Mind Readings: AI, Morals, and Safe Usage

In today's episode, let's talk about artificial intelligence and morality. Can AI be moral? The short answer is no. The long answer is no because we can't agree on what moral is. We can agree on some utilitarian aspects of morality, of ethical behavior.

We can agree that generally speaking, uh, a machine should probably not do harm. Um we can generally agree that machines probably should not say outrageously wrong things, like the earth is flat or gasoline is safe to drink. Um beyond that, you start running into issues of whose morals you want to use to guide the machines. And this is where this is where a lot of models will need fine-tuning and perhaps different versions to match the different cultures that they operate in. You know, think about uh a very simple product like alcohol, like wine, right?

There are some places in and some cultures where it is not permitted, it's outright not permitted. There are other places where that's something people indulge in a whole lot. Think about something like marijuana. There are places where it's legal, places where it's not legal. How do you teach a machine that supposedly has universal reach to localize for each of these things?

The answer is you really can't. Not well. You can absolutely get at them to, again, to avoid causing basic issues and basic problems, but above and beyond that, you really can. So think about AI kind of like kind of like a chainsaw, right? A chainsaw has no morals.

It's a tool. And that tool, and the results we get is contingent on the user. You can, and chainsaw manufacturers do, include lots of safety instructions and some basic safeguards, right? Like there's a hand guard so that your hand doesn't get hit by shrapnel as you're sawing through something. There's a circuit breaker so that the motor doesn't overload if you're using an electric chainsaw.

But beyond that, there's not much else you can do to make something that is inherently very powerful and sometimes very dangerous safer, right? It's like a firearm. Take your average firearm. A firearm is designed to shoot things, right? There's there aren't a whole lot of applications for it that don't involve it shooting something.

So you can have safeties, you can have indicators in the barrel to show you if around is chambered. You can uh regulate to some degree the kinds of ammunition that you can use. But at the end of the day, it's a it's a thing that's designed to shoot things. Um there's really no way to make that safer. Um with AI, particularly generative AI, there are things you can do to tune the models.

There are things you can do to avoid putting garbage data in in the first place, and that that is where a lot of discussion is going to be happening in the very near future about whose data is in these things, whether the manufacturer of these models had permission to use that data, and whether or not that people can retrieve their data from it or remove it. So there is definitely moral and ethical things you can do on the creation of the models themselves, but not on the use of them once they're deployed. So the short answer to this very challenging quandary is twofold. One, if you want an AI model to have a specific point of view, you will have to train it or fine-tune it to have that point of view. The stock off-the-shelf models probably will not have those.

And if you want it to embrace a specific viewpoint in the world, you're going to have to fine-tune it. And then that's what you use rather than something that's off the shelf for that the general public would be using. And two, um, we should be asking for regulation, not of the models themselves, but of what goes in them. To say, like, yeah, if you are if a model is considered a derivative work of copyrighted work, that's going to change the game for a lot of generative AI. That's going to dramatically change what the machines are capable of.

And it will make permitted data collection a lot more valuable, right? Individual authors could give licenses to machine learning companies to say, yes, you may use my works in exchange for this fee, um, as an example. So that's where we're kind of are with AI and morals these days. Uh ethical AI. There isn't, because that's just not how these tools work.

There absolutely can be safer tools that have obviously morally incorrect things removed, again, like recommending someone go drink gasoline. Um, but after that a certain point, it's going to be up to the individual to take responsibility for what they want the machines to do rather than demanding that manufacturers create custom models for every single person in the world or even just every single group of the world, you'd be still be doing a lot of uh processing that is impractical and inaffordable unless there's very deep-pocketed folks funding those things. So that's the show for today. Thanks for tuning in. We'll talk to you soon.

If you like this video, go ahead and hit that subscribe button.