[0:00]Catch this video, ad-free and uncensored, only on Nebula. Is it going to destroy the world? That's the question on a lot of people's lips. AI is coming, they say. It's inevitable, they say. We'd better teach it to be good and hope it listens. Can we make ethical AI? The future.
[0:43]The AI people are most worried about is artificial general intelligence or AGI, a digital consciousness that surpasses humanity and might decide to do away with us. A lot of money, effort, and anxiety is right now being poured into making sure AGI aligns with human values. But I'd like to put the alignment problem to one side for the moment. We will come back to it later. AGI does not currently exist. What we do have are a lot of very different tools that are all called AI. We have algorithms that judge people on their peroral hearings, job applications, and credit scores. We have security systems that can recognize faces and voices and we have text and image generators like Dali and chat GPT. We are going to be talking about all of that because before we get anywhere close to Skynet, there are ethical dilemmas that arise for regular AI. We know that AI can produce unethical outcomes. For example, in 2018, Amazon experimented with using it to review job applications and rate them from one to five. They had to scrap it when it started downgrading applications from women. So, we know that things can go wrong, is there anything we can do about it? There are two broad approaches. The first is build it better.
[2:20]Just build better AI. For example, data scientist Faber says, we need to recognize algorithm decision making has two steps. In step one, the model makes a prediction, like this resume is a 5 out of 5. In step two, it allocates something on the basis of that, like, therefore, you should give this person a job. These are really two separate tasks and therefore, there are two different ways things can go wrong. In step one, prediction, the model might be inaccurate. Like Amazon's AI. It looked at all the resumes submitted to the company in the previous 10 and learned the patterns, which might sound good, but tech is a very male dominated field. So the system concluded from its limited training data that men must be better applicants because they get offered more jobs, and that's not true. So if the problem is accuracy, then more data and more diverse data could help. In step two allocation, the model might be unfair, like Amazon's AI offer jobs mainly to men. So we just constrain what the model can do. For example, tell it that if 30% of applications come from women, 30% of jobs need to be offered to women too. If the problem is fairness, just tell the computer to be fairer and we've done it. We've made ethical AI and all it took was one bit of code, if evil, then don't.
[3:57]But maybe it's not that easy. Here's an dilemma. Suppose we are building an AI to help a university decide which students should be offered scholarships based on their resumes. The university wants to save money by accurately predicting which students will achieve a certain grade level that represents a good use of funds. And suppose at this particular university that students of color drop out more than white students and are therefore more likely to represent a loss of scholarship. The AI we build will learn that pattern. However, suppose that students of color drop out more because there's problems with faculty racism and lack of affordable housing and harassment by campus police. They do drop out at higher rates and race is a predicting factor, but it's not their fault. So, should we alter the model to ignore race? That seems fair, but it would make step one, prediction, less accurate and undermine the purpose for which the university wants to use the AI. As for step two, allocation, we could tell the model that if 30% of applications come from students of color then 30% of offers need to go there too, if racism then don't. But that would mean some scholarships get offered to students who really weren't the most likely to succeed, which again, undermines the point of building it. And you might say, well, that doesn't matter, it's fairer that way, and personally I'd agree with you. But now we're not talking about technology, we're talking about political philosophy. Because it turns out there's a trade-off between accuracy and the kind of society we should live in. And the trouble with that is, not everyone agrees what kind of society we should live in. This dilemma can come up a lot. Another example might be, if you search Google images for CEO, should it show you pictures that accurately reflect most CEOs, which is to say pictures of men and therefore risk reinforcing sexist biases? Or should it show you images of the kind of world we want to live in where CEOs of all genders are guillotined for their crimes. So, okay, how to make ethical AI? We've got build it better, and there are some ideas there, but also some trade-offs. So maybe we also need police it better. AI can never be perfect. So, when it goes wrong, we need to make sure that people have options. In the novel The Trial by Franz Kafka, the protagonist Joseph K. is arrested, and as he moves through the criminal justice system, he's never told what crime he's accused of. The story is horrifying, because Joseph is rendered powerless by the bureaucracy. He can't take action and he can't even find out what action he could take until eventually, he is executed and never told why. Philosopher Kate Brendenberg says that when AI goes wrong, it's like a cave novel. You get denied something and you don't know why. And if you don't know, you can't improve. If your resume gets rejected by the recruitment AI that doesn't give you any feedback, you can't write a better resume next time. Not only does that make you powerless, she says it'll probably undermine people's trust in institutions that use AI, which could be a big problem if we're talking about, say, a court. So she proposes the right to an explanation. If an AI makes a decision that affects your life, you should be entitled to know why it did that. In some places like the EU, this is already a real legal right. But here's another dilemma. If I own the recruitment software that rejects your resume and you say explain this, what kind of explanation, do you want me to give? Do you want me to tell you how the model works generally? Like, do you want me to show you the code, or tell you what it did specifically in your case? There might be some practical trade-offs to be made here. If I'm a big company, then I might not have the time to give you individual feedback. The most helpful kind of explanation would probably be what's called a counterfactual one. That's where we say, if you had done this differently, the AI would have given you what you want. Like if you hadn't spelled the company name wrong on your application, it would have given you the job. That kind of explanation is useful because you can do better next time. Obviously though, if we're giving counterfactual explanations of algorithmic decisions, those counterfactuals need to be true. It needs to be true that if you had spelled the company name right, the AI would have given you the job. Otherwise, that's not actually a useful explanation. And that can present a problem because a lot of AI's are what's called black box models. They use multiple layers of non-linear programming and they're so complicated that their outputs are a mystery even to their creators. We could guess that if you spelled the name right, it'll give you the job, but how can we prove that to you? One solution to this is quite fun. Have you ever heard the nursery rhyme about the old lady who swallowed a fly, so she swallowed a spider to catch the fly? It turns out you can kind of do that with AI. If we have a black box model whose behavior we want to explain, we can build a second surrogate model to approximate a simpler version of what the first one did and get counterfactual explanations from that. It's kind of like when police hire actors to do crime scene enactments and show what probably happened, like it's not bulletproof, but it is an option. However, here's another dilemma. Suppose we build an AI whose outputs we know are unfair, like we deliberately build the racism machine. And we put it in charge of our company's recruitment. It's a black box model, so when a candidate of color gets rejected and requests their right to an explanation, we turn to our surrogate model, which tells lies. We built the racism machine, and then we build the Uncle Tom machine to tell everybody that the racism machine just has some legitimate concerns and by calling it the racism machine, you're actually silencing artificial voices, which is not very tolerant of you. So you use your right to an explanation, but the explanation you get is bullshit. And you probably don't have the time or expertise to prove it. This process is called fair washing. And it was first proposed by a group of scientists in a paper published in 2019. Not content with establishing the theoretical possibility, they did it. They built an AI, they knew was biased. They deliberately built the racism machine, and then they built a second surrogate model that they called laundry ML to laundry the outputs of the first one. Turns out, you can make the results seem fair, even when you know they're not.
[11:13]Weirdly, after that paper was published, laundry ML got a job writing for the Telegraph.
[11:22]Okay. How to make ethical AI?
