哈佛教授分享三类工具：让作品对人可读、对 AI 不可读并污染训练数据

以下是视频口播的自动转写，可能有错字。

I am a harvard-trained professor who studies the impact of ai on society. And recently, i've become incredibly interested in how i can break ai systems just by publishing false data on the internet.

It's actually quite fun and quite easy. Would you like to learn? Yes. All right, we have three different categories of tools here.

The first is to hide your browser data. The second is to hide your artistic content and the third is to hide the thoughts that you write down and publish on the internet.

Let's start with text. If you're publishing your ideas on the internet and you don't want ai to be able to steal your intellectual property, you're going to want to use a text called shield font.

It's a font that looks perfectly normal to humans, but it's completely illegible for ai shield font works by using something called ligatures, which is a technical term for when you take two letters in a font and you combine them into one glyph.

But instead of doing this with letters, shield font does it with entire words. Thankfully, shield font also has screen reader support, which means that the font is still accessible for humans who are visually impaired.

And the more people who use fonts like shield font, the worse the llms get over time, because they end up just gobbling up a bunch of mumbo jumbo that doesn't make any sense to that.

But what if you're not a writer, and instead you take photographs or images? Then you're going to want to use something called nightshade.

It does think pretty similar shield font where it leaves your original image completely readable to human eyes, but then makes it illegible to robots.

And if you're worried about ai-powered editing or manipulation of your images, you're going to want to use something called photo guard, which makes that much more difficult.

And finally, if you want to poison your web browsing data, so that machine learning models aren't able to deduce your spending patterns or your psychological state.

Do you need some money? You want to use tools like track me not and adnose m. They basically work by injecting a lot of misleading browsing or search signals that confuse a machine learning model about what the user is truly doing.

so save this video and send it to a friend of yours that's an artist or a publisher, and make sure you stick around for the next video.

keep wondering why.