Jobs that in the past would have been at risk of automation in things like factory work in previous generations, now that same phenomenon is going to play out in knowledge work. Right now, it's like the AI industry has a gas pedal but it doesn't have a brake pedal in the car. And what we're saying is we want to do some of the work required to build that brake pedal. So we in the world have an option. Many of the regulations and policy frameworks of today were not built on the assumption that artificial intelligence systems would exist, let alone get this powerful. I am worried for my kids if we as a society don't have a serious conversation about what the implications of AI's continued advances mean. Develop a hobby. Anyone who has a hobby has something that they're passionate about and that they know more about than most people. And with that hobby, you can have curiosity, you can have ideas, and you can use that to really get the most out of these AI systems. And I am sure turn that into amazing jobs, jobs that don't even exist yet. And it requires you to experiment with the systems and have that curiosity. So that's my message.
Are you seriously saying that your systems are on the cusp of essentially building themselves with minimal, perhaps no input from humans?
We're saying that the dream of the whole AI industry, which is 70 years old, of being able to build a generally intelligent system that might be able to help do its own science, now seems like it's in sight. And we're trying to talk about this early because it's an important issue for the world to understand.
And just unpack what that means, that you have new measurements of the progress from what's going on inside your own company and the progress in terms of coding, for example.
So what I'm seeing is a dramatic increase in the amount of code that is being moved around Anthropic as an organization. Coders here are now writing about eight times the amount of code than they did in 2021, 2022, 2023, and 2024. We've had this dramatic acceleration and that has caused us to see this increase in the rate of research and the rate at which we're covering new ground. So something is beginning to speed up about AI development itself.
Okay. And in practice, what does that mean in terms of how much of this is derived from the AI and how much of it's derived from human coders?
So 80% of the code that goes into Anthropic now comes from Claude. It comes from Claude and Claude-related systems. So the majority of the code in our organization is now written and derived from our AI systems itself. And that accompanies this dramatic increase in the productivity of all of the people working here.
And this could get to 100%. This is when it starts to essentially make itself, the AI makes itself, improves itself.
That's exactly why we're sharing this information. That's not certain, but it would have huge implications if that did happen. And so our belief is share this information with the world. That's what I do at the Anthropic Institute and talk about these issues early because they have amazing implications for both science, robotics, the general pace of discovery in the world, and also they have some potential risks which we're going to need to discuss as well.
Well, give us some examples then of what it could mean when we have this 100% AI generating itself.
So, I'll give you two examples. One, biology. Biology is a domain that's quite difficult for AI to work within because you need to run lots and lots of experiments in the real world. For AI systems to get much smarter and be able to build themselves, they'll also be capable of doing science in the real world. And so you can imagine human scientists working with AI systems to cover a lot more ground in biology than they have before. Similarly, robotics. Robotics is an area that is really, really tough for AI systems today. You have to do a huge amount of work of adapting the AI system to be able to work in the real world. It breaks all the time. It's a hard science problem. AI systems capable of this kind of recursive self-improvement would be able to adapt themselves into domains that AI has struggled with so far.
So just to be clear though, what you're saying is that this is the moment where the AI sort of decides its own pace of discovery. It does it itself with minimal human input.
The world we're looking towards is one where it's as if we've added millions more scientists to the world and we get to make choices about where we direct them. So it's not that the AI system is going to be making all of the choices itself. These things will be governed. They will have policy frameworks around them. Labs and governments and others will work with them. There will be real choices for society of which parts of science do we want to go faster, which parts of the economy do we want to go faster? We will have choices that we haven't been able to make before, but will have amazing implications.
But you're also stressing the potential problems here, aren't you? The potential serious problems. And you seem to be raising the question of control. I mean, are we still in control of this process?
Yes. I think a good way to think about it is what would happen if you times by 10 the number of people working at the BBC right now. You'd suddenly have these questions of how do we know what they're doing? How do we validate that the work they're doing is good? How do we make sure that the reporting they're doing is up to BBC standards? And there'd be this huge amount of work we do to make sure that these new kind of virtual employees were working on our behalf in ways that made sense to us. That is the same problem that we're going to have to deal with and work on in many, many domains. All organizations and societies are going to have to contend with a world where we're getting a lot more done and we need to figure out how we trust these systems that are doing that work for us.
I just want to, you used to, you described this as if you were having thousands more employees. Is that how it works? Are these, I think you call them agents. Are they sort of like having employees? Is that how you treat it? And each human has like, in your company, has like 10, 20, I don't know, 100 of these people working for them?
I think that's a great way to think about it. You know, I went on paternity leave in November of 2025. I came back in February of 2026 and I discovered that some of the teams I work with, I had colleagues who were now working as if they had large teams of people around them, but in fact they were just working with many, many copies of AI systems. And similarly, I'm building a bunch of new teams right now. And we've changed how we're approaching hiring. We're still hiring people, but now instead of hiring, say, a big set of engineers, we're now hiring a bunch of interdisciplinary experts like lawyers or philosophers or others because the technical work has been taken care of. So smaller teams, faster moving teams, and everyone at the company is more like a manager than an individual employee now.
And so you're saying this for coding, but does it apply? It surely does apply in other areas particularly of knowledge work, creative writing, creative directors, legal work, accountancy.
It feels like it's going to apply in most domains. And in part that allows everyone to do a lot more than they were able to do before, but it's also a challenge. I think many people working today don't think, if they're an individual practitioner, what would it be like if I had five colleagues or 10 colleagues, what could I get done? And that's the kind of mindset change that this technology implies.
And what sort of jobs therefore are under risk, likely to be replaced in this world of AI agents that can actually kind of improve themselves without human input?
So as we say in this report, I think there are open questions about whether AI systems can be truly creative, truly come up with interesting off-the-wall ideas that there is not really evidence for that yet. And what we see happening with work at Anthropic is we're now limited more by the ability to generate good ideas than the ability to do the engineering to turn those ideas into reality. So everyone who generates a bunch of ideas, is creative, has kind of entrepreneurial ideas, is going to be advantaged by this. And if you're in rote jobs, highly repetitive jobs, jobs that in the past would have been at risk of automation in things like factory work in previous generations, now that same phenomenon is going to play out in knowledge work. And we need to figure out how do we help people change their mindset about what sorts of work they're doing and move to these more creative jobs where they're going to be able to get a lot more done.
Well, I mean, you use this phrase knowledge work. I mean, that covers large swaths of, for example, the British economy. That covers the sorts of jobs that many graduates think that they are going to create careers in and you're suggesting with this report that they can't count on that.
We're hiring many, many people including early stage graduates here today, but the sorts of graduates we're hiring are the ones that have a bunch of creative ideas that they want to put to work and a bunch of kind of an entrepreneurial mindset. I think that story is going to play out across the larger economy where we need to figure out what are the skills we're teaching people and what are the types of roles that companies hire for. So the implication of this technology and AI's advance in general is significant amounts of economic change. We can't predict exactly how that change manifests but surely some of it will be a change in the makeup of some jobs. That's what's happening.
Give us a sense of that. An AI agent very quickly, in the next, before 2030, could replace a lawyer, an accountant, a copywriter.
What we see today is, on our legal teams and our accountancy teams here, these teams have the property of being hybrid teams. They have a bunch of people that work on it and they have a bunch of instances of our AI systems that they're working with to get stuff done. So I think the story is generally any profession is now going to be able to work with AI systems and change how work happens as a consequence which also means any single profession has some potential for AI to come in and augment or automate different types of work and we're trying to measure that right now. There aren't clear measures that say that augmentation or automation is happening at large scale, but clearly that's the implication of this and we're trying to share data about it so that we can see if this shows up in the broader economy.
So right now you're saying it's not showing up, but you're trying to prepare society for a shock, a shock to the jobs market and what people consider to be work.
Exactly. We're trying to prepare the world for potential change. What we see at Anthropic is the very early suggestions of both the technology getting better and the sorts of teams that we're building changing and the ways that people work are changing. Now, how robust that is, how long term that is remains to be seen. Our philosophy and that of the Anthropic Institute is get this data out there early, share it with the world and sort of discuss this with the world and then figure out if this has big macroeconomic or societal implications.
Well, it seems that it would, but it's also having an impact on safety, security. There are big concerns about this. Just want to go back to this issue of control here. If an AI starts to improve itself, 80% you say already internally, gets to 100%, what, in the next year or two? Is that realistic?
It's plausible, but I think that it's also a choice as to whether you let AI systems get that far. Something which we discuss in the post and why we're sharing this is we think this is a topic that the world should be talking more about and a topic the world should make decisions about, not only private companies.
So we might want to stop the 80% that you're seeing internally in terms of coding being done by AIs getting to 100% because the consequences would be what for safety? The AI could start deceiving us, could start developing itself in ways that we can't keep track of, that we do lose control.
Yeah, you want the option to be able to take your foot off the gas and put your foot on the brake. Right now, it's like the AI industry has a gas pedal, but it doesn't have a brake pedal in the car. And what we're saying is we want to do some of the work required to build that brake pedal. So, we in the world have an option. It's not obvious today that you want to do that, but absolutely, as you say, at some point in the future, you might say, let's get all of the benefits we can for say biology and medical research and let's take a pause or take a moment on AI research where we can absorb the societal changes implied by this.
So as it gets closer to 100% and you say you want to do loads of biology, can your systems stop a sort of self-governing AI from developing a bioweapon with this biological or medicines knowledge for example?
Yes, I mean that's work that we and others in the industry have been doing for many years now. Bioweapons, cyber weapons, governments pay attention to this. But this is all the kind of thing that we're going to need to talk about more as not just companies but societies and eventually figure out what regulatory frameworks are such that you kind of bind industry around common standards here because no one wants an out of control system building bioweapons and you can create laws to prevent that and we have a technology that prevents that today.
And obviously we've seen this play out with your product that's yet to be released, Mythos, which sort of percolated out into a very tight-knit group of companies and has just gone out to a few European companies too. So it's right to say that it had capabilities that you did not predict and capable for example of finding security holes in banking systems in all sorts of critical systems.
Last year we started writing, just as we are writing now, about RSI, about the potential for AI systems to get a lot better at cyber and we built the evals and tests that would tell us if that was the case. Then when Mythos came out we discovered our AI systems had got a lot better at cyber. That led to us rolling it out differently because we want to give the world time to deliberate about how to use this technology, work out how to get the most good from it, and then as we're doing now, expand coverage so more of the world gets access to this technology to use to harden cyber defense. I think that's the story that's going to play out with AI in the coming years is AI systems will become better at society relevant things, not just commercially relevant things. And we as both an industry but as governments and as people are going to have to make choices about do you just generally release this, do you release this generally after a period of experimentation which is what we're trying to do with Mythos. These are the sorts of questions coming up for us all.
