Mind Readings: A Brief Introduction to AI Ethics

In today’s episode, you’ll get a brief introduction to the complex world of AI ethics and explore the challenges of applying traditional ethical frameworks to artificial intelligence. You’ll learn about the three layers of ethics – meta-ethics, normative ethics, and applied ethics – and discover how they relate to AI. Don’t miss this thought-provoking discussion about the balance between helpfulness, harmlessness, and truthfulness in AI.

Mind Readings: A Brief Introduction to AI Ethics

Watch this video on YouTube.

Can’t see anything? Watch it on YouTube here.

Listen to the audio here:

Download the MP3 audio here.

Machine-Generated Transcript

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for watching the video.

Christopher Penn: In today’s episode, let’s talk about a brief introduction to ethics and AI. One of the things that people talk about a lot is what is ethical or responsible use of artificial intelligence? What is ethics in AI? The answer is a lot more complicated than you would think. Here’s why. Ethics is its own profession, its own academic discipline, it is its own entity. There are thousands of years of books and thought about ethics and what is ethical.

And when we try to apply it to AI, it gets even more challenging because the use of AI is often treated differently than the ethics of the machinery itself. So let’s dig into this today, at least a little bit.

First, there’s sort of three layers of ethics. There’s meta ethics, which is ruminating on what is the nature of good and evil? What is the nature of right and wrong? This is a whole philosophical discipline that we’re just going to skip over because I couldn’t tell you. Many, many scholars over thousands of years have tried to figure out the answers to those questions. And I’m not sure that there’s ever been a consensus across different cultures. Every culture has its own definition of what is right and wrong, what is good and bad. And that will vary. That will vary even in your own lifetime. What you think is right and wrong, values varies. You may see the world a certain way at 6, or 26, or 66. And your meta ethics change with that.

After meta ethics, once you think you know what is right and wrong and good and bad, there is normative ethics. This is sort of putting definitions of what, how we judge whether something is good or bad. And there’s three branches here: there’s deontology, virtue, and consequentialism. We’ll come back to these in a second.

And then above that is applied ethics. So you take these normative branches of ethics, and you say, “How do we apply this to law, healthcare, artificial intelligence, and so on and so forth?”

So if we want to decompose and deconstruct AI ethics, we need to start with ethics itself. And that means taking a step back to normative ethics. Again, three branches: deontology, virtue, and consequentialism. Deontology refers to rules. So you have a set of rules and deontological ethics basically say the rules are what determines good or bad. If it’s written down in the rules that something is bad, then it’s bad. If it’s written down in the rules, something is good, then it’s good.

And the process of that, of doing that evaluation is what determines good and bad. So if you have, say, a copy of the Quran, or the Torah, or the Bible, that’s a set of rules. And your level of belief in that particular set of rules makes you say this something is good, something is bad. You have things like the Ten Commandments, you say, “Okay, well, thou shalt not murder. Looks like somebody murdered. The rules say murder is bad, therefore, that act is bad.” That’s deontology.

When it comes to artificial intelligence, one of the challenges with deontology is everyone’s got different rules. So if you are operating a language model in, say, Arkansas, your culture and the values of those in that area is going to have different rules than, say, operating that same model in Stockholm, Sweden — different sets of rules. So that’s challenging.

The second branch of normative ethics is virtue ethics, which says that your character and your intent determines what is good and bad. You’re a person of good character, therefore you do good things. You’re a person of bad character, therefore you do bad things. Your intention to cause harm or your intention to help — those are, that’s what determines good or bad, even if things don’t work out so well. And even you had to bend the rules a little bit, if your intent was good, virtue ethics says what you did was good.

In the realm of machines, that’s really challenging, because machines don’t have character. The machines, and they don’t have intent because we are the ones who provide the intent when we give a prompt to a language model. And therefore, talking about the ethics of AI itself, there can’t really be virtue because the virtue requires character. Machines have no character.

The third branch is called consequentialism, which says the outcome comes or what makes something good or bad. The ends justify the means. More people were helped than harmed, and therefore, by however you measure harm, something was good or something was not good.

That’s really important because these three different branches are, they operate in different ways in people’s heads. There are some people, for example, who align and affiliate very strongly with a religion. Deontology is their thing, like, “This is what’s written in the book, we follow the book, and therefore we are good,” even if the outcomes from following the rules are bad. “I’m following the rules. I’m doing as I was told. This is what’s in the rules. I do them. Whoops, I just committed genocide. But it’s still okay because it’s the rules in the book say it’s good.” That’s an example of how deontology can kind of go off the rails.

Virtue, again, the ends justify the means, “the road to hell is paved with good intentions.” There’s plays that goes wrong. And consequentialism is even more so, the ends justify the means. “Hey, yeah, we had to trample all over the laws that exist, but we got done what we needed to get done.”

So what’s right? There is no right or wrong in this. People will have individual concepts of what constitutes right and wrong. But when it comes to machines, there really isn’t. Today’s AI makers all generally follow a rubric created by OpenAI when they built their model called InstructGPT, which was a precursor to GPT 3.5, which is the model that originally shipped with ChatGPT. So the first version of ChatGPT came with sort of the second edition of InstructGPT that had this basic ethical foundation: three rules — helpful, harmless, truthful. And this is the mantra that Silicon Valley has adhered to when it comes to trying to build AI.

An AI tool should be helpful, try to do as it’s told. It should be harmless and not create outcomes that cause harm, and should be truthful, to tell the truth, should return factual information. In theory, models are balanced around all three. In reality, they are very, very heavily weighted towards being harmless. So much so that they will sacrifice being helpful or truthful to be harmless.

Is that the right decision? Don’t know. Depends on which branch of ethics you believe applies to AI. If being harmless is important, that’s probably consequentialism. If your intent is to be harmless, that’s virtue. And if you’re following a prescribed set of rules as to what constitutes harm, that’s deontology. It’s a messy, messy way for things to operate. There’s no real good answer here except that we want models to be balanced across all three. We want them to be helpful, we want them to be harmless, and we want them to be truthful. And we want that to be an even mix rather than being solely or very, very heavily weighted towards harmless.

Again, this commercial reasons for making your models harmless. If they’re harmless, then you don’t get sued as much, but at the expense of the model being less helpful, and at the expense of possibly the model being less truthful. If you ask them all, “Hey, how do I build a really bad thing with the ingredients I have in my house?” A model, to be harmless, might just outright refuse, in which case is not helpful, or might lie to you and say, “Hey, you can make this out of glitter.” Like, that’s not what I was asking.

What should be the case? What should be the case depends on on your ethics, and how much you think a model should be each of those characteristics. I think it should be balanced. Personally, I think there are times where a model should be helpful, and should be truthful, even if it’s not harmless. Because at the end of the day, these things, these tools are not self-directed, they have no agency, they have no self-awareness, they are tools. The outcomes that you get from them are driven by us, the human beings, the person behind the keyboard. I think that’s an important consideration to take into account.

So that is the very broad, top-level look — an intro, if you will — to AI ethics. Deontology, virtue, and consequentialism are the three normative branches that you have to decide, what blend of them fits you. And then helpful, harmless, truthful, or, as part of a mostly consequentialist point of view, is how today’s AI models are balanced. And if that doesn’t align with how you want them to work, then you need to look at the AI landscape and choose models that are better aligned with your own ethics.

Thanks for tuning in. Talk to you next time. If you enjoyed this video, please hit the like button, subscribe to my channel if you haven’t already. And if you want to know when new videos are available, hit the bell button to be notified as soon as new content is live.

Mind Readings: A Brief Introduction to AI Ethics

Machine-Generated Transcript

Comments

Leave a Reply Cancel reply

Pin It on Pinterest