You Ask, I Answer: How to Regulate Generative AI?

In today's episode, Carolina asks, what in your opinion should be done for generative AI not to go too far? How can we take benefits but within control? It depends on what you mean by control. If you mean the ability to uh turn these things off, to stop using them, we're past that point. If you mean the ability to regulate them uh through a central authority like uh a company like OpenAI, we're past that point too, because uh there's now a wave of freely downloadable open source models that are very high quality.

I talked about this in a recent episode of my newsletter, um based on Facebook's Llama or Stable LM, open source models that are in the wild that you can download and put on your computer today, and no one can control that system, right? No one can control you having that software or not. So that's largely moot. The reality is that large language models are here to stay, and that the technology itself really can't be regulated because it's now so open. You can download one of the base models and then fine-tune it, train it to do whatever you want it to do.

You could train it to only do clam chowder recipes, right? You could train it to spew hate. You could train it to create propaganda and misinformation. And because these models all are small enough and today's personal computers, your gaming laptop is powerful enough to do that fine-tuning, there really is no way to regulate that, right? Any more than you can regulate how someone's going to use a chainsaw, right?

Yeah, you can put safety warnings all over it and stuff, but if somebody wants to go all Friday the 13th on somebody else with a chainsaw, there's not really anything a chainsaw maker can do to stop somebody from doing that, right? So what do we do? What we do is we look at the outcomes and we regulate the outcomes. Um, for example, in the USA, which is where I'm based, we have laws that add uh essentially more penalties onto a crime if that crime is done within a certain context. For example, we we have a category called hate crimes, where if you commit a crime, there's a sort of base level of of punishment for that.

And then if it can be proven in a court of law that it was a hate crime, that you did that crime because of the person's race or sexual orientation or veteran status or disability, you get extra punishment, right? You get extra punishment on top of the punishment you've already gotten. And so having laws that would essentially restrict what people do with these models would be the way to go. And we're not talking about saying you can't write certain things, whatever. We're talking about uh essentially things that are already against the law, just in a enforcing those laws, which is a whole separate conversation, and B, maybe adding an extra bonus penalty for if you use machines to do it perhaps at scale.

So for example, fraud is illegal. Uh scamming somebody out of money, illegal. If you used, say a machine to synthesize someone's voice to create a fake ransom call, that's still illegal. It's just more illegal. And so you could add a penalty saying if you've you if you misuse technology, in addition to the twenty-five years of jail time you're gonna get for fraud for in this case, uh I know I'm just making up these numbers.

Um you would then say, and we're going to add an extra 10 uh on 10 years of penalty on to this because you'd you used AI to do it, right? So it's kind of like well, I mean, there are there are many different laws that have multipliers or variables that change the severity of the punishment. If we want AI to succeed, if we want AI to be useful, uh, and we want people to not abuse it, we have to A, enforce the laws we already have, which is always a uh always a treat, and and and do so in a um in a coherent, consistent way, meaning that some people don't get a pass because of their background or who they know or how much they bribe the judge. Um, and B considering multipliers on existing laws to say, like, yeah, if you use AI to do this, the crime is worse, right? We consider the crime to be worse, therefore the the punishment is worse.

That's what we can do, because we cannot control the mechanisms themselves. Any more than you could control spreadsheets, right? If you think about it, if you use a spreadsheet to commit a crime, you can't just turn off spreadsheets. It's impossible, right? You there's no way for you to stop people from using spreadsheets.

There's open source ones, there's Google Sheets, there's Microsoft Excel, and yeah, Microsoft could maybe turn off your specific copy of Excel, uh, if they had the licensing information, but it just downloaded download open office from or LibreOffice online, free. Um, runs on your computer, very capable, and now the person's got a spreadsheet. And if you didn't want them to have a spreadsheet, you're kinda out of luck. But you can say, yeah, if you use a spreadsheet to commit this crime, we're going to add an extra five years of penalty, you know, or whatever the however the legal system works in that country. That's essentially where we are today with large language models, with generative AI uh in general, is saying, yeah, the tools are out there.

Now we gotta regulate how people use them and and make it clear there are criminal penalties for misusing them. Not the general misuse of them, but if you're committing a crime, if you just use AI for it, we're just gonna make the penalty worse. So that's it. Um it's a heavy topic to talk about, and it's one that I feel like a lot of governments, a lot of legislators, a lot of elective officials do not understand. And they will propose legislation that is impossible to enforce.

And so, like many other things they've tried to legislate and and very difficult to enforce, regulation of the technology itself probably is not going to be super successful. Anyway, that's uh the answer, or at least that's my answer to the the question. Thanks for asking, and we'll talk to you soon. If you like this video, go ahead and hit that subscribe button.