AI安全开关可行性与监管路径探讨

AI anxiety is nothing new. Open the pod bay doors, HAL. I'm sorry, Dave, I'm afraid I can't do that. But now that the technology is here, AI dread has hit a peak. First, a tweet from a former OpenAI and Anthropic researcher went viral. The people building AI earnestly believe that it could kill us all by the end of the decade. Then the top AI executives called for a slowdown. A real fear that some of these companies, or some companies developing AI, could get too much power and be able to sort of exert undue influence on the economy, push a worldview out on people.

And I think the world is right to be afraid of this. The danger of AI is very significant at this point. When a lot of people from Anthropic and from OpenAI are telling you that their models are very dangerous, I think we should believe them. This has led a variety of experts to call for an AI kill switch, or a safety protocol that could stop AI from doing something catastrophic. When machines go off the rails, humans need to be able to turn them off.

I think it absolutely makes sense for us to think about having an AI kill switch in place. As we saw with with the OpenAI Hugging Face incident, we are shockingly close to the 2001 A Space Odyssey scene where HAL says, I'm sorry, Dave, I'm afraid I can't do that. I know that you and Frank were planning to disconnect me, and I'm afraid that's something I cannot allow to happen. The systems are trained, they're built to imitate human behavior.

And that's where this behavior comes from. But is an AI kill switch even possible? And are we too late? We spoke to a variety of experts to find out. So historically, a kill switch is a big switch on the wall of the factory that shuts off the power and then it shuts off every machine in the factory. And the question is, can we create that for AI? And the answer is, of course, it's more complicated. If a hyperscaler, Anthropic or OpenAI, had just one huge computer or one data center with tens of thousands of computers, then cutting the power for that is a real possibility, right?

The simple reason that we can't have a single kill switch for everything is that it's all distributed. So there are tens of thousands of machines, thousands of data centers. A possibility is to have standards and coordinated standards, right? Because turning off a system in one country leaves it running in another. So we really need international coordination for this. But standard, ideally hardware based kill switch that is, every chip could be delivered with an operational code that must be provided every minute.

And if it doesn't get its operational code, it's simply ceases to operate. So that's an idea for a hardware kill switch that requires the chips be designed in a certain way. We've seen proposals in Congress for potential kill switch bills for AI in particular. Is that something that is even possible? I think directionally, it's the right thing to do when you're talking about having some type of controls over AI. But I'm not sure it's going to be a panacea to solve all the the myriad problems that AI is presenting.

You're now talking about having to have thousands or hundreds of thousands of software manufacturers build backdoor kill switches into their software. And I think logistically, that's probably hard to impossible to do. I think a question that it also raises with me is who controls the kill switch? Is it a human? Is it an AI itself? Who regulates the idea of putting in controls is an interesting thing, but probably isn't something that realistically is going to work.

There is no single kill switch. I think what we actually need is regulation around data privacy, resource extraction, labor protections, child safety, all sorts of other areas that AI touches and that this is not special or exceptional as an industry. Anthropic co-founder Jack Clark told the BBC this week that the industry is working on a kill switch of sorts. Most labs have different ways of being able to pull the plug.

And yes, of course we have systems in place for this, but this is the kind of thing that you want to feed into the larger policy conversation. Should you mandate that companies definitely have a kill switch? And is that kill switch verifiable by a third party? I think that's the kind of thing that society is going to want to know and might want to eventually pass rules around. Lawmakers from both sides of the aisle have introduced bills that would regulate a kill switch.

It basically requires the large AI companies to have the capability to turn off their own AI advanced models. It creates a process with appeal rights for AI companies when the government orders a slowdown or shutdown of these models, and we give the primary jurisdiction to the Department of Homeland Security. I'm very skeptical of the framing of a kill switch, and so I'd be skeptical of legislation that filled out that framing.

I'm much more interested in legislation that looks a lot more like how we regulate other industries that create harmful products. Because I think the kill switch framing leaves a lot of ambiguity that tech companies can exploit to have this work in their favor. The EU AI Act has some real teeth to it, and you can see these tech CEOs fighting back quite a lot. We're off to a good start with the EU AI Act, for example.

There's regulation as part of our transport and food and drugs, but the one exception is software. Software companies disclaim liability for any harms that might come from the use of their software. Whether or not we have a kill switch, if we make the companies that create and sell these services liable for the harms that may come from their use, then I think that would be a game changer. The challenge with regulation is that by the time they are formulated and proposed, the technology has moved much farther and it becomes much harder for you to future proof every aspect of AI systems that may come into existence in the future.

It is really the agents that I'm more worried about, and I think we need to stay more focused on. Of course, the models themselves need to have guardrails, and there are various efforts to make sure that they do not let you maybe create a bioweapon or some such. Right. But at the same time, we should really focus on the agentic automation, because these are the ones that have access to the tools. And using the tools is how they start performing bad stuff.

Are we all going to die in 2030? Maybe it won't matter. We are not, we're not going to die in 2030. Safety is an engineering problem. Testing is an engineering problem. We should create products and properly test them. And if they're not ready to be released, just hold on to it and keep testing it and keep engineering until it's ready. I think what we need to avoid is mistakes that have happened in the past with like the tobacco industry or oil and gas, where companies have been able to either self-regulate or sort of flood the media with misinformation or fund their own ethics research.

This is happening in tech, too. And so I think it's important that we not let tech companies regulate themselves because it simply doesn't work. Although President Trump does not see the need for guardrails. Robots will not be taking over. The AI will not be taken over the rest of the world. The whole thing is a hoax. Elon Musk suggested the industry work together on a solution while he spoke at the All In Summit this week.

I mean, what I think would be wise to do as soon as possible, if not immediately, would be to have the major AI competitors test each other's models. But the thing that we do most immediately, and probably get agreement with China, would be peer review, where the leading AI companies all test each other's models before release. There are a lot of ways that we could be restricting and limiting model development and imposing more rigorous testing frameworks before things are released, and I think they should be audited by external audit

原文超出正文长度上限，此处截断——上游还有内容，完整版见上方「原文 ↗」。