Why CEOs Are Getting AI Wrong — with Ethan Mollick | Prof G Conversations

I don't think that age-gating is a particularly bad idea for synthetic AI characters that try and act like people because we don't know what the effects are. I think it's easy to be alarmist and catastrophic about it. The effects may end up being very good. I don't know, but neither does anyone else.

I'm outside of beautiful Philadelphia, Pennsylvania.

I'm at my home. Yeah, that's my game collection back there.

So I think that there's always debates, right? There's like external facing, but when you talk to these people internally, I think Anthropic is fairly sincere about their views about how AI works. You may or may not agree with them. He actually has a pair of essays, one on the bright future ahead of all of us and the other about our potential doom, and pointing out issues that may actually occur. So it always is a question of weirdness that you're building this thing if you're so worried about it. But I think it is a sincere anxiety.

So I'm in a weird boat here, which is I think that there's a lot of worries about the existential risks of AI and so people leap ahead five years, assume the current path continues, and there's no sign yet by the way that AI is slowing down development, but there's a move towards sort of existential, you know, Dario in that essay talks about what would a group of geniuses in a data center, they're smarter than any human, what would they do? I'm actually much more concerned and thinking about how we guide the next few years to make AI help people thrive and succeed rather than the negative consequences that could happen. How do we mitigate those negative risks? So I think there's a nitty-gritty path between here and some imagined future. We don't know if AI is going to get there to sort of super powerful and autonomous, but we do know it's disruptive today. So I worry a lot about how do we model the right kinds of work so that when we start using AI at work that we do it in ways that empower people rather than fire people. How do we think about AI in education so that it helps students learn rather than undermines learning? How do we think about using society in ways that don't lead to deep fakes and dependencies? I think there's two sides to each of these coins and we need to get very nitty-gritty about which things we care about.

I mean sure. I mean, but some of it is also they drink their own Kool-Aid, like they believe this stuff whether that serves marketing or not. But I do want to take a step back. I'm a business school professor, right? You know, like you, and so I've been doing a lot of work with my colleagues on impacts of AI at work. And there's a few things. One is there are fairly large impacts. In any randomized control trial we see — we did an early experiment with my colleagues at Harvard, MIT, University of Warwick, at Boston Consulting Group. We found 40% improvements in quality using the now obsolete GPT-4 with people who weren't even trained. 26% faster work. Penetration rates are up there. It's interesting. Companies, people are using AI, but they're not talking about it. They're not using the corporate AI. So about 50% of American workers use AI. They report, by the way, three times productivity gains on the tasks they use AI for. They're just not giving that to companies, right? Because why would you? Like you're worried you'll get fired if AI shows you that you're more efficient. You look like a genius right now and maybe you know the AI is the genius. You're doing less work. So I think that there's a difference between what companies are seeing about adoption, what's actually happening with adoption.

So I think the big picture overestimation and underestimation is work is complicated and organizations are complicated, right? So you can get lots of individual productivity gain, but if that's producing 10 times more PowerPoints than you did before, that's not necessarily going to translate to any actual benefit for the company. So leadership needs to start thinking about how do you build organizations around this. At the individual level though, huge impacts, and you know, coding especially has taken this massive leap. We have earlier evidence that you saw about a 38% improvement in the amount of code people were writing when they started using agentic coding tools with no increase in error rates. But that's even increased further. The newest coding tools — both the people in charge at the research level at OpenAI and Anthropic have said 100% of their code is now written by AI. That's actually quite believable given how good these tools have become. We're seeing similar things, you know, managerial tasks, medicine. We're seeing impacts in scientific publication, super interesting area. People who started using AI early to write scientific papers, and we know this because there's a great study that looks at when they started using the word delve, which was a dead giveaway you were using AI back in 2023. If you used delve a lot in 2023, then you actually published about a third more papers in higher quality journals afterwards. Now the question is, is that good for science to have more AI writing? Separate issue. So that's the sort of process versus detail problem, right? People are becoming individually more productive, the system isn't built to handle a mix of high quality and low quality, and is more work, and that's where the bottleneck often is.

So I think that equation starts with a realization, which is nobody knows what's going on, right? Like I talk to all the AI labs on a regular basis. I don't take money from them, but I talk to them all. I do research on this. I talk to policy makers and CEOs, and it's not like there's a playbook out there, right? We're a thousand days into after the release of ChatGPT. Like everyone's figuring this out at the same time. I'm seeing companies getting incredible amounts of benefit and other companies struggle, and part of that is how much they're willing to embrace the fact that they have to do R&D themselves. So part of the value of giving people access to these tools is experts figure out use cases, right? If you're doing something in a field you know well, it's very cheap to experiment with AI and figure out what it's good or bad at because you're doing the job anyway and you easily look at the results and see whether they're good or bad results. If you're paying someone to do R&D for you, that's a very expensive process. So people are inventing uses all the time. So the most successful case I'm seeing are a combination of what I call leadership, lab, and crowd. The leaders of the company have a clear direction, set right incentives to make things happen, think about process. They give the crowd, everybody in the organization, access to these tools to use advanced tools like Anthropic's tools or OpenAI or Gemini. And then they have an internal team that is actually thinking about what you build. So they're harvesting ideas from other people. So I'm seeing this happening everywhere — there's certainly a lot of stuff happening with internal processes, security, customer service, lots of stuff on analytics. The AIs are quite smart. So if you let them do analysis work, you can actually get really big impacts from that as well. Just wide ranges, but very different across organizations depending on where their expertise is and how aggressive they are about trying to experiment.

It's a great question. First of all, you know, you started this off by talking about marketing. Part of the problem with AI is there's, anytime a new phrase comes out, there's a blur of confusing different interpretations of it because everyone wants to sell AI products right now. So it's really easy to get bogged down. So agents basically can be defined as an AI tool that is given access to tools. So it can do things like write code, search the web, and do things that when given a goal can autonomously try and accomplish that goal on its own and correct its course if it needs to. So an agent would typically be something where you could say, hey, I'm going to have Ethan on this podcast, research everything about him, come up with a pitch deck on why we might want to have him on the cast, talk about interesting things that he might have said before and then boil this down to five really good questions to ask and it would go out and do the research and 20 minutes later you get kind of a complete result. That's an agent at work. So agents are basically the chat bots that you use today when you go to ChatGPT Plus, we call an agentic harness a set of tools and capabilities they have — searching the web, writing code, connecting to your data — that lets them do more work. So when you combine those two together, that's where you get semi-autonomous AI.

Okay. So, the good thing about AI is it's very democratic, right? There's no better model than the ones you have access to today. You or every kid in Mozambique has access to the exact same tools that are at Goldman Sachs or Department of Defense or anywhere else. There's no better models. They're basically being released as soon as they come out. That being said, the really good models tend to cost you at least 20 bucks a month. So you are probably going to want to subscribe to either Google's Gemini product, Anthropic's Claude product, or OpenAI's ChatGPT product for 20 bucks a month. And you're going to want to, when you do serious work, pick their advanced thinking model. So GPT 5.2 thinking is important to use. Anthropic 4.5 Opus and Gemini 3 Pro. Those are the sort of starting pack of tools you can use. They're all capable of doing agentic work. You can access them through the chatbot. And I always recommend people just start by trying to do stuff they do for their job. Ask it for everything you do that day. Just ask the AI to generate some ideas for me, give me feedback on this, help me write this email, create the presentation. That will help you map the jagged frontier of what AI is good or bad at. And it's a really good starting point. Like there's a lot of other complicated stuff. If you want to do research, the deep research tools for Google are currently better through this product called NotebookLM. And that's free and that's very good. If you want to do coding, you probably want to use Claude Code, which you have to download. But the basics are pick one of the big three, pay the 20 bucks a month, and then start using them. You need eight or 10 hours of just talking to it like a person and seeing what results you get.

Yeah. So to take half a step back, on what drives the underlying dynamic is something called the scaling laws. And the scaling laws basically tell you the larger your AI model is, which means the more data you need to build it, the more data centers, the more electricity, the more chips, the better your AI model is. And it's very hard to build a small model to compete against the larger model. They're just better at everything. You could build, you know, once you have one of those, you could do all kinds of variations, but you have to build a big model. And there's a bunch of other tricks that you could do on top of that, but that's pretty critical. And because of that, there's only a few companies that could actually play in this space, right? So in the US, we mentioned the big three, which is Google, Anthropic, and OpenAI. There's also Elon Musk's X, which has been scaling quite quickly — xAI — and there's also Meta, which has been quiet recently, but is spending a lot of money in this space. Outside of that, there's a lot of people with smaller competitive, but they're not really competitive. Amazon, Apple, they don't really have their own models that compete. There's also three or four big Chinese companies that are producing very good models or releasing them for free to the world and one French company in the same boat. So within that dynamic, there's this competition about who could build the biggest data set or who could train the biggest model because bigger models are smarter, who could put the most research and tricks into them. And it really is interpersonal in some ways — like the heads of these companies are really out to get each other, right? Like they do care about winning this race. They think they should be dominant. And so there is a lot of resources being put into getting ahead of the other people in this space one way or another. So right now the sort of three most polished models are Google's, OpenAI's, and Anthropic's, and again which one is better is changing on a day-by-day basis or at least week-by-week basis as each one releases new approaches. And then we're waiting to see if anyone else kind of catches up to them. But those three are in a very tight race. As soon as one of them comes up with a product that uses AI in a new way, the other two copy it, right? So Claude Code is currently the very hot coding tool. OpenAI has Codex, which is a very similar thing. Gemini has its own set of tools. Deep research was invented or first came out from Google. Now there's deep research projects from Anthropic and from OpenAI. So you could kind of pick any of the three of them and be in good shape as long as they could keep growing and spending money and they don't hit a wall in development, which hasn't happened yet.

I can. Right. What I worry about is trying to talk to all the various levels, right? What do you do if you're just starting off? Pick any of the three, you'll be fine. But I think people who use them a lot, they have personalities, right? Those personalities are shaped by the companies, the way they train. I mean, it's amazing that they're all so similar to each other that things basically work across all three. Like, you wouldn't expect Microsoft and Apple to produce a system that works exactly the same. These are similar enough that for most people it doesn't matter, but if you care, right, Opus 4.5, Anthropic's models are tend to be known as the best writers of the bunch. They're often quite good at sort of intellectual topics. They're a little fussy in terms of, you know, they have high ethical standards relative to the other models. ChatGPT is really two different flavors of models. There's a set of chat models that are really optimized for you to have conversations with and roleplay and be friendly. I don't tend to use those much because I tend to focus more on the work aspect. And they have a series of very logical, very good at long task models that are very good at producing a lot of work. And Gemini is an interesting set. Very smart overall model. Weirdly neurotic — like it actually gets kind of self-flagellating if you tell it it did a bad job. It apologizes and kind of grovels. Weird kind of dynamic there. So they all have their own sets of personalities and approaches.

Well, I wouldn't call it regression to the mean. We are seeing a race, right? There is huge impact. Each model generation is much more capable than the one before, right? So we keep crossing these lines where oh, the AI can't work with Excel and suddenly it works with Excel better than most bankers and does a discounted cash flow analysis. Or the AI can't produce a PowerPoint and suddenly can do that. Or it can't do math and suddenly, you know, last year two models won gold at the International Math Olympiad. So there's not a regression to the mean because there's no drop-down of ability level — the ability levels keep going up. But all of the companies in the space are on roughly the same development curve, right? Their models keep leapfrogging each other by a fairly predictable amount over time and you can draw a pretty good curve on any benchmark that you want that shows the same exponential gain in AI abilities. So which raises the big question of like so what happens in the long term? And I think that depends on one of what the long term of AI looks like. There's one version where we just keep having a race of capabilities and you need to stay ahead and you pick a model maker and as long as they stick with you you keep paying them money. There's a version where one of them achieves what's called takeoff — their AI models become self-improving and they build the smartest possible model, no one can catch them, and build artificial general intelligence, a machine smarter than a human at every intellectual task. That there's some sort of apotheosis or endgame. Or there's a version where everything sort of plateaus out and then people spend billions of dollars building models and eventually free Chinese models or another company catches up and there's no money and it becomes commoditized. I don't know which of those three scenarios dominates.

I mean there's something interesting going on because an open weights model, which is a model that you release publicly to the world that anyone can run, right? So if I want to use ChatGPT, I have to go to OpenAI and use ChatGPT to do that. If I want to use one of the Chinese models like Qwen, any company in the US can download that model and run it themselves. So that model based on open source made sense for software because I could give away my core software for free but then sell you services. It doesn't actually make a lot of sense for AI companies because they're building a model and giving it away for free. There's no ancillary benefit to that. They don't get a gain in the long term. They're not selling other solutions. They have no special prize or tool left behind in most cases. So there is a little bit of weirdness about how long will Chinese companies sustain releasing free models. They're about eight months behind and consistently eight months behind the frontier of US models. And what's driving that, right? Is this a state-sponsored effort in the long term? Right now it's not clear that it is but it might be that there is some sort of dumping kind of effort. On the other hand, the degree of intelligence is fungible. Like if you are talking to a CEO and they're saying we're going to use a Chinese model because it's cheaper, the cost of models has dropped 99.9% for the same intelligence level in three years, you'd be like, you actually for most applications want the smartest model that's most capable of doing tasks as cheaply as possible. So fixating on a model that's not as good may end up being a problem. Like this isn't an equation where we're done yet and we could pick among roughly equivalent products because we're racing up a curve of ability that's still changing over time.

Yeah. And there's a few of them, right? And they're kind of jockeying against each other. So as you pointed out, data centers are the sort of choke point, right? How fast can I build one and especially how fast can I power one and can I get enough chips to put in one, right? So the power and building and chips are all a big deal. For a while, data was the bottleneck, but AI companies have increasingly found that they can make their own data. So it turns out as long as you have some human data, large language models can create their own data and other models can train on that and you get good results. So data is not the choke point — it was, but it could be again. There's also a research choke point. There's a lot of things that LLMs do really well, but there's some parts of the jagged frontier that are still very jagged, right? LLMs don't have memory. They don't learn things over time. So I have to instruct them every time. It's like I'm talking to an amnesiac every time I speak with an LLM. So continual learning is a problem that gets in the way of building these amazing models for the future. They don't keep learning, like humans keep learning, otherwise you have to train them every time. So there's research bottlenecks, there are energy, power and data center building bottlenecks. And those are the sort of big ones right now. From a policy perspective, energy is the big one that all the AI labs are worried about. They could reliably turn energy and chips into money right now. And the question is how fast can they build those data centers?