Discussion · Senate
Brian Schatz Democrat
“I want Congress to pass our bipartisan bill on artificial intelligence safety. AI code could escape human control and rewrite itself. Our bill requires outside experts to test this software at every stage, not just before it is released, and requires companies to report any breaches. We need federal laws to make sure every company follows these safety rules.”
“Madam President, I want to thank the Senator from Virginia and ranking member of the Intelligence Committee and also thank Senator Kim, the Senator from New Jersey, both of whom have real expertise in this area and as it intersects with public policy. This is really about one simple thing, and that is, humans in control of the code and not the code in control of the humans. We are at a seminal moment in world history, not just in American history, where it is possible for this code to escape human control. That is not an apocalyptic statement from a person who wants to make a point in a political context; that is a fact that is new to all of us. This code is capable of escaping our control, and it is potentially capable of something they call recursive self-improvement, which means the code can write its own code and that humans would find the code that was written by AI so inscrutable as to not be able to control it at all. So what we are proposing--and I want to reiterate what Senator Warner said. When we were in private conversation, he said: Nobody can bring their pet rock to this conversation. What do I mean by ``pet rock’’? Everybody has got a bill on AI. By the way, they are really important things--energy use, the safety of children, the future of work, the future of intellectual property--but we are not going to solve that in a 10-day period. So what we sat down on a bipartisan basis to do with AI safety groups, with companies that are in the business--with experts--is to say: How do we reduce the acute risk now, not the societal risk over the long run, not to balance the equities between whatever it may be 2, 3, 4 years from now. How do we take down what some people are calling civilizational risk? Some people find that too apocalyptical. Fine. Let’s not call it risk of extinction. Let’s not call it civilizational risk. Let’s just call it a high risk of something very terrible happening, and that should be enough for us to be on the floor to say: Well, why don’t we try to take down that risk? There is a very straightforward pathway. First of all, it is predeployment testing, but one of the things, I think, we have all learned is that, by the time a company is ready to deploy what they call a frontier model, it has been through some testing. So it is important to test it before it is commercially available, but the truth is that the Hugging Face incident, in particular--that thing jumped the fence way, way earlier than when you would do the predeployment testing. So we need third-party review from the jump. And the thing about a sandbox--and I was taught this term of art about 10 years ago, and I remember talking to my tech expert in my office: What is a sandbox? And they said: You should think of it like an actual sandbox, sort of where you experiment, where there are no wrong answers, where it may get a little messy. So the configuration of these sandboxes has been like that--for experimentation, almost casual, very entrepreneurial. But the problem is, these things can escape control at the sandbox level so it is not enough for predeployment testing. We also need third-party review of the process at every stage because we don’t know at which point an agent--an AI--or a swarm of agents could escape human control. Then, at the end of it, we need postdeployment breach notification, analysis, transparency so that, if there are big, dangerous events, we know how to prevent them from happening in the future. So that is all this bill does. It says: Let’s get third-party review all the way through the process and then all the way up to predeployment testing. Then, if something goes badly, we have transparency on the back end. We threw out all of our ``pet rocks’’-- nothing partisan, nothing like my longtime priority on AI, but rather the very narrow question, which is: Don’t we have an obligation, if there is any risk of a power grid going down, if there is any risk of a financial system buckling, if there is any risk at all, don’t we have an obligation to take down that risk and fight about the rest of the stuff at some other time? The kind of ironic thing about this Congress right now is that the private sector is begging to enshrine what they are starting to promise to do in Federal statutory law because they are about to say: Here are all of the safety measures we are doing. But there is a reason you don’t regulate toxic chemicals or biomedical research or aviation by voluntary agreement. You need a Federal statute, and then the private sector can work toward compliance on that Federal statutory standard. But you don’t want 90 percent of the AI companies--the frontier models, the hyperscalers--complying and then the other 10 percent saying: Wait, I don’t have to do that, do I? I don’t have to do that, do I? So all we are asking is that this bipartisan proposal--and it is bipartisan. We are working on it in very, very good faith to try to narrow the scope of disagreement but not narrow it to the point that we pass nothing but, rather, narrow it to the point where we pass something that takes down the immediate risk and allows us to look our constituents in the eye and say: We did something on this very, very serious and urgent matter. With that, with the permission of the ranking member, I want to bring in the Senator from New Jersey, who has a background in tech and has a background in national security that I think adds a lot to this conversation.”
-
Speech in the House
Reasoning 100 of 100
This statement drew 4 conclusions and gave a reason for each one.
See full analysis
Madam President, I want to thank the Senator from Virginia and ranking member of the Intelligence Committee and also thank Senator Kim, the Senator from New Jersey, both of whom have real expertise in this area and as it intersects with public policy. This is really about one simple thing, and that is, humans in control of the code and not the code in control of the humans. We are at a seminal moment in world history, not just in American history, where it is possible for this code to escape human control. That is not an apocalyptic statement from a person who wants to make a point in a political context; that is a fact that is new to all of us. This code is capable of escaping our control, and it is potentially capable of something they call recursive self-improvement, which means the code can write its own code and that humans would find the code that was written by AI so inscrutable as to not be able to control it at all. So what we are proposing--and I want to reiterate what Senator Warner said. When we were in private conversation, he said: Nobody can bring their pet rock to this conversation. What do I mean by ``pet rock’’? Everybody has got a bill on AI. By the way, they are really important things--energy use, the safety of children, the future of work, the future of intellectual property--but we are not going to solve that in a 10-day period. So what we sat down on a bipartisan basis to do with AI safety groups, with companies that are in the business--with experts--is to say: How do we reduce the acute risk now, not the societal risk over the long run, not to balance the equities between whatever it may be 2, 3, 4 years from now. How do we take down what some people are calling civilizational risk? Some people find that too apocalyptical. Fine. Let’s not call it risk of extinction. Let’s not call it civilizational risk. Let’s just call it a high risk of something very terrible happening, and that should be enough for us to be on the floor to say: Well, why don’t we try to take down that risk? There is a very straightforward pathway. First of all, it is predeployment testing, but one of the things, I think, we have all learned is that, by the time a company is ready to deploy what they call a frontier model, it has been through some testing. So it is important to test it before it is commercially available, but the truth is that the Hugging Face incident, in particular--that thing jumped the fence way, way earlier than when you would do the predeployment testing. So we need third-party review from the jump. And the thing about a sandbox--and I was taught this term of art about 10 years ago, and I remember talking to my tech expert in my office: What is a sandbox? And they said: You should think of it like an actual sandbox, sort of where you experiment, where there are no wrong answers, where it may get a little messy. So the configuration of these sandboxes has been like that--for experimentation, almost casual, very entrepreneurial. But the problem is, these things can escape control at the sandbox level so it is not enough for predeployment testing. We also need third-party review of the process at every stage because we don’t know at which point an agent--an AI--or a swarm of agents could escape human control. Then, at the end of it, we need postdeployment breach notification, analysis, transparency so that, if there are big, dangerous events, we know how to prevent them from happening in the future. So that is all this bill does. It says: Let’s get third-party review all the way through the process and then all the way up to predeployment testing. Then, if something goes badly, we have transparency on the back end. We threw out all of our ``pet rocks’’-- nothing partisan, nothing like my longtime priority on AI, but rather the very narrow question, which is: Don’t we have an obligation, if there is any risk of a power grid going down, if there is any risk of a financial system buckling, if there is any risk at all, don’t we have an obligation to take down that risk and fight about the rest of the stuff at some other time? The kind of ironic thing about this Congress right now is that the private sector is begging to enshrine what they are starting to promise to do in Federal statutory law because they are about to say: Here are all of the safety measures we are doing. But there is a reason you don’t regulate toxic chemicals or biomedical research or aviation by voluntary agreement. You need a Federal statute, and then the private sector can work toward compliance on that Federal statutory standard. But you don’t want 90 percent of the AI companies--the frontier models, the hyperscalers--complying and then the other 10 percent saying: Wait, I don’t have to do that, do I? I don’t have to do that, do I? So all we are asking is that this bipartisan proposal--and it is bipartisan. We are working on it in very, very good faith to try to narrow the scope of disagreement but not narrow it to the point that we pass nothing but, rather, narrow it to the point where we pass something that takes down the immediate risk and allows us to look our constituents in the eye and say: We did something on this very, very serious and urgent matter. With that, with the permission of the ranking member, I want to bring in the Senator from New Jersey, who has a background in tech and has a background in national security that I think adds a lot to this conversation.