Why is this important?

- All right, so let's get started with AI ethics. So why is this important? I believe that everyone has a right to know what is going on behind the scenes because these models can perpetuate harm when users are not aware of how they work. We are going to touch on many different aspects of AI ethics through many different mini case studies throughout this talk, and there are definitely some topics that we could go more in depth on, but this talk is sort of meant to introduce you to many different issues and open questions in this space. So let's talk about what happens when people don't understand the models they're using. The first example we have here is there was a lawyer last year in 2023 I believe it was, that used ChatGPT to help him write a brief that he was going to present in the courtroom for a case that he was trying. And he basically asked ChatGPT to help him write this entire document, right? And what happened was ChatGPT wrote the document and cited a bunch of cases helpfully in the document, but all of the cases that it cited were completely made up. And so the lawyer obviously got caught and faced, you know, fines and the, like, the law firm that he worked for faced sanctions and all of this. But really it was, it was kind of mostly just embarrassing for him, right? He said he did not comprehend the chatbot could lead him astray. He thought it worked more like a search engine and just like a really, you know, useful high powered search engine that would be able to write up the brief too after doing all the research for him. But that's not how AI models work. He didn't understand that, right? Another example is a bunch of teachers, one concrete example of this is one professor in Texas flunked all of his students after he ran his papers through ChatGPT asking ChatGPT if it had written the students' papers to check if they had used GPT to write their papers, and ChatGPT claimed it had written all of them. ChatGPT is very greedy. It will claim that it has written a lot of things. I don't know if this is still true, but you used to be able to pass the Constitution and then ask if it had written it and it will say, "Yes, I wrote this," right? And so this teacher, this professor thought that if ChatGPT wrote something, it would be able to definitively without any mistakes, completely accurately tell him if it had written something or not. And so he flunked every single student in the class and this was really a big deal for some of them. They were about to graduate and this messed up their graduation plans and it was all reversed once someone explained to the professor how ChatGPT worked. And he, you know, reversed the grades and everything. But another example, and there's many examples of teachers doing this, not completely understanding how this works and accusing students of using ChatGPT when they didn't. In programming too, this can be an issue. So there's a lot of pretty good AI code helpers that will help you generate code. And I use them in my work, they're actually very, very helpful. The CEO of Google says that more than 25% of the code at Google is now AI generated, which is a lot of code. But there's also been studies showing that people who use AI code assistance are less likely to write secure bug-free code. AI assistants are not quite there yet in terms of writing code that is foolproof, but they are good at writing code that looks very good. And so it's easy to have AI write a bunch of code that would've taken you an hour to write, do it in a couple seconds and you glance at it and you're like, great, that looks good. And I'm definitely guilty of that as well. But because it's not foolproof, you're more likely to have bugs that you don't, that you would've caught if you had written the code yourself, have security issues and things like that. I want to take a moment to step aside for a second and talk about how machine learning AI works in general. For those of you who are maybe watching this talk who have a vague sense or no sense at all of how this works. And if you already know how AI works in general, you can feel free to skip over this. But let's talk about how machine learning or AI models work. So much like a baby learning how things work by example, by trial and error and correction, this is how models learn as well. So let's say we are trying to create a model that is able to tell if there's a cat or a dog in an image. So like whether an image is of a cat or whether it's an image of a dog, right? How this works is that we have a bunch of training data, a bunch of images that we input into the model and then we ask it to give an output, a guess, a cat/dog prediction, right? So maybe at first the model will see its first image, image of a cat. And because it has no knowledge of anything about cats or dogs, we don't give it any other information than these images, it might guess wrong, right? It's just going to guess willy-nilly. So it might say dog, then it's told it's wrong. It's given this correction, right? Because we have these images a key part of the training data that we're feeding into the model, these images we're feeding in are labeled. We know, humans know and have, you know, somewhere in the dataset that this is labeled as a cat and that other images are labeled as a dog. So we can very easily say, "Nope, that's wrong, this is a cat." This is all done through code, right? It guesses dog. We say, "Nope, that's wrong." And then it will update itself so that next time it sees a cat, hopefully it'll get it right. And this takes a while, right? Like it takes a while to get to be really accurate. Once it sees a bunch of images, a bunch of different cats, a bunch of different dogs, it's corrected when it's wrong, it's told it's right when it's right. It slowly builds up a model of what makes a cat and what makes a dog. It learns to recognize the patterns of what images of cats look like versus what images of dogs look like. And again, this is much like how we learn when we're young and seeing the world for the first time. Like at no point does a human say to us when we're a baby, this is what a cat looks like. Whenever you see something with pointy ears like that, that's a cat, right? That's not really how we learn. Again, that's not how models learn either. There's no rules in there saying whenever you see this tail, whenever you see these pointy ears, guess cat, right? So when we are young, we might point to something and someone will say, yeah, that's a cat. Then maybe we'll point to another fuzzy animal and say, cat, and someone will say, no, that's a bunny, right? And so in the same way, we learn through lots of input data, lots of trial and error and correction, and same way the models do it. They get better and better. Their accuracy improves as they learn more about the patterns of what makes a cat and what makes a dog, right? Again, it starts completely random, then hopefully reaches a point where the accuracy is good enough to deploy the model into the world for whatever we need to discern between cats and dogs for. That is how machine learning works generally, general overview. We can, you know, head back into the main portion of the talk now with that behind us, because that'll be helpful to know when we talk about other things in AI ethics and safety. So when we create large models like large language models like ChatGPT or image generation models like MidJourney or any of those, any of those image creation models based on prompts or anything like that, they're trained on immense amounts of data, lots and lots of data, and they're the product of the data they are trained on. So models are trained on all of the data available on the internet, like all the textual data available on the internet for language models, all the image data available on the internet for computer vision models. And that means that they are subject to seeing and learning from the best of society, the worst of society, the brilliant, the untruthful, the curated, and the raw. Anything the internet has to offer, these models will soak up. And since they've learned from that data, they're a product of that data, right? So they learn from some of this really, really great data and some of this kind of not so great data, some like untruthful or hateful data, right? So this is all really, really important to keep in mind when we talk about AI ethics and safety. A lot of the issues people talk about in AI ethics kind of all come back to this one fact that models are the product of the data they are trained on.