you guys because I and all of you there because I send you generalists and they become specialists same way we've done specialists for 4,000 years apprenticeship, right? And apprentichip has often been a really good deal. I send you someone who's really junior who you can underpay and they'll do the tasks your middle managers don't want to do and they'll show try and show you whether they're good or not. You can choose to hire them later and in return they learn the ropes, right? They get feedback. Hopefully a really good mentor. Sometimes the mentor yells at them. Either way, they learn what they're doing and it it pays off, right? And they get paid for it. And that all broke down because if you're an intern, you've realized that the AI is better than them at better than you at doing the work because of course it is because you don't know anything. So you you're you'd be dumb not to use AI for everything because you want a job. And if you're a middle manager, you've realized that you can assign work to the AI around the intern. You don't have to manage people and people sometimes have feelings and stuff and you don't want to deal with that. So I'd rather have the AI do all the work. And companies are like, I want efficiency and learning is inefficient. So if you everyone use AI for everything, we get higher efficiency and therefore we don't need, you know, who cares what happens long term. And the trading pipeline just snapped entirely. And the way forward in this has to be starting to realize we need to build inefficiency in some ways into what we learn. Either you're going to need to turn to universities to help you think about how you teach people, how we assess people, and we'll have an extra year of Salesforce University at various schools like the way we do with MBA programs or else you're going to have to take time and make teaching actually central to what you do and say we're going to have to assess people without AI and we're going to have to have moments where they're inefficient and they're getting feedback from humans. And I think the companies aren't always ready for that. The companies that do will will thrive though.

I mean, it it it's bad in a bunch of ways. I mean, one of them is nobody likes verifying stuff. It's the worst job. Like the worst job in the world, like every time I write something and there's a proof read like I feel so bad for them, right? Like there's no way I we've all pro proof read stuff. It is awful work. So if your job is reduced to all the interesting works done by the AI and your job is s is both proofing it and then agreeing that you will be fired if it's bad. That is the worst possible outcome in an AI world, right? And it doesn't have to be that way. I worry that if you we just let the sort of AI world progress the way it will, that's what we end up with is like humans are you get fired um and you're responsible. Like it's just awful. And instead we should be thinking about the way we do this with other labor saving or drudgery saving techniques. How do we get AI to bring us the most interesting problems and do the worst stuff so that we our life is a series of interesting decisions and moments where we can be additive and useful ways. And I I worry a lot that unless we do design work around that, we end up with the default which is how do we automate work, right?

No, I mean I agree. I mean, I think I just worry about everything being a temporary state of affairs. Like, you could design a world right now with the current the current systems are jagged enough, right? They're good and bad enough stuff that you cannot hand over large job functions to AI yet, right? Like it makes judgment like all the comments are sort of the same. It sort of looks the same. It's like any of you who have turned to like everybody who turned to AI writing six or eight months ago, like it looked good for a little while until now everybody gets plottish at first like and your eyes start to bleed as soon as you read a LinkedIn post now. It's like, you know, tells you it's carrying the weight and that, you know, the um and like it's just the worst, right? Like, but we're not going to stay there forever, I don't think. And part of this is how do we build for the future systems we want rather than just being reactive to what the systems could do.

So we have a lot of studies of this creativity is really hard to measure. Okay. So we measure it with things like the rat test or mode association test where you're supposed to name f you know 10 very different things from each other or you know how many uses can you how many uses can you make of a mouse trap or that like those are the kind of measures we have and you know they they're good for human creativity like every other benchmark they sort of break down with AI but AI out compete almost every individual human in terms of the creativity we have however we have a lot of research that also shows including some being done by my lab that shows that the ideas tend to cluster more than a human right like the AI Yeah, if you you're used to this, right? Like it's like AI is actually not a bad writer. It's just one kind of writer, right? AI is not a bad thinker, but if you use AI a lot, you probably see that it has like the obsessions of a 2022 era Silicon Valley person. Like it loves NFTTS and VR and will always like suggest those kind of like and you know and and carbon neutral setting. Like it has a bunch of stuff it's really interested in. Um and it comes back to those themes. Now I can tell you we have research that shows better prompting techniques and other approaches will make the AI more diverse, more like a set of human ideas. But it's still not as diverse as really bringing humans to the table. We generate ideas people don't. So one source of human advantage is variation. We bring more v variation to the table and you don't the whole world isn't clogged.

So there's a few things. One is it has access to the data but it's trained on that data. And then there's a series of training steps the AI goes through to enforce the way it thinks. And one of the weirdest things about AI, by the way, about these systems is whether you're talking about, you know, OpenAI Anthropic or you're talking about uh the the Chinese open weights models like Kimmy or you're talking about Grock, which was specifically built to be very different. They all have almost the same kind of political views and value system. And we don't 100% know why that's the case. Some of that has to do with the training system. Some of it has to do with the data. Is it possible to push AI from its sort of centrid area? And by the way, the wrong thing to think about is that AI gives you the average answer of the internet. It does not. It gives you its own thing that's much better than the average answer of the internet for a bunch of really interesting reasons. But regardless, it tends to be fairly similar. Now, you could destabilize that a bit by being giving interesting prompts, but then you're also adding your own variation to this. That might be a solved problem at some point. Like, but right now, that is a very large jagged frontier gap.