The Ethics of AI – AI Times EP 1

you Hello and welcome to the inaugural episode of AI times from the AI infrastructure Alliance also known as aiia, I'm Mike vizard along with my co-host Lee Baker General Secretary for the aiia. It's our pleasure to bring you the first and what we become a series of ongoing episodes focused on all things AI. All right.

Let's just jump right into it Rebecca. Welcome to the show. What's your perspective on AI ethics these days and where are you from and walk us through all that?

I sure. Yeah. Thanks for having me.

So yeah big questions, I guess I'll start, you know, trying to tie those two together. I I started off. At Stanford and studying AI but actually I started out before that when I was when I was about seven or eight when I was an extra in a movie called Bicentennial Man and any of you remember it but it is Robin Williams movie.

About a robot trying to become a human. And so that was my first sale. That's my first start with AI and my acting career did not work out.

End up at Stanford studying and studying AI. And kind of you know, the question continued. What does it mean to be the human and air world what is mean to be?

Artificially intelligent and so some of these big questions were really really exciting to me. And back, you know back when I graduated not not too long ago about 10 years ago. I was really trying to get a lot of people to to talk about ethics and I organize a seminar can only get a few people to come.

And I was back at Stanford last week and they actually have programs now. So every CS student for every CS class has to there's an Ethics component to every single course and to graduate you have to take in an Ethics seminar and all I think so, I think you know, I think To the question of can we can we get really practical about ethical AI I think. It has to be part of the culture.

And it has to be part of the mindset and and I think we're seeing more and more that as more, you know, these these case studies come out and we start to see how things could potentially go wrong. We are we're getting ahead of it and we're starting to starting to become part of the zeitgeist. And I guess back to my background I ended up in I started out in AI.

Implementing and then in about 2018 Quantum Computing came onto my radar and I implemented enough to really understand where the limitations lie in terms of today's Hardware. At least at that time and so the idea of being in Quantum Computing which is going to unlock so many different use cases when it comes to AI when it comes to technology broadly that was really exciting for me. So I got into Quantum Computing and I'm now I've been in this space for a few years for the opportunity to be at the Forefront of something that will really really Shape where technology goes and so that's that's me.

I currently work in I started a company called Q secure and we do cybersecurity to to address advancing threats one of those being Quantum computing. All right, no ethical issues around cypruscurity these days whatsoever. Definitely not and I you want to introduce yourself where you're from and go from there.

honey, I'm with you and to kind of both answer your question and also leaving my experience in the area of AI ethics. Um, so the answer is absolutely yes. We can have kind of AI ethics in a practical form but one of the challenges that we have is a lot of Frameworks a lot of principles that we see out there.

There are either high level or very abstract. So for organizations, it's hard for them to implement those Frameworks or bring those principles into their workflow. So part of my role and responsibility is to work with our customers.

To create the applied version of these Frameworks that would be applicable to their use case the way that their processes work and also think about what kind of features what kind of tools we can have in our product at data robot that would support our customers better. If it's around privacy, whether if it's around bias and fairness how we can put these tests and tools into our own products and provide it to our customers. All right.

Thank you very much. And Mike. Let's kick it over to you and Everything they said was wrong and I can't wait to shred it.

It's hard to disagree with those points. My background was video games military simulation about PhD computer science, but in the VR side, not Ai and I retired and got into AI unretired to focus on Black Box because I think it's so important that we're using techniques that may or may not be ethical but nobody knows to apply to such safety critical human critical decisions, whether it's parole or food distribution or weapons start getting and likes so I've got to start a dive playing where we built fully explainable machine learning solution. So I think of ethical AI it's sort of two fronts one is the notion of the sort of social issues of ethics and how ethics vary so greatly from country to Country.

And so the notion of let's build something ethical on a global scale doesn't make any sense Google invites you or both global companies, right? So how do you handle the ethics? Well, it's really about allowing the inspection of what's going on from Soup To Nuts.

I liken it to a coffee where if you want to responsible cup of coffee, you have to start with responsibly sourced organic beings and not underpaid the farmers and not exploit anyone along the way through transport all the way to not having your birth to make three dollars an hour and work 15 hours shifts. The whole thing has to be ethical for you to have a cup of coffee you feel proud of and AI is the same way you start with good clean data that you know, what's going on where you got it. The Providence is safe everyone along the way believes that that data should be in your hands for the use that you have for it all the way through to an expectable transparent algorithm at the end to Complete the task you want to complete?

How's that? That sounds great. And of course, my co-host is Lee Baker who's from the association Billy explain your role and how you got into this whole function and jump us into the next question.

Thanks, Mike appreciate that. So yes, I act as general secretary for the AI infrastructure Alliance or Aya as we call it Aya is designed to be a forum an alliance for ISB software vendors in the machine learning space specifically kind of mlot space because we see the need for best of breed components when you're building out your machine learning pipeline not least for reasons of ethical best practice. So hopefully Aya is not ducking the big questions.

I think we're starting to get into some of that today. So what better place to lead off because some of the notes that are yes touched on already is there are a number of different kind of Frameworks or ideologies or regulatory kind of principles for for ethical AI. How do we how do we Mark our homework?

Where's there one abiding principle that we should be daring to or do we build out our own kind of appropriate AI data governance? What does what is the panel? Think about how we govern this from a corporate and an individual responsibility?

I like things how you kind of prompted that I'm gonna kind of push that to you. So, I mean there's so many different ways to skin this cat, right? I would say that.

It this isn't an individual concern other than building a good culture, you know, Rebecca talked about a training in ethics and we just released a project to the data and Trust Alliance, which is a super neat organization of Fortune one and Fortune 100 companies focused on literally data and trust and the project was m&a due diligence if you're going to buy a data or AI company or if you're going to invest in one, what should you be looking for? It's a simple as that and that's a hard question to answer if you're Insight or if you're GE and so much of what we build in that framework and it's I think publicly available at this point was about the culture of the organization because at the end what's really going to matter is that every implementer throughout the chain is thinking about how does this how am I treating this data? And what is the algorithm I'm building and how are we double checking?

We haven't built something by us. And what is our audit process and you can Check each of those boxes. But if the reason for it isn't there, it's still not going to succeed.

So it comes down to does this company have a mission statement that involves trust it does this company understand customer that the data where they're getting what the value is and so I guess it comes to this individual level of understanding what ethics are and then trying how do I Implement those on an everyday practice to hit the Gestalt of what you're And I can I follow up on that with you. I'm what exactly is the difference between ethics and say just trying to optimize something and I'll give you an example. The Department of Justice is looking into some SAS application that is designed for landlords.

And it optimized, you know rental property prices on behalf of those landlords. And I guess the question then becomes are they manipulating something at the expense of tenants? But you know from the perspective of the landlords, it's just an optimization.

They're trying to figure out that you know, what is their best price capability and then using an algorithm to do that. So is that a good instance of AI or should the tenants get together and build their own countery to get their own ethical set up going where they're gonna say Hey, you know, we're gonna figure out algorithm to get the lowest price rental property and you know, are those ethical considerations or is that just the fundamental nature of capitals So the way that I would look at it is, you know, when you're building AI part of the process is optimization. We are to trying to optimize for a solution for problem that we have and get the best outcome out of it.

But what we miss here. Is have we considered human as part of these process? Um, you know to be able to understand what would be the impact of this system.

That we are going to apply whether if it's on tenants better if it's on. employees who work on shifts But would be the outcome of this optimization on them. Are we providing something that is be considered their feedback.

We considered their kind of aspect into the process or be purely thought about optimizing the cost or increasing efficiency. So that is the aspect that's missing in these parts. So we didn't bring tenants as a stakeholder into the process of building it and then we are thinking about the ethical aspect of it one component of it is also bringing all these stakeholders to the table.

Then we are thinking about building any ice system and also Thinking about running an impact assessment of the system how this system what are the benefits of the system? What are the risks of the system and the impacts that is going to have and how we can actually think about maybe it's a good system. But we need to tweak some aspects of it to make it work both for the landlord side and also for the tenant side.

So these all these considerations should be part of building in even when we start the process even prior to building thinking about these aspects of it. So for me the you know, the the part that in many cases we might be missing here is the consideration of the human aspect of it. So that just seems it seems so difficult to do it's a little bit because that's a purely adversarial relationship in some ways.

If you're thinking short-term if you're thinking long term then yes, you want Happy tenants over 10 year cycle, but if what you're trying to do is maximize profits in the short term you're absolute job is not to make the tenants happy but to maximize profits which is probably maximizing rent the idea like I love that notion of stakeholder, but it's sort of like saying let's make sure we bring the Targets in of this military targeting system, you know, the guys who are building the missiles to shoot at our ships. We should really talk to them first before we set up an AI to defend against missiles, right? It doesn't make any sense because their goal is not aligned and all I think that system is just simply echoing the goal.

And I think you there's sort of core problem with if you have a purely capitalist society that doesn't particularly care whether the tenants are unhappy or some people can't live there anymore. But I don't think of it as exploitative to do that if the laws and our ethics think of that as acceptable because someone has to disagree it's a panel. So was it what does that mean?

Then in terms of role and functions? So, you know, we hear a lot about kind of the chief data officer or kind of data governance roles, you know Rebecca and you'll mind what is the function look like that can kind of own this on behalf of the the organization or customer. So yeah, it's a great question from I think there's there's you know, the both questions bring up kind of the I guess the question of I guess altitude or different groups these different stakeholders, right?

So I yeah as part of a group that authored some governance principles outlining and taking a lot of learning is from HPC and from AI outline. Principles that governing bodies can can start to think about and Implement, right? So you've got your governing bodies and then you know on the other side.

I think we have to be thinking about one thing that we talk a lot about is a huge part of this is empowering the people so, you know in your previous question the tenants but anyone who's data is going to be used I think there's you know, I think we're evolving past this but traditionally there has been a lot of kind of gatekeeping in Ai and deep Tech right? You can't understand it unless you you studied it but I think popularizing this and and getting across the awareness to everyone who this is going to affect so that people can sell Advocate and but back to your question Lee, you know in my in my company, we're relatively small. We're 50 60 people.

I'm Chief product officer and I so I've kind of taken it on myself we've Things like have a product management checklist for each stage of the design to build to test to iterate process where you're thinking. Okay, where am I getting this data and my representing all demographics that are going to be affected, etc. Etc.

So I think you know, right right product managers have a good responsibility. I think it's often got to come from the executive level as well. Hmm.

So Mike following up on that do you think that you know, ethics is a slippery slope? So do we need to just kind of say there are certain use cases involving say life and death issues where we need to clearly have some sort of ethic supplied but it's not this, you know, golden metric that can be applied to every transaction or every application in the world. We need to figure out some sort of balance.

Is that where we're looking for here? Well, I mean, you know from our perspective writing explainable AI I'm not trying to Target advertising optimization. 3 my percent difference between the Facebook ad price of Target for let's say literally the company Target is black box or not, you know not really worried about the ginger bias that might occur in that.

I'm a lot more worried about what happens in residential loan decisions or parole or the others. So yes, I think the the level of let's say ethical scrutiny and time spent with stakeholders is much higher when it's more safety critical human critical. So yeah, I think absolutely that that notion of playing by the rules is a tough one because I think we have bad rules right now.

And so if you're playing by rules to gather data, let's say your LabCorp, right? What's the date of your gathering on people? You've got their full DNA.

There's a whole lot of numbers in there that sorry statements in their Eula where they're allowed to do all sorts of things with that data and then resell it right. I don't love that but they're playing legally within the rules right now. And so are they bad people for using that as their primary revenue Source selling my DNA information without my permission.

I don't know that they are I don't like it. I would sure like that there were big warning label. I'd like even more for me they'll delete it but as it is those are the rules here.

With that is sadly ethics encoded into regulation. So Rebecca talked about kind of owning the process kind of from a product pointed view. Should that be the case that she marked our own homework or should there be kind of a higher power to kind of audit this this activity?

What's your second hand? What's what's your view on on kind of who wields the big stick? So part of this would go through, you know.

The internal process having a thorough documentation of everything that we've done all the decisions that we made so that would help us really be able to audit ourselves. And also be able to trace back if we see any issues any anything that we want to review we would have that documentation. We would have that way of knowing what happened in the process.

So that is one way of we can think about auditing it. Another aspect that came up was in terms of providing that information as to external stakeholders, but if it's the users better if it's the you know people that the system is going to be applied to so having that level of transparency and explainability that here's what we did. Here's how we better if it's collection of the data, but if it's how the system was built and also kind of providing the the possibility for them to be able to Opt out if they want to but in terms of you know a higher system to do the auditing process, for example, we see that coming out of AI so in some cases they want.

To have independent Auditors for confirmity assessment aspect of it. So it really depends, you know, there are cases that the stakes are very high. So You may want to focus on having external Auditors to also review the system, but for their the ones that they're not.

You know as AI act colleagues, you know high risk if they're not high risk, maybe your internal auditing as long as you have a framework that's aligned with what The Regulators want. You would be able to you know, do your own auditing process? Rebecca and you want to follow up on it?

I mean in your studies. Is there going to be maybe someday some sort of certification that will be issued by some body that will audit and AI model for ethics and fairness. And is that feasible?

Yeah, like an organically grown kind of sticker on the box. Yeah, and yeah, I think I I hope so and then also I'm you know, I'm both an optimist and a pessimist right? I'm an awesome Optimist and that you know, I really I really believe in a world where AI represents our better better Angel so to speak and and that we can you know build towards that but I think you know when you when you really get into it.

For example, when we were putting authoring those principles for governance you get into kind of these like philosophical rabbit holes, right and it's you know, it's probably similar to any other a lot of other areas of law, but I think you know and Mike was kind of saying this it's hard to make a one size fits all and I think there are certain things that you can do. There's low-hanging fruit and you know, for example parole make sure that your auditing your your models for you know, any kind of prejudice those sorts of things. I can we'll start to see it.

We have as part of a group that did a course air course for an ethical technologist certification. So I think you'll see these one-offs. I think it's challenging to to do it from a regulatory level, but I think we'll see we'll see parts of it for sure.

I'd say from the other side. We don't see that Good Housekeeping seal approval on software systems and software systems are fully inspectable. Like they can be complicated.

It can be a million and a half planes of code, but the algorithms are there and they're inspectable when you're talking about 160 billion artificial neurons. Nobody can inspect that doesn't make any sense and it's a harder problem. Right?

Because you're trying to certify The Good Housekeeping of both the data and the model you've built with the data. I think it's possible and that's the reason we don't have it right now is is this software ethically built again, right? That's the trivial example of doing this for AI and we've been writing software systems since the 50s and we don't have that at all.

So if we cancel our software, we're not going to solve it for AI. So what's left then ideally at least you have the sort of auditability in in the rear view of well, we build this model. We use 20 years of parole data.

We did a fantastic job left nothing out and then it turns out of course, it's completely biased by the after lunch. Of the judges where they're tired and cranky and it's biased racially because you took data from the whole nation and blah blah blah and at least we go back and look at it and we can say oops. Yep.

I can see here. This decision was indeed racially biased or time biased or whatever. It might be oops.

That should have been in there and then we fix it and move on at least we do that with software. I don't think we'll ever get it with AI on the path. We're on it's a dead evolutionary path for ethics.

How's that? So is that is that a source of optimism that because the bias can be coded and made transparent that's better than it living in my head where I can you know apply my bias kind of subconsciously if explores the model and we understand its outputs and we can shine a spotlight on its bias. Should that be a source of optimism for us or should we be kind of pessimistic that we're just kind of a man coding existing biases?

Yes. So imagine you have the racially biased judge, right? So how do you pull that out only after years and years.

Can you go back and figure out? Oh, wait a second. There was a problem here because any individual case you can make an argument that it's okay that she didn't get the credit card and he did even though they're married but it was this one case, right?

So you have to have this pattern of behavior much easier. If you train an AI built on that judge's decisions, you can more quickly determine whether or not there's a bias issue from that model then you could from the visual because ideally software or an expectable inspectable. Ml model will be more inspectable than anyone's brain, right?

So yes, there's a case for optimism there right now though. What's happening is we are scaling human fallibility as quickly as we possibly can. We're scaling it like crazy without any ability to understand it any more than a human.

To that point and I will we need AI models trained in ethics to inspect AI models built by humans or bias and you know, can we use AI to police AI? Um, but there are ideas around that but again, it comes down to first of all, what are our values that you're trying to build it, you know? To kind of piggy back on the conversation that we had just now in terms of biases even many times.

How do we Define our bias? Right. So before even thinking about you know using AI to actually try to reduce bias our first step is first to figure out what bias means in that specific use case that we have.

Right, so when we figure that out and it's a challenge of as much as we want to be optimistic about things being easy on the Epic side, you know even just defining fairness. Is a challenge based on use case because you know when you pick one definition and you want to be fair in that one, you're going to be biased with another definition. And when you know in these situations, someone would always be unhappy about the results.

So first, we need to figure that out on our own and then we can think about how we can actually use AI to help us. Mitigate the biases I'd be seeing the system. So there are tools that we would be able to use to mitigate if you see a bias in our model or if you see a bias in our data from the technical perspective.

So we already have those two. It's a matter of figuring out from ourselves how we should be thinking about the ethics how we should be thinking about the bias aspect of it before even thinking about using AI in this regard. I think hiring is an example where you always want bias of some sort.

You're trying to build it. Right? Like I'm trying to build a more diverse Workforce.

So I'm purposely biasing in a certain direction. It happens to be a legal direction to bias. Right but I am biasing by protected class a little bit because I'm trying to build purposefully more culturally and gender-verse Workforce, which is tough in an engineering company, right?

But so it starts with what are we trying to do and then build a system and then double check that that systems actually doing what we want it. We want to younger Workforce. We want a more experience Workforce.

These are reasonable things to say summer illegal some are not right but at least understanding what you're trying to accomplish. Let's really hard problem like you say because it tends to be that that strategy comes out of action like it just you notice over a few years. We've been getting a younger Workforce and you realize it wasn't on purpose at all, but it happened in the tactics.

So let's roll back from the model a little bit and talk about kind of the data. You know, the data science true isn't is garbage in garbage out? And so you see that there is a a an appetite an avarice or as much data as possible which inevitably creates data race symmetries, you know, some people have access to data that organizations don't how do we how do we govern the firehouse and kind of the data capture?

We've got on my side of the pond we have gdpr as the regulatory function that's not exactly been wildly successful in managing Bad Bad actors. You know, how do we go beyond that to kind of understanding who who has their hands who has their greedy pause on the data in the first place? Well, I'll throw something up semi-technical Garden like while everyone stares that I think large data sets are sort of an artifact of a bad transformation process for current Black Box.

Ml. Right, you know, why do we need a hundred million data points to be able to understand credit card transactions? Because you're using a lossy method of compression, right?

Essentially. That's what you're creating. Is this lossiness.

We throw a lot of stuff out and ideally when that models created it doesn't throughout the good stuff and keep the bad stuff and then overfit to something incorrect or learn shallow, really or whatever else but ideally more mathematical driven AI won't require the 100 million records to try to solve a problem now mind you there's some problems you actually need that because the shape of that hyper-dimensional information space requires it because it's so complex, but most often you don't so I'll just say maybe that'll get solved in the next couple years and reduce the Haves and Have Nots totally different answer your question, but that Rebecca sometime to think well, let's follow up. That with the questions they were back here because I'm on this side of the pond. One of the most famous quotes in history is what are the most nine dangers words in the English language and it's Ronald Reagan saying, you know, hello.

I'm from the government and I'm here to help. So what is the role of government going to be in this business? Because you know, we have gdpr would make success but our government's going to get involved and you know, I find it hard to believe that some Congressional office somewhere that's going to write a bill is going to understand how AI algorithms work.

So, you know, this doesn't vote well from my perspective, but what do you think? Oh man, that is a good question. Lot of the in this Irish security space we have to get a lot of compliance is very important.

Right? You've got sock 2, if you want to work with the government, you've got, you know this whole idea of fedramp and your nest standards. And so I think you know, I think there there are areas like that where you get a bunch of really smart people together who know their stuff and you get them to to put forth ideas how you can how you can keep these things under control.

And it's it's really a hard question. I think you know the hard thing is AI is moving. I'll speak for myself AI is moving faster than I thought it would things are things are really starting to accelerate.

We had a few huge milestones in the last year for anyone who hasn't gone out and played with GPT 3 or Dolly. It'll blow your mind. I promise so these you know, these these big models these big things that were really starting to be able to do and and one of the hardest things for AI has traditionally been deal with context and we're starting to be able to do stuff like that.

So bringing it back. I think it I think AI is moving faster than the speed of government traditionally has and so that's that's kind of where I stand I think it'll catch up but we gotta get ahead of it in in other ways as well. I just say that there's the transition of taking old technology decisions and Regulation and trying to apply an AI.

I hate the open source push in AI, you know, if only your AI is open source, then it must be gooder and betterer for society and it's got no impact whatsoever on bias. In fact, it just literally biases software development contracts towards ad supported companies that happen to make AI you know, Facebook's Google's whoever else they're not the AI business. They're in the giving it away business to drive the size of the internet because Google gets a cut of the internet, right and that's their whole business model.

Whereas little companies like mine are trying to sell AI we can't open source it. Yeah we can't do but we're not allowed to make software in New York now, even though we have explainable ai that does exactly what they want because a well-meaning regulator decided that gosh what we need is open source to protect the people and that's just it's sort of a faulty analogy, but I get why they're grasping straws like you Wreckage it's moving too fast. I don't understand what's going on in Ai, and I do it every day because it's moving too fast.

So that end Mike so what we're seeing and and Rebecca referenced some examples of this is the ability for kind of your citizen data scientists to be able to create models using natural language text. Is that a good or a bad thing? And we're making citizen data Sciences of everybody or actually we further abstracting what's going on behind the curtain and making it harder to interrogate what's going on.

What's your view on that? Ooh, that's a really good question. I spent a lot of time trying to democratize video game development.

So you didn't need 50 hackers in order to build something cool. It could be two or three people in a garage could make a mobile game and literally somebody just made nine billion dollars off of a mogul mobile game that was looking at so that's I love the democratization of tools. But anytime you're using python instead of writing machine assembly.

You're too far away from the hardware to understand what's going on right great question and for sure people using mid journey and Dolly and chat GPT have no idea what they're thinking about. Ethically where those models were sourced where the images were sourced the fact that you can you can literally create copyright and code within what's the co-pilot the Microsoft tool it will spit out copyrighted coach chunks for you because it trained on copyrighted code chunks and we'll help you break the law, right and there's a warning on the box that says that that's I think let's just start with people understand what AI can do because I think that's beautiful thing and we'll train them later what it is that can do wrong as opposed to not exposing. How's that?

All right. He tried humans guys. We're down to our last question.

I'm going to give it to hanaya. Will there be a backlash because of all these issues and will people decide that they don't trust the AI models and then once they stop trusting the AI models, they'll stop relying on them. And then the whole thing will come for no particular purpose.

So for this one, I would put my pessimistic hat on because pretty much you're using AI on a daily basis, you know, it's on our phone be busy all the time. So as part of that definitely, you know, there are backlashes that we see from unfortunately many many headlines that we see about the AI gone wrong. right, so Part of the problem is that you know because AI is so embedded in our lives it's hard for people to really separate the two but one thing that I would add you to that is, you know, because of all these things that headlines that are making people are now paying more and more attention to it.

So they are getting more and more educated on the these aspects of it as well. So for example, if you look at five years ago, if a headline came out, The level of awareness for people about issues. Let's say like Instagram Facebook issues or you know, all the other companies that made headlines.

They may have not really thought about it. It's just one of thing, but now they it's constantly on their mind and they don't need to be technical even you know, People who are using AI whatever we want to call them citizen data scientists or or just you know AI users now they know about these things more. So part of the process is that it's not necessarily would be a backlash that they would stop using it now.

They're using it in a more conscious way. They're paying more attention and would demand more so it won't be halt but rather now, there's more kind of feedback and engagement between the creators and the users. So we would see that aspect of it moving forward.

So it would be a slower process in terms of you know with regards to you know, severe backlash then the creators need to rethink everything that they've done but rather slow process of improving it because the users are demanding more ethical and more trustworthy systems. All right. Well, it sounds to me like we need a lot more AI education out there.

So the users can ask those intelligent questions. Hey guys. Thanks for being on the show.

This is our first episode. We hope you'll all watch the next episodes as we go along here with my host Lee and with that any kinds of suggestions. Feel free to share them with us as you find your way to our website and various other resources provided by our friends over it the AI Association Network co-hosting this way folks.

Thanks again for being on the show.