OpenAI’s warning: Washington isn’t ready for what’s coming

Yeah. First of all, these are ideas. I think an agenda is too strong of a word. We want to put these things into the conversation. Some will be good, some will be bad. But as you said, we do feel a sense of urgency. And we want to see the debate of these issues really start to happen with seriousness. Again, there probably be many more solutions, better solutions than we can think of ourselves and the broader world. But we do think that the time is now to get serious about debating these issues so that we can begin to have some solutions in place before the change that we see coming really comes down. We may be wrong. It is possible that even if the technology stays on the path that we expected to stay on, society takes much longer to adapt to some of these changes than we would expect. But it does feel like now AI is beginning to do increasing amounts of real work, and some of the transformations that we've been wondering about over the last few years feel like they're starting to happen. So as these models get very good at writing code, at doing other kinds of knowledge work, particularly as they get good at doing kinds of new science, the economy, the way our society functions, the shape of work, we think those things are going to change. And we'd like to see people begin to think about how that should go.

Part of what we try to do throughout the history of OpenAI is just talk about, with appropriate uncertainty, what we see coming. And we have not always been right, but sometimes, pretty often, I think we have been. And I think we have a responsibility to play a sort of role as an educator on what we see coming down the research pipeline. For a long time, we've said things that at the time we said them originally were considered like, you know, some more polite version of crazy. And again, sometimes we've been wrong and we'll be wrong about some things in the future, too. But part of our job, and I think part of how we can contribute to society is say, here's what we see coming. We may be wrong. We may be right. If we are right, here are the kinds of ideas that we think might work given the particular shape of what is coming. But it's obviously not up to us to decide. So our goal here is to put some ideas out early for debate. And in the same way that the ideas we were putting out a few years ago were early, but some of them turned out to be important, I think some of these ideas will turn out to be important too.

We talked about this earlier, but already the models are incredibly capable and having, you know, real impact on science and the economy. Two of the most important drivers of how we live our lives, our quality of life. But the next generation of models will be a very significant step forward. And as we prepare to get ready to launch those, I think this is an important conversation to have.

Let's say that the current models can help scientists make small discoveries. I really don't want to overstate what the current models are capable of, but you see these amazing things from scientists on, you know, Twitter or other platforms saying, I use GPT-5.4 and it helped me do this amazing thing, or that amazing thing. And I would expect that with the next class of models, you start to see people say, this helped me make the most important discovery of my decade or maybe my career. So that kind of level, like, you know, maybe not win a Nobel Prize on its own, but like a significant, career-defining discovery. On the kind of knowledge work side, I suspect that, you know, current models, can you hear people say maybe they're like twice as productive or three times as productive as they used to be as a coder. And maybe you'll start to hear people say, I'm able to do like the work of a whole team with these tools. So if it's, you know, me and x hundred GPUs, we can do the work of a whole software team. So that's quite significant. On the scary side, the main areas we currently track in our preparedness framework are cybersecurity, bio, and actually any issue of time. Let me stop at those two. I suspect in the next year we will see significant threats we have to mitigate from cyber. And these models are already quite capable and will get much more capable. And then on bio, this is something we've been talking about a lot. The models are clearly going to get very good at helping people do biology at an advanced level. Wonderful things are going to happen there. We'll see a bunch of diseases get cured. Someone is going to try to misuse those. And for now, when the models, the frontier models are all sort of in the hands of pretty responsible companies, I think we can mitigate those by the companies aligning the models and having good classifiers and good safety stocks. But we're not that far away from a world where there are incredibly capable open source models that are very good at biology. And the needs for society to be resilient to terrorist groups using these models to try to create novel pathogens is like, that's no longer a theoretical thing, or it's not going to be for much longer. So part of the reason that we are trying to push also in this blueprint ideas around societal resilience is a realization that AI safety or safety in a world of powerful AI cannot be done by the companies alone.

I think that's totally possible. Yes. I think to avoid that, it will require a tremendous amount of work also in a sort of resilience-style approach. Again, it's not just like make one AI model safe, it is defenders. We have this thing called a trusted access program. Other companies have other things. But, you know, cybersecurity companies, the major platforms, the governments using this technology to try to rapidly secure their systems, the open source stack, all of that. That's quite important.

I actually think in a different time, we used to say this a lot, the government should be doing this kind of work. And in a different time, I think it would have happened. If you look at some of the great expensive infrastructure projects of history or just scientific progress projects, things like the Apollo program, the Eisenhower Highway system, the Manhattan Project, these were government projects. And in a different time, I think the creation of AGI would have been a government project, too. I don't think in the current way the world works, that is likely to be successful. And so I think the biggest case against nationalization would be that we need the US to succeed at building superintelligence in a way that is aligned with the democratic values of the United States before somebody else does. And that probably wouldn't work as a government project. I think that's a sad thing. However, I do think absolutely, the companies developing this and the government have to work extremely closely together. I don't think this works as like a standard or a company and you're the government. There's going to have to be very, very deep partnership here. But that's not happening now. What we're trying and I think we're working very closely with the government.

Well, you mentioned a great one already, which is cybersecurity. I think biosecurity is another really important one. And then there are some other things here, like building our infrastructure, some of these economic ideas in the blueprint, that maybe doesn't have to be closed this year, but pretty quickly.

We're close enough to AGI that the precise definition matters. Some people would say we're already there. Some people would say we're not there yet. The fact that AI is discovering new, legitimately new scientific knowledge and the fact that AI is doing serious, valuable economic work at real scale, means wherever you think we are on that curve and where if you want to label AGI on that curve, we're quite far along. Like this is, we are in a new paradigm here of some sort. Faster than people expect, faster than people are prepared for. I would say already, major swaths of knowledge work have been transformed. The world has, you know, adopted that at different rates. But certainly what it meant to be a coder at the beginning of 2025 versus the beginning of 2026 were very different things.

First of all, not literally every. I think there will be many things, and I think this is wonderful, where we only care about other people, we will spend more time with other people. People will remain the most important part of our lives. You know, we'll have these incredibly smart machines doing stuff around us. And the most fundamental part of what it means to be human, that's not going to transform at all. And I think this is really important. Like the society may change in a lot of ways, what it means to be a person, what it means, some of the civilian life, how we choose to spend our time, what we really want. We'll have more flexibility and more ability to achieve. But, you know, for hundreds, thousands, tens of thousands of years, we've been driven by a lot of the same things. We want more of those. We want higher quality connections with people. We want more time to, you know, spend in nature, do what makes us happy. And I think we'll have that. The core of what it means to be a human and to have a fulfilling life, I think that's pretty deep. And that's not going to change. A lot of other things about the economy, the way society works, those may change a lot. But I think it's really when you get that point across, on the trust point, it's incredibly important that people building AI are high integrity, trustworthy people. And I think almost everybody involved in our industry feels the gravity of what we're doing. And so we all take that responsibility very seriously. We feel that weight every day. We also think it's very important that no one person is making the decisions by themselves that are going to impact all of us. I don't think we should have to trust a single person to get every decision right. We really believe in the democratization of AI. We really believe in putting this tool in the hands of people and letting people understand it, figure out what rules society collectively wants to put around it, and most importantly, how we're going to integrate it into our lives and have all these incredible benefits that we think are possible.