终结危险AI的真正解法：打破技术特权，回归通用软件法律追责

You saw the title, yes, there's an easy way to prevent dangerous AIs that very few people ever talk about, and certainly no one in the AI industry, even many of AI's biggest critics don't talk about it, more on that later. What we would have to do is this: treat AI software just like any other kind of software. That's it, this is the way that's all it would take, and we should because they really are just another kind of software. But there are a lot of different things that would have to change in order for that to happen, and there are a lot of objections the industry is going to have to have to that happening. And talking through those is what the rest of this video is for. And the people in the AI industry do not like this idea at all, but after all the crap they put the rest of us through the last few years, as far as I'm concerned, they can get F-----.

This is the Internet of Bugs, so my name is Carl, I've been a software professional since the late 1980s, and I'm trying to do my part to make the Internet a safer, less buggy place. You can find more of my stuff on my Patreon, my Substack, and InternetOfBugs.com - links down below.

So let's start with the most recent "AIs went rogue and hacked Hugging Face" headlines. So if you were to write a piece of non-AI security software that could hack websites, a computer virus, a malware, ransomware, call it whatever you want, and you let it run on an Internet attached system for a couple of months without paying any attention to it and it hacked some websites out there, what would happen to you? Well, they put out a warrant for your arrest, and if they could find you, and if you weren't hiding in some country where they couldn't reach you, they'd arrest you and put you on trial. There are lots of examples of that happening. But the last few weeks, we've had some counter examples where the arrest warrants for the people to OpenAI and Anthropic. They built systems that have hacked other companies, but they're not in legal trouble. And that's because we don't treat AI software like we treat any other kind of software. It's also because the most visible target of the recent AI hack, a company called Hugging Face, appears not to have wanted to press charges, put a pin in that we'll come back to that later.

So why don't we treat AI software like other kinds of software? Well, for a few reasons. But the most important are one, that the public has been convinced that AI is some new kind of thing that really isn't software, and two, AI spending is literally propping up the U.S. economy right now, and without it, economic growth would be very negative. Now all I'm going to say about the "propping up the economy" thing in this video is that the amount of money being invested in this thing is orders of magnitude more than it's worth, and it's only a matter of time before people realize that and the bubble pops. Treating AI like normal software would make the bubble pop faster, but as painful as that would be, the longer this goes on, the bigger the bubble gets, and the more economic damage it's going to do when the inevitable crash happens. Fundamentally though, I'm a tech guy, not a finance guy, and there are people out there that cover the financial side of this a lot better than I can, so go check out Ed Zitron or Nobody Special Finance if you want to know more about that. Links to that down below too.

Which leads us with just point one. The argument that AI software isn't just software, it's something else entirely, and the rules for normal software shouldn't apply to it. They have a bunch of different ways they try to make this argument, some are more explicit than others. The first thing they do is by claiming that these AIs are as smart as humans, especially "human PhDs in every subject", they love that one, or that the AIs have reached AGI or artificial general intelligence, their name for human level intelligence, or something to that effect. This is nonsense. I'm not going to go into that right now because I have a number of videos and posts in which I'll explain why this is nonsense and I'll put links to them below.

The next way they want to convince you that AI software isn't just software, is by convincing you that AI is going to kill us all. There's been a lot of that recently, and again, I'll put a list of previous times I've argued about this nonsense in the past, but as I'm arguing in this piece, if we treated AI like it was any other kind of software, it would no longer be dangerous, which makes this line of reasoning circular. AI isn't normal software, they say, because it can kill us all, but it can only do that if we accept the premise that it isn't normal software and then we don't treat it like it's normal software, which means it can't kill us all.

Anyway, yeah, there's a related argument here that says that we need to regulate AI software in a special AI-specific way by doing things like managing government review of AI or having government slow down AI progress or argue that some AI companies need to be prevented from moving so fast. This of course, we're supposed to need to do because, as I said before, they want us to believe that AI is so special and different that it will kill us all, and their arguments about AI killing us all and needing to regulate it in an AI-specific way provide the AI industry another benefit, it allows them to control the narrative. It keeps the public so busy talking about AI-specific regulation and AI-specific danger and human extinction that it keeps us from having a conversation about why AI software shouldn't be treated like any other software.

So let's have a quick conversation about that. Imagine for a moment that we were to treat AI software like it was normal software. What would happen? Well, to start with, our reaction to OpenAI and Anthropic have each had software they wrote and get out of their control, steal information from and damage the computers of other companies wouldn't be, "Oh, we should have a conversation about slowing down development." It would be, "When is the FBI investigation going to determine who is criminally liable? When is their trial going to be? What internal documents proving company negligence are going to be released to the public? And how much jail time are those responsible going to serve?"

And the AI industry really, really, really does not want that conversation or that investigation or that trial to happen. So much so that they would rather have us all believe that they are working on a system that they all believe has a good chance of killing them and their families and their friends and everyone they've ever known. So what do you think is more likely? That they're actively trying to build something that they truly believe will destroy the world? Or that they're only saying that because they don't want the same rules to apply to them that would be applied to any of us, where we do the same thing?

At this point, I've spent more than 3 and a half decades writing software for a living and trust me when I say that the way that programmers behave when the code they are writing exposes them to criminal legal liability is very different than the way programmers behave when there are no consequences to being reckless. And in my professional opinion, none of the AI companies are exhibiting any evidence of worrying about any kind of liability. But you know what, we could change that. And if we did, the safety problem would fix itself because the AI researchers would not be doing what they're doing if they thought there was a chance it could send them to jail.

The business people at the AI companies currently doing the circular financing stuff, they'd probably still be doing what they're doing. They already know it could send them to jail, although based on the fact that no one was held responsible for all the financing irregularities that caused a 2008 financial crisis. They probably think they're safe and unfortunately they're probably correct. But the crimes coming from the software at OpenAI and Anthropic and SpaceX are very different kinds of crimes.

For one example, lots of people have been sent to jail for hacking. In fact, between 2014 and 2021, 2,590 individuals were sentenced in the US for some form of hacking, just in the US. But that small potatoes, hundreds of people in the US every year are sentenced for child... Okay, so I have to say this indirectly so the censors don't flag this content. Let's say "for images with children with too few clothes on" and the average sentences for such offenders is 275 months in prison. There are whole databases and websites dedicated to tracking court cases where AI was used to generate images of an intimately abusive nature. And there have been a number of news reports about abusive unclothed images generated by AI from SpaceX and OpenAI - links to all of this down below. And if we treated AI software like normal software, those news reports would have been followed by news reports of FBI agents raiding the offices of OpenAI and SpaceX to seize the images used to train OpenAI's DALL-E and SpaceX's Grok to see if there are illegal images in there of underage children in illegal states of undress. And chances are if those AI's are as good at generating the illegal imagery as the news reports imply that they are, there's likely training data depicting those kinds of images, which is a big crime. And there are specific reports that OpenAI has such images on file as witnessed by Kenyan workers who are paid less than $2 an hour to view and classify such images. Again, links below. If you or I did that, we would be in big trouble. But the AI companies have convinced society that the same rules should not apply to them.

Now the AI companies and the AI employees are going to raise objections if we actually try to hold them accountable to the rules the rest of us have to live by. First off, they're going to say that they're already doing all that they can do to make the system safe. That's a lie. We'll talk more about that in a little while. Then they're going to say that they have to keep going or someone even less scrupulous than they are is going to make it to super intelligent AI first and it will kill us all. Again, the killing us all thing is nonsense. At least anytime soon. There may come some point in the distant future when real artificial general intelligence is a possibility, but there are a number of serious deficiencies in the current state of the art technology that the AI industry will have to overcome before then. Again, see the list in the comments about the previous work I've done explaining why we're not anywhere close to AGI and what breakthroughs the industry would have to do in order to get to AGI.

They'll probably also try to make some arguments about the need to use their models to defend against the open source models - you know, the ones that they don't make money off of. I could make a whole video on the implications of open source models, and I might one day, but for now, here are two things to consider: item one, if the open source models are good enough to be used on offense by the hackers or the attackers, then they're also good enough to be used on defense so we can use them and we don't need the frontier models from companies like OpenAI and Anthropic and SpaceX to help. But most importantly, there's item two: The frontier model AI companies like OpenAI and Anthropic constantly complain that all the people that release the open source models do is steal their work and distill it into the open source models. If we were to take them at their word, then the only way the open source models would get any more dangerous would be if the frontier models got more dangerous first and the open source models distilled what made them dangerous to get more dangerous themselves. Therefore, if the leading AI companies stopped training dangerous models, the open source models wouldn't have anything dangerous to steal, meaning we wouldn't need to defend against them. QED.

Now, let's talk about the "we're making things as safe as we can" lie. This is the one that really drives me crazy. The AI companies say that they're trying to make AI's that are aligned with humans and that they're trying to put guardrails in place to prevent the AI's from doing bad things. These so-called guardrails, they say, were turned off during the Hugging Face hack, which is one reason why it can do what it did. Here's the problem. So if you want a toddler to be safe, you don't train the toddler how to stick a fork in an electrical socket and then generate guardrails to try to make it harder for the toddler to do so. You keep all that stuff away from the toddler in the first place. They could do that with AI, they just won't. All of the filters and guardrails that they have put between the trained AI and the public could also be applied to the data that's about to be used to train the AI before the training starts.

Here's one example: Here's a new report of an engineer at MIT who got past OpenAI's guardrails and tricked ChatGPT into explaining how to make biological weapons. So here's a question: Why the HELL would you train a chatbot how to make biological weapons in the first place? If you keep that knowledge out of the training data, then you don't have to worry about the guardrails because there's no information there that the chatbot can be tricked into divulging. And they've already done work to try to filter that information out. It's just that their filters are easily tricked. But if they applied those filters before the training, there wouldn't be anybody at that point that would be trying to trick them. So they should work just fine at preventing that information from getting into the model in the first place.

But the AI companies won't do that. They want to use all the training data that they can possibly get their hands on. Why? Well, there are three reasons. First, because filtering that information out of the training data will cost them money and slow them down. Boo hoo. Second, because they have a misguided belief that if they put enough training data into a model, that it will transcend its training and become super intelligent. There was a point three or four years ago when that seemed maybe that idea might be plausible. We now know that it's not going to work, link below. And then there's the third reason, the real reason: they want the AIs to know how to make biological weapons so that they can make money renting the AI to people trying to defend against biological weapons. It's a great racket. They create the problem so then they can make money from the solution to the problem they created.

This is why frontier models are so good at hacking - because they're trained with enormous amounts of past hacks and past vulnerabilities and code patches and proof of concept exploits and hacker capture-the-flag games. Why would you train a model and all that stuff? Well, so you can get money from the people threatened by the AI that you created. And if your model happens to get loose and start hacking the Internet, well, that's great advertising for why people need to pay you to defend their networks from the AIs just like theirs. And when the inevitable public event happens, then just turn into an opportunity to try to convince everyone that AI is the only way that they can defend their own network. And as long as they aren't held responsible for any of the hacking their software does, they're going to make a ton of money on that.

And the government certainly shows no signs of holding them responsible. Politicians as far right as Josh Hawley and as far left as Bernie Sanders have bought the "AI goes rogue" story hook line and sinker. And the leading idea for legislation at the moment was written by the ControlAI organization. This is the group that paid Hank Green and SciShow to lie about AI development being faster than nuclear power. There's a video on it. The bill they wrote is all about creating government oversight to prevent super intelligence, which we're nowhere near. And it does nothing to help the underage girls whose Instagram posts are being turned into naked pictures by AI, the depressed teenagers being discouraged from seeking human support before making life-ending decisions, or anyone else being currently harmed by AI or any companies that are being hacked by AI.

There is one thing that could rain on the "AI gone rogue" parade though, if, hypothetically, and I told you to put a pin in this back at the beginning of the video, if a company who had machines compromised and proprietary information stolen -like say Hugging Face - by an AI company that admitted to the public that their AI was responsible. And if a company like that, like Hugging Face, were to press charges against OpenAI or to file a public lawsuit against OpenAI for damages related to the hack, that might just set a court precedent that the AI company should be held responsible for what their software does. And if THAT were to happen, that might spook investors and pop the investment bubble, which would be a shame for some company like, say, oh, Nvidia, who is raking in the money from the AI bubble right now. But I'm sure it's just a coincidence that the company who stood to lose the most from a court precedent being set just happened to buy the company best positioned to force that court case to happen. Eh, move along, folks, nothing to see here.

And while we're speaking of "move along, folks, nothing to see here", there's one more aspect of this that I find astonishing, which is that almost none of the most vocal AI critics are calling for straightforward legal liability. Instead, even the critics are talking about how we need to figure out new ways to get AI to be regulated and to have the government step in and slow things down or some crap. NOW. Less than two months before the midterm elections. As if anything besides blow-harding jawboning and posturing was going to get done anytime soon. But we don't need that. All it would take is some prosecutor out there willing to do their job, or enough of us throwing a fit to push law enforcement into looking into corporate crime for once. Wouldn't that be nice?

Now, I do have one caveat. There's one problem that treating AI software like regular software can't solve: AI can't be prevented from making mistakes. So, for example, there was recently a case where AI gave some hikers incorrect information that almost got them killed. Treating AI software like regular software won't fix that - because regular software already kills people by giving them bad directions. It's a phenomenon called "Death by GPS" - happens all the time, lots of links below. So, to be clear, treating AI software like every other kind of software doesn't make it magically perfect. It just makes it no more worse than the existing non-AI software already is, which, unfortunately, is the weakest link in this scenario because historically, we've definitely not done enough to regulate software written by large corporations. And more importantly, on those rare occasions the government has tried to enforce anything against large corporations, It's just a "cost of doing business" "slap on the wrist"-sized fines with "no admission of wrongdoing."

In my defense, though, if we never solve the problem of enforcing existing laws against large corporations, then there's nothing whatsoever that can be done to make AI, or, in fact, anything done by any large corporation ever, any safer at all. At least not WITHIN the law, but I'm not saying anything more about that. But we do need to do more about holding companies - not just AI companies, but all tech companies - responsible for the software they write. That's the core reason that I made this channel, why I'm making these videos, because the quality of software on the Internet was already going downhill - and fast - before ChatGPT - and especially vibe coding - became a thing, and it's getting worse even faster now. And the Internet already had far too many bugs before AI came along. And treating AI companies like the rules of software liability don't apply to them is making everything much, much, much worse. And anyone who tells you differently is probably trying to sell you their AI so you can defend yourself from... well... their AI.

Thanks for watching. Let's be careful out there.