How to Develop an Org Wide Ethical Approach to AI

My guest today is Jonathan Conradt. He recently left Amazon where he worked as a principal AI scientist. We're going to be talking about AI and machine learning today, and what needs to happen in terms of education around these technologies that are being folded into the work people are doing every day. Jonathan, welcome.

So eventually at Amazon, what I discovered was that there were great AI projects that weren't being approved. And there were kind of bad AI projects that were being approved. And the principal problem is that Amazon has really smart people, really smart vice presidents. They're well educated. They come from great backgrounds. They know their business inside and out. But AI was coming out of left field to them.

They didn't really know how to deal with it or what it involved, and they were struggling to make good decisions about what should we invest in. And so they were missing out on great ideas and sometimes investing in things that weren't going to work out very well. So my, about my last year at Amazon, most of what I did was prepare curriculum and teach vice presidents, directors worldwide at Amazon about machine learning and AI. And then ultimately I decided.

If it's this big of a mess inside of Amazon with these people inside of a tech company, what is it

like outside of Amazon? And so I decided to leave Amazon and do things like this, try to help people understand what AI is, where they need to be cautious, where they need to be aggressive and help them understand how to make the best of it. It's an interesting and very poignant time to do that. There's a lot of things that I think AI and machine learning can do, and some of what companies have cast their eye on it is sort of tied to it being the shiny new thing, right?

Folks have some really big ideas, but there's a lot of practical solutions as you and I were talking about before we got on this call that are much more practical and can help companies tackle immediate needs. I was wondering, can you give us some good examples of this and where can HR implement something for immediate impact? Yeah. So generative AI is that kind of shiny ball that's catching everybody's eye, right?

But it turns out traditional machine learning has also made big advances in the last few years. One of those things is called automatic machine learning. It's basically what happened is scientists were looking at the work that they were doing. And they realized that they were basically following the same steps every time. And there was a potential to just automate all of that. And there's a really good open source package called AutoGluon. And it's quite remarkable.

When I looked at it, I had a bit of an existential crisis. Because I realized that here was a package that Someone who had decent SQL skills, who knew a little bit of Python. They could produce a world class model overnight that would have taken me potentially months to create. That's really useful. So in HR, there's a variety of things you need to do. Like, you might want to classify things. Like, for example, you might want to identify people that would benefit from, say, leadership training.

Or employees who are at risk of attriting. Those kinds of things. And these models can do that for you. There's also types of predictions you want to make, like they relate to numbers, like how many resumes do we expect to have, or how many people do we need to man the warehouse in two weeks, right? Those kinds of kind of predictions of numbers, it also does really well. And it can also do time series. Time series is useful for things like Understanding, you know, the seasonality of data.

So I'm not sure how that relates directly to HR, but I'm sure there is some things that are seasonal in HR that they want to be aware of.

So what's remarkable about this is in three lines of code, someone who has the data can basically create a world class model. I think that's just remarkable.

HR. Let's talk to us a little bit about the biggest ethical considerations with AI and HR, and what are some of the things we have to solve for in these models before getting it more involved in what HR does?

For HR, these kinds of laws and concerns are top of mind. And so there's a unique opportunity for scientists and people doing models to incorporate HR and ask them good questions, like things about how do we want to manage gender and what sensitivities do we have around race, those kinds of things HR people know inside and out, right? But the other thing that could happen is within the HR project itself. So let's say an HR organization wants to go and do something.

All of those issues are very sensitive. So the classic example is, you want to automate resumes as they're coming into the organization. And you want to classify them, or you want to sort them in some way to be more efficient. And there's dangers there, because Machine learning models, they cheat. They're just incessant cheaters. They're going to find the shortest path to the answer that they can get.

So you can see things like, the model will learn that if they mention in their resume that in high school they played lacrosse, those people get hired and promoted within the company historically. And that could be a historic bias. And it's something that you don't want as part of your filters for, uh, hiring or interviewing people. And so, you know, you wouldn't necessarily expect that word to end up being important to a model. But like I said, models cheat.

And they're going to find all of those little things. Responsible AI is, it's a whole suite of ideas around how do we best manage the relationship between what our goals and what people and the people that are involved, right, maybe their customers or their employees or job applicants, how do we manage their data, how are we fair across subgroups, how do we, you know, expose to them honestly and transparently what the machine learning model is doing.

And so this is an interesting area of ethics and again, HR professionals, they spend a lot of time thinking about ethics and how to treat people well, you know, how to treat them correctly. So responsibility is a fascinating area for HR and it touches upon a lot of HR solutions. Like, let's say you wanted to build a model that was going to predict who of your current employees is going to become a vice president in 6 years. That's a useful model, right? Because you want to retain those people.

You want to develop those people. That's a good model at the same time as you're creating that model. You need to be very careful because let's say all of your existing vice presidents that are your training data, mostly graduated from Ivy League schools or, you know, maybe their mail and all of those kinds of factors are in your training data. And so to do the right thing in terms of responsible AI, you need to remove the factors that shouldn't be considered for, for consideration.

Don't let the model see or make inference on things you wouldn't consider in an actual promotion. Yeah, Responsible AI is, it's really interesting, and there's classes about it online. There's lots of papers about it online, and regulation is coming. It's already in the EU. It's coming to the U. S. quite possibly. You know, it's been introduced in Congress a couple times. And this is an area, again, where HR is probably ahead of the curve for a lot of the scientists.

And so you can be aware of it, and you can help them understand.

It's like everything, every action has a reaction, right?

Like one of the signals is, Did you restate the question? Let's say you ask OpenAI a question about something about, I don't know, the Korean War, and it gave you an answer, and then you basically ask the same question in a slightly different way. That's a good signal to them that, Oh, we didn't do a good job. And they eventually asked the question and got an answer that they appeared to like. So that's a good signal to us.

But of course what happens is, if there's a gender bias, or income bias, or geographic bias, In those interactions with the system, the system is naturally learning to do the thing that Dom and a group likes a great example. That is actually and I'm not sure this is true, but this is this is the rumor. Open eyes models had some peculiar language that they were using. Like, there was words that they were using that were happening more frequently in their responses.

Then like you and I would have a normal conversation and someone finally said, Hey, this looks like the way people speak in Kenya. I think it was Kenya, but it might have been another African country, but it was some African country. What happened was during the process of creating these models, there's a process called reinforcement learning with human feedback. And so what you do is you ask a question, you generate five responses from the AI, and then you choose which one is best.

And a couple of things were happening. One was people were choosing the longest one, which wasn't necessarily the best, but it just seemed like it probably should be the best. And so what happened was the model started producing longer and longer responses. The other thing that happened was people would naturally. Click on the ones that sounded like the way they speak.

And so we had this large group of contractors in an African country, and they appeared to be responding to the language usage and tone that was familiar to them in their own country, which is slightly different than American English and slightly different than British English. And so this got picked up and people were starting to create tools that could pick out AI responses. Based on this weird distribution of words, I mean, weird within the U. S. context, right?

So there you have an example of, well, biases can creep into A. I. in a whole bunch of variety of ways, right? So if more men are using it, it might end up responding in

a way that is more satisfying to men than women, perhaps. Or it could be take a, a deeper interest and responding to things that the more popular, the more common population is, is talking about, or it could end up mimicking back to you the dialects of the trainers.

I guess my question to you is how much can we actually put guardrails on it with employees? And is there an element of executive education around these tools that needs to happen before anything gets implemented and handed to employees, essentially?

There'll be guard models so that when someone types in, you know, what is the best formula for building a bomb with fertilizer, When that question comes in, there are guard models, and they're small, they're fast, and they're highly tuned to, uh, identifying violence, or illegal things, or unethical things, and they just take the question and drop it on the floor, because there's no point doing the expensive processing after that, because we're not allowed to answer

those types of questions, right? And then there's another guard after the A. I. is responding. And this guard's job is to make sure that you didn't trick the A. I. into telling you something you shouldn't tell you, right, like how to build a bomb. And so there it attempts to identify the fact that A. I. has gone off the rails and drop it on the floor. So there's kind of a technical term for this. Yeah, guardrails for employees is an interesting problem because everybody has a phone.

And in my own life, At home, I've built a machine that can run an AI and I've, I've got a way that I can get to it on my phone and I can do anything I want with that machine. I could talk to it about anything. I could send any documents online. And of course, if you completely cut off employees, you provide them with no way to get to these types of tools. They will just find a way. Because the tools are too valuable for them.

It takes the 20 minutes of writing an email into five minutes of writing an email because they can, you know, get it started and then go back and edit it. People are going to want to do that. The big problem with cutting people off completely and not providing any outlet is also that they're going to use models that you have no access to or control over or visibility to. And they're going to do really dumb stuff, like they're going to upload documents from your company to these, to these models.

If you go look at the terms and conditions for open AI and Google's model and probably all the other ones as well, it says anything that you enter into this model can be used by the company in perpetuity for free for improving or developing the model. So you really don't want to. You know, someone uploading your business plans for the next year or your budget or list of employees like that would be a nightmare. Fortunately, there are ways to provide nice guardrails.

Like you can get corporate accounts now for most of these AIs. My former employer, Amazon, they have a really nice way for you to have basically your own AI that everybody can get to and it's, it's sectioned off. So things aren't being uploaded and shared with Amazon and. I think that's a really good way forward. Most companies are going to provide the same thing, but what about executives? So now we've got the problem of executives.

Well, first of all, we know that executives are just as likely to upload sensitive documents to AI as everybody else. In fact, they might be worse than others at it. And, you know, there's a limited understanding of like, where is that AI actually sitting? It can be hard for people to tell whether it's Say it's within our own IT, it's protected, it's safe, or it's a random third party on the internet, right?

It can be hard for executives to figure that out because, you know, they're experts at their work. They're well trained, they're, they're smart, they're well educated, but they're not IT people, right? So then you run into the, the issue of, okay, once we've kind of trained them, phishing training that we all have to go through, right? Once you train them, don't click on that link, don't upload that document. It's a really good idea to try to provide them access to something.

Something that they can use that's useful for them, that's, you know, well protected, that's official, that's, you know, logged, and all that kind of good stuff. But those aren't projects. So projects are, hey, we're going to replace our entire customer service department with this AI. Those kind of initiatives that Some companies have already attempted, right?

So there's, there's those early adopter companies that ran really fast, spend an enormous amount of money, had enormous difficulties, probably weren't successful, but they were early adopters. The companies that are thinking about these projects now, I would term to be more fast followers, right? So they have an opportunity to look at, okay, what is working in the marketplace, the prices have come down dramatically.

So they're not going to spend quite as much and there's been some good research about how do we measure the impact of these kinds of things. But, you know, just like let's take this, the customer service thing, I don't know anybody that's like, I really hope they push me off to chat so I can talk to a machine about my problem. I don't know anybody that is looking forward to that and definitely not your best customers. And so while there is an opportunity to.

Maybe make some things more efficient and use AI in those realms. You're going to have to be careful about it. And I would suggest that one of the first things you want to do is have a good model in place where you can differentiate between an expensive serious problem that you can recover from and these calls that are costing you a lot of money. But, you know, are are really simple things that are easy to answer. Like, how do I change my password? Those kinds of things.

But yeah, executive guardrails are that that's a real problem. And that's 1 of the reasons that again, why I left Amazon was so I could have those conversations with companies and have that deep conversation about, okay, well. You know, what data do you have? What experience do you have? What problem are you trying to solve? Are you trying to make more money with this? You're trying to save money with this. What are your goals? And, you know, have you tried something simpler?

You know, that's also a possibility. And so I invite people to, you know, put 15 minutes on my schedule. We can have those conversations and decide if that's something that I can help them with. But yeah, it's gonna be very hard. I think there's three kinds of companies. Like I said, there's the early adopters. They've been trying stuff, spending a huge amount of money, taking big risks, and probably not succeeding. The second group are the never evers.

They're the ones that are convinced that, oh, this is another crypto. It's like, you know, it's a big fad. It's, it's not going to work out. We've avoided these things in the past, we're going to keep doing everything the way we're going to do it. Those companies in particular are in big trouble, because it's not that, this is more like the coming of electricity.

And then there's the fast follow organizations, who are just now, looking at the data, figuring out how to measure what they're going to do, you know, starting pilots. They saved a lot of money because they, the prices have come down and now they're kind of getting into it. I think those are the interesting groups that are going to succeed and do really well. And by the way, they're going to catch up with all of those early adopters. Because the performance of the models is going to hit a peak.

Okay, this is as good as they're going to get for a long time, right? There's going to be this plateau they're going to hit. The early adopters are going to get to that plateau first.

Even if you started late, the rate at which these things are improving is so steep that someone who is a fast follow company is going to get there just in plenty of time.

And you all use that, and you give the list of columns or rows that you want to add, and it takes that as an input, and it produces an output. In AI, that is all it is. So if you think about that sum function, In between times that you use it, it's not wondering about what it said most recently. It's not plotting against you. It's not daydreaming. It's not, you know, it doesn't exist in some ways, right? It only exists in that moment that it's being used and it goes away.

AI is fundamentally a mathematical object that's a function. And so it takes an input, processes it, produces an output. A lot of the amazing magical things that we see that it can do are actually augmentations and code that we've written around the AI to organize data and things for them. And that's all complicated. It's a little bit like expecting everyone to understand exactly how a fuel injection system works. Most people don't even know what that means.

And yet they can drive, and they can do their lives, and they've got these reliable systems. And so, I think the, the, what we need to do is for each group or subgroup within an organization, we need to help them get to the right level of understanding. Right? Some people are going to be users, some people are going to be technicians, some people are going to be deep in the weeds with these things, and some people are, are going to be just beneficiaries.

They never interact with the system, but it's just doing good things on their behalf. So that's going to be tricky. HR has got a lot of, you know, HR is usually responsible for training, right? You know, they got to figure out who to bring into train and what kind of training to invest in. And it's going to be tricky, but you know, having a peanut butter solution across the organization, like everybody's going to get AI trained and we're all going to provide the same thing won't work.

When I was at Amazon, one of the things I worked on, you know, I'm teaching these vice presidents and these senior vice presidents. And one of the questions was, what can I do for the people that drive our trucks, the people that put the products in boxes? You know, a lot of these people have, they don't necessarily have a college degree, they're hardworking, they're good people, they're smart people, but you know, they're not technical.

How do I bring them up to speed on AI so they don't, um, number one, so they don't fear it, right? The company needs to adopt machine learning and AI, right?

You don't want to be left behind because you were scared of it or your workers were terrified of it, right? And so to get over that, there has to be training that's appropriate for them. It meets their needs and answers their questions. But also, you know, there's that training at the executive level, which is quite a bit different.

A long time ago, I was at the Gallup organization and I helped develop StrengthsFinder. And so this was an instrument that was about helping managers understand their employees. And through that deeper relationship, the goal was that everybody in the organization would benefit from that. There would just be a better understanding of each other, greater support for each other. And a lot of goodness comes out of that, that today I think we would refer to as part of wellness.

But yeah, and what we're trying to do with AI is, first of all, in a more interactive way, you know, learn from people about how they're currently doing. One of the things that's frustrating in a lot of assessments is, you know, you have to click the boxes. And I think at the end of the day, I usually feel like I'm not sure they're going to interpret my clicks the way I was thinking about it, right?

And so what we're doing is we're opening it up where, besides that, the kind of forced choice data, which is still useful, we give employees an opportunity to just express themselves. It's like, okay, tell, you know, as you were, as you went through that last section, kind of talk to me about. What is it you were thinking about? What struck you as most important? What do you wish we, we were learning from what you were just saying? And this text is super valuable.

One of the ways that I saw how text was super valuable was I created a machine learning model that would take research papers and it would try to predict whether I would find them interesting. And so I had several hundred papers that I'd marked as interesting out of thousands and it worked okay. And then I used a part of an AI. So AIs don't understand words, so you have to take the words and convert them into basically a mathematical object called a vector.

And you might remember that from high school, a vector is a mathematical object that has a direction and a magnitude, right? And so you take words or you take sentences or even entire paragraphs and convert them into these mathematical things. With basically no effort on my part. I just took all the titles and abstracts from all these papers and I used the first part of the AI to create a vector and now I had more data and so I provided that to my same model and it improved by 14%.

And so that's a big jump. And likewise, if you could think about it, if we have an employee satisfaction survey, which is part of what we're working on, and you can understand their open ended responses, and it gives you 14 percent better understanding of your employees. That could be a really big deal,

Ask for information, maybe hit a database, hit an application, do a web search, right, that kind of stuff. And so what I think is going to happen is people are going to gain essentially executive assistance. Like, imagine everybody in the whole company has an executive assistant.

And that executive assistant manages all of the kind of overhead of being at work and simplifies things like when you get an email that's making a request, the agent reads the email, organizes the information you need to respond and says, here's some things to think about in the response. And maybe you could say something like this. But you, as a human, make the decisions, right?

You can look at the data and say, Okay, I want to describe it a different way, or there's a reason that we want to say something different, or, you know, the AI misunderstood it, whatever. I think those are going to be really powerful instruments. And we're going to see that at the Here's the funny thing. AIs are best at replacing vice presidents. Because if you think about a vice president's job, It's about information synthesis.

There's the entire organization that's gathering information, and that information bubbles up through the organization. And then you have this vice president who has to make kind of these strategic decisions based on this really wide view of the entire organization. And it turns out that type of information synthesis is where AIs are much stronger than humans. I think we start off with, with these kind of Microsoft calls them co pilots.

You might call them an executive assistant or an AI assistant. I think they're going to be really useful and powerful. And they're going to save you time and allow you to have a better understanding of the business. By the way, I think it's a mistake for an organization to say, okay, we gained this efficiency, right? So now let's look at computer programmers. If my computer programmers use AI to help them write their code, their efficiency goes up pretty dramatically.

And there's two ways you could kind of respond to that. One is you could say, whoa, okay, we don't need as many programmers to get the same amount of work done. The other way is to recognize that in every tech company I've ever worked at, There comes a time when you prioritize everything you want the programmers to do. And then you have to draw a line and say, well, we don't have the time or money to do anything below this line.

So you have to get rid of things that would benefit customers, would benefit the company. Well, if AI increases productivity, essentially what it does is it pushes that line down. So the smart companies are going to be the ones that are like, look, let's make everybody more efficient. Let's help them get their stuff done faster. Let's help them reduce the overhead of just being at work. And let's produce more for our customers. And those companies are going to accelerate past their competitors.

So the companies that choose to downsize and remove people and stay at the same level of work with fewer people are going to get crushed by the companies that keep the same number of people, but dramatically improve the amount of work that you get done and the amount of things they can do for the customers.

And help them understand this is a machine learning process that you could use to make your life better. This is how you can apply AI. So I'd love for people to go on Calendly and schedule some time with me.

Here's the process that you're going to go through to do basically the whole thing. I mean, that's really the biggest thing that. That has changed. It's sort of changed the nature of content creation.

I wouldn't say in some ways, I wouldn't say it's easier because we're still in like that learning phase of figuring out how to, how to like get the most out of it and what you would actually want it to say, or particularly with your audience profile, depending on what it is, getting it to match that. I cannot believe how far it's come since the first time I typed something into chat. I mean, It's remarkable, isn't it? It's changed