Ep 83 - Is algorithmic pay the future?

[00:00:00] And we're live! Welcome to another episode of Comp and Coffee. Very excited to be joined by a bunch of amazing folks today. Ruth, as always, but also Sarah and Lulu are joining us as well from the Payscale team.

[00:00:33] And we're going to jump right in here in just a minute. But you know, Ruth was pointing me to this article she was reading. And I have to say, I started watching some of the videos related to it and my youngest daughter overheard and she's like, what really?

[00:00:47] And she started watching it too. And we were all fascinated by this. And the article is the premise of this article is two brothers, right? We're watching both either. Either they, I wasn't sure if they themselves drive for Uber.

[00:01:00] I think they do or they drive ride shares, right? And they were kind of testing using their ride share apps as drivers to see if there was a difference in route revenue they were going to make. And they're obviously brothers.

[00:01:17] So most of that information is probably the same. Now the video I watched, and they were getting different numbers, right? So brother A would get like an $8 fare and brother B would get like a $7 fare or a $6

[00:01:28] fare for the same destination at the same time, which is sort of really interesting. And now there was some nuances to it. So sometimes I try to come at these things as a little bit of a skeptic.

[00:01:39] One of the things I noticed was that one of the drivers was driving a, happened to be a Tesla and it was a Tesla they owned. And the other driver was driving a hybrid and it was a hybrid they were renting.

[00:01:49] I don't know if that factors into the algorithm, but it kind of comes back to, you know, Ruth started reminiscing about this idea or ruminating on this idea that algorithmic pay, right? Is it, this is all set up by an algorithm and how does the algorithm work?

[00:02:03] And do we have any visibility into the work and the algorithm? How does it all come together? And how do we know it's being equitable, right? Like if it is based on whether you own a rent that has certain implications for people from certain economic status, right?

[00:02:15] So what's happening there, what's going on there and is this an indicator of some broader problems we could have if we move to a compensation world where a lot of things are figured out automagically by computers. So excited again by the guests we have today.

[00:02:29] And I wanted to start things off here with Sarah. You know, you are, you're our director of data science here. You're all things data. And I'd love to get your insight into like when someone says algorithmic pay to you, like

[00:02:41] what does that mean and does it exist today? Or is that kind of stuff actually happening now? Sure. Yeah. Algorithmic pay describes making pay decisions based on a mathematical combination of compensable factors. There's a real spectrum there, right?

[00:02:56] Classic examples that are very well accepted within our field are adjusting pay ranges or determining pay ranges based on an incumbency or some experience or where they're located. Sometimes we pay people more if they have more experience coming into the job and less

[00:03:13] if they have less experience coming into the job. In the case of Uber, at least from the popular media, neither the compensable factors or those input features or the methodology are transparent.

[00:03:26] So it's hard to tell whether all of the features they're using are things that we would consider fair features to base compensation on. But the term algorithmic pay on its own doesn't mean it isn't fair. It just means that you're combining these different features in some mathematical way

[00:03:43] to end up with a pay range. Got it. Do you feel like there's, do you think people would define some of the things we do here at PayScale or maybe comp professionals do in general as algorithmic pay? Sure. Absolutely. Absolutely.

[00:03:59] The example I like to use here is maybe you're hiring a nurse in flat Alaska and there are no other nurses in flat Alaska yet. The first nurse that you're hiring for this town, there's not a population of nurses to

[00:04:12] determine what the pay is for that new hire in that location. And so to get a fair pay range, you have to use an algorithm roughly. And we use that in a couple of places, an HRMA, for instance, if you ask for a pay range

[00:04:27] for a nurse in flat Alaska, we're going to take the national median pay for a nurse overall, the whole country, and then look at the difference in pay for most jobs when they move from

[00:04:39] the national level to flat Alaska and make an adjustment there so that you're getting a localized pay range without having access to an underlying data pool there. So anytime we're extrapolating, and HRMA is the primary place where we do this, but

[00:04:56] lots of comp analysts and professionals do too, anytime you're extrapolating from a data set that isn't quite the right fit or triangulating, that is considered algorithmic pay. In the employee reported data set, we try to break pay down by the kinds of compensable

[00:05:13] factors that affect it, the years of experience, the skills that an employee is bringing to the table, the industry that they're working in to allow our clients to build back up using those building blocks and come up with a consistent and transparent algorithm that

[00:05:28] describes how a person is paid for the job that they're doing and the value that they're bringing to the company. So we consider those algorithms and that transparent, explainable compensation strategy to be a best practice.

[00:05:42] It's a much easier thing to explain to your employees than a less algorithmic approach where maybe your comp analyst is being more artistic and coming up with an individual pay range for each job, but the way that they're doing that and the data sources they're using

[00:05:58] for each of the jobs is a little bit bespoke and not consistent between the jobs. We think that if you can use one algorithm to describe how everybody within the company is paid and how those decisions are made, that you're being more fair overall. Oh, great.

[00:06:14] Thanks for that detail. Really helpful. I think you jumped into the point you made about being transparent is really part of the magic. And I think that's one of the challenges we see in this Uber example. Ruth, I know you found this thing.

[00:06:27] Thanks so much for sharing it because it really was a fascinating video for folks to check out. Did they ever manage, Ruth, to uncover sort of any hypotheses around why the rates were different in some of those things they were doing? Was there any transparency like Sarah describes?

[00:06:40] I think Uber gave an official response which told everything that they didn't include in it as a compensable factor. So, Sarah just mentioned compensable factors. So, they said they don't personalize fares to drivers based on race, ethnicity, acceptance rate, total earnings or prior trip history.

[00:06:57] So, they were not considered when calculating fares, but they didn't really say what was. So, Russ, you mentioned at the beginning potentially the one difference that you could identify from some of the research you did was whether who owned a car and who didn't.

[00:07:11] But I think we are moving to a point, and I know we're going to discuss it a bit more later, where along with the general drive for pay transparency, there are greater calls for transparency into the models that are being used and how we're coming up with the

[00:07:25] results that we're doing. Yeah, I think that feels like such a core opportunity for us as we think more about algorithms and then even if you extend this to generative AI and everything else, we need to be able to show our work, right?

[00:07:38] I think that sounds like it. Lulu, I'd love to get your thoughts on some of this. Hearing the lack of transparency, hearing some of the things that are implied by some of this, what are some of the things that you as a lawyer who specializes in the realm

[00:07:56] of compensation, what do you feel like gets you nervous about this? A lot of things get me really nervous about it, but it's also kind of exciting, I guess, and the generation and the world that we're living in.

[00:08:07] But some of these firms have found in their own research that these practices can actually lead to women earning less than men. Shocking, computers are still discriminating based on gender. They also saw in their algorithms and top-down acquisition, so recruiting and things like

[00:08:24] that, if you just dump your resume into a job-vite or something like that, that those algorithms can actually screen out women's CVs and resumes over men, which is a huge problem. At the end of the day, it's going to be a concern if these algorithms are recreating

[00:08:42] traditional wage differences based on gender alone or even based on race or any other protected class. These computers cannot take into consideration, at least right now, they cannot take into consideration non-discriminatory compensable factors like years of experience or education

[00:08:59] level or location or any other tangible non-discriminatory things that you can bring to a company that could actually increase your pay and get you the job a lot more easily. So if that's the case, if the computers are screening out applicants based on gender or

[00:09:14] race or any other protected factor, then it actually could be seen as per se illegal under a million employment laws. So when somebody wants to bring a lawsuit under employee discrimination, things like that, all they really need to show is that they were paid differently or they weren't

[00:09:32] offered the job based on a protected class alone. And when we talk about protected classes, we're talking about gender, race, religion, ethnicity, sexual orientation, disability, all the gamut. That's all they have to prove. And then what happens with organizations?

[00:09:47] Well, then they can win sizable sums from employers who solely rely on computers to make that kind of decision. And I'm going to say this probably more than once during the podcast today.

[00:09:58] This is a huge reason why it is still going to be so important for HR folks, real humans that blink, breathe, gesture. Y'all can't see me gesturing, but I'm gesturing why it's going to be so important to still have these human checks on what decisions computers are making.

[00:10:18] Kind of hopping across the pond a little bit, the use of algorithmic pay has also led to some employees asking for more information in Europe under the GDPR. That is the General Data Protection Regulation. After litigating and filing lawsuits against their organization, some workers have

[00:10:34] actually won the right to have access to what data companies are actually taking and extracting from their work to determine particular pricing. We don't have we haven't looked behind the curtain on that yet. That hasn't been revealed quite yet, but stay tuned.

[00:10:51] And then on the US side, there are similar privacy laws like in states like California. We haven't seen litigation around this yet, but we have been seeing attempts from regulators to try and get this information.

[00:11:04] But it's been met with resistance from companies, which should not be a surprise to anybody. Companies are saying to regulators, you know, this is about privacy. We don't want to disclose our practices because it might lead to information about workers being leaked and revealed.

[00:11:21] But they also maintain, which I think is real. I'm not an IP lawyer, intellectual property lawyer, but this is a fascinating argument as well, is that if they built these systems in-house, they're claiming like, hey, federal

[00:11:32] regulators, we're not going to give you access to this because this is our personal intellectual property. But when we see these lawsuits start popping up, if we see these lawsuits start popping up, employers can't hide behind these arguments anymore because

[00:11:46] during litigation, they're unlikely to keep this information behind closed doors. Wow. And so much great information there. We're going to include some links in the show notes to some of these laws. Lulu referenced, thank you so much, Lulu.

[00:12:00] This whole thing sort of really seems at odds around the pay transparency movement, right? If you're if you're being paid based on an algorithm that is completely insulated, as well, you just described this whole problem with intellectual property. And so it's my algorithm.

[00:12:16] I'll go. So no, you can't see it. Seems like it's sort of a big impediment to that. Ruth, what's your take on that? Yeah, I mean, I think we've talked a lot about pay transparency here at PayScale and

[00:12:26] we are seeing this push to really get to the point where we can understand the compensable factors that we are using to drive pay, because I don't think we've intentionally always focused on those. We've kind of maybe we've geared towards the market rate.

[00:12:41] Maybe we've thought that we're applying for a level of experience, but we didn't really have the data to support that. And now we're having to not only post ranges in pay ranges in job adverts, but also

[00:12:53] get to the point where we're having to communicate with employees more openly about how their pay has been determined. We are going to need to be able to explain this is why you, employee A, are at this

[00:13:05] point of the PayScale and this is why you, employee B, are at this point of the PayScale. It's not always an easy discussion to have. There's also a whole load of kind of like history that pay history decisions that mess up a lot of those decisions that happen.

[00:13:21] But I think we are moving more closer to a world where we will have more defined compensable factors. And there has been talk in the equity world as well about having more defined formulas, again, which is essentially a manual algorithm where you

[00:13:37] will say if you've got this number of years experience and you're a new employer, you've got this experience and you're in this location and you've got this particular skill, we're going to put you at this point on the pay range.

[00:13:49] And that's essentially, as you said, Sarah, just another algorithm, isn't it? So, I mean, I'm optimistic of algorithms when it comes to the topic of pay transparency and pay equity, because I do genuinely believe that using technology to

[00:14:06] make decisions at scale is important and also to ultimately help subjectivity in making compensation decisions, because there's an awful lot of subjectivity. I worked in investment banking running compensation for about 15 years, and I can tell

[00:14:22] you a lot about subjectivity in pay decisions from my time living that life. So, there is so much subjectivity. So, at least if we can get to a base level where the basic factors are kind of being

[00:14:34] taken into account when you get to a pay decision and then, yes, a manager who's making a pay decision then has some variability to decide just exactly maybe there was something in that person's history that wasn't captured in a compensable factor.

[00:14:47] So, we might need to adjust or make this decision. I think we're getting to a good place. I totally agree. I think that the algorithms can give us a chance to reduce that subjectivity and that variability based on our own decision-making biases.

[00:15:01] But we have to be really, really cautious. Ruth, as you and Lulu both mentioned, much of those same biases are codified in the data that are used to train the algorithms. So, an algorithm that's trained on this bias dataset can perpetuate that bias unless it's

[00:15:17] transparent and unless we can correct for it. The example I like to use here, so in many jobs, men make more money than women. We know that. And most pay algorithms, no pay algorithms, can legally use an employee's gender to inform

[00:15:31] their pay. But there are other ways that an algorithm can learn that gender bias and encode it. So, for instance, if we added the compensable factor skill in hunting, like being good at shooting deer or something into an algorithmic pay model for a software

[00:15:45] engineer. For a software engineer, being able to shoot a deer, bag a deer, whatever you say, shouldn't impact the employee's performance in the role, shouldn't impact the value that they're bringing to the company, shouldn't impact the pay, I would say. And generally, more men than women hunt.

[00:16:02] Right. So if we built a compensation model that allowed that hunting skill as a compensable factor, the model may misattribute the underlying population gender pay gap to that skill. So now instead of predicting that men should make more than women, which is a

[00:16:17] really obvious bias that we wouldn't allow, the algorithm might suggest that a software engineer who hunts should be paid more than one who doesn't hunt. When models are transparent and we can inspect the biases of each of those compensable

[00:16:31] factors and the outcome, we can identify those issues and correct for them. When the model is super opaque or is so fancy that we can't trace the impact of each of the compensable factors, then we run the risk of introducing that undetected and

[00:16:46] uncorrected bias back into our models because they're there in the data that we're trading on. That's such a great point, Sarah, the idea of undetected bias. Right. And we can and the example you gave is such a good one about the

[00:17:03] hunter. And people might be like, well, no one would put a hunter skill into that job description. But the interesting thing is it also really depends when you start to train these models, especially if they're AI driven.

[00:17:15] And when you train them, all of a sudden the legacy data, the data you train them on, if it turns out that historically, like back in the 80s and 90s, most highly paid software engineers happen to be hunters because they were mostly male.

[00:17:28] An algorithm might and especially if it's told to ignore gender, it might conclude that there's somehow some kind of correlation between hunting and being a good developer. And all of a sudden you'd have the situation Sarah just described.

[00:17:38] And so as we think more and more about training AI driven models, and we talk a lot about this on the show, right? We had a big episode about chat GPT. A lot of software companies in our space are thinking about generative AI and how do they

[00:17:53] leverage it. And so is this something we should be concerned about there too, Sarah? Can you dive in a little bit about how we think about it and maybe some of the things we're doing to try to avoid some of that?

[00:18:02] Sure, sure. So chat GPT is a natural language model. One of the coolest things about it is that in addition to understanding our English and being able to generate English, it can also speak SQL and Python and React, Java and other computer and human languages.

[00:18:21] The big advantage I see to chat GPT and the benefit I think we're going to see over the next couple of years is that it's going to make it really easy to query your data and your market data and all of the data that's available with natural language.

[00:18:34] It'll make it easy to plot and visualize and interact with that data. I think it's going to grant everyone the keys to the world as if they were database experts and programmers and data visualizers and superpower Tableau users and really

[00:18:46] good at Excel without having to learn the specific languages for each of those tools. We'll be able to interact with that data to query and plot and write code in basically any programming language. Chat GPT and other large language models are not good and not

[00:19:03] trained to be good at generating numbers like you would want to do for a compensation strategy. I think it will hallucinate numbers and you can see some good examples there if

[00:19:14] you search the news, but I think it will be added to many tools in the HR space and elsewhere, but it won't be the back end algorithm that is generating the pay ranges. It'll be a way to interact more naturally with that algorithm and the underlying data.

[00:19:29] However, I will say that large language models like chat GPT are a subset of a bigger field called deep learning, and we're seeing the same kind of leapfrog advances in the non-language applications of deep learning.

[00:19:44] Deep learning is an old math concept that is now tractable thanks to super fast computers, and it allows us to build more and more complex models which can often explain the world better and be more accurate.

[00:19:57] But the cost is that they're harder to understand and they're not as transparent as some of the simpler models. So, for instance, in a classic algorithm like one you might be using now, salary could be

[00:20:09] modeled or determined based on as a function of a few input features, maybe the job title, the location that the job is in and the years of experience that the incumbent is bringing. Each of those values is contributing to the expected pay for the job.

[00:20:26] With a deep learning model, we can model this with tons more complexity. So instead of having location affect the job in a single way, the model learns that location in combinations with the other features like location and skill together or

[00:20:38] location and industry together or location and skill and industry together impact pay. And it can add in these tens or hundreds of combinations of each of the factors. So in general, we can describe the world better that way because it's a lot of levers to

[00:20:52] move and to learn and come up with a really, really rich description of why someone's paid the way they are. But then it's hard and sometimes impossible to say exactly why the pay range is predicted the way it is or recommend or we're recommending that pay range

[00:21:08] to a new client. So it's more likely that a model like that could use a feature like that hunting skill in a way that biases the output, but it's harder to tell than it is in a simple model.

[00:21:21] So I think it's possible in this case that Uber is using a complex deep learning model under the hood that could be encoding some of that bias. Maybe it's the rental versus owning car that is separating on something that we wouldn't

[00:21:36] actually consider fair, but it's buried so many layers deep into the model that we can't see the impact of that on that final pay range. So as this field matures and I'd love to hear Lulu's thoughts, but I expect that

[00:21:48] compensation modeling will be similar to other kind of sensitive mathematical models in our society, like resume screening, for instance, is more regulated. Credit score models are fairly well regulated. What that will the society and our legal system will demand transparency and demand that kind of bias checking.

[00:22:11] So we're certainly on the viewpoint that these models have to remain inspectable and testable for bias at every step of the way so that we can make sure that we're not codifying the biases from the underlying data through components that we can't see and can't fix.

[00:22:33] Yeah, I think just listening, wow, so that was such great information in there and thinking through everything you were just describing, it seems to me that companies like Uber may not even know what's actually happening inside that deep learning model.

[00:22:47] And so Lulu, that just seems to me to like scream all kinds of potential legal stuff. You know, what's your what's your take on the whole has ChattGPT rolls out into these different aspects and we get more generative AI.

[00:22:59] What do you think is going to start to happen from a legal perspective? Well, personally, when ChattGPT first came out, I thought it was cool. I went in and I asked if it knew who Lulu Cycli was.

[00:23:08] And the only thing it got right was that I am a lawyer, a Lebanese-American lawyer based in Dallas. But ChattGPT also thought that I was like a singer, songwriter and an artist that presents my work in West Texas. Y'all don't want to see me at Carrie.

[00:23:23] Do you have any Oscars? I mean, y'all can get my autograph next time you see me because I'm apparently very famous, according to ChattGPT. But it got a lot of stuff wrong. But I did struggle because Lulu as a member of society, Lulu versus Lulu as the

[00:23:39] employment lawyer, were at odds with one another because as an employment lawyer, my employment lawyer brain said, OK, wait a second. Isn't this just another way, another avenue to perpetuate discrimination? Are organizations going to start asserting the shaggy, it wasn't me defense?

[00:23:58] Because they're going to say it was the computer who made the decision and discriminated. It wasn't me, I promise. So that made me nervous. But I also wasn't the only, I guess, employment legal mind to think about this because a few

[00:24:13] months ago, officials from the Equal Employment Opportunity Commission, the EEOC, they're the federal organization, the federal department that oversees employment discrimination in conjunction with the Department of Justice and the Consumer Financial Protection Bureau and the Federal Trade Commission.

[00:24:31] They all came out with a joint statement basically saying companies are put on notice federal discrimination laws are going to apply if you use AI to make decisions or chad GPT to make decisions or any sort of algorithms to make decisions.

[00:24:46] And then they went on to say the step further was actually something I was really interested in is that these organizations are going to keep your organization, your organizations accountable for discriminatory decisions, even if those decisions were made by a computer.

[00:25:02] So I promised you I was going to say this more than once, which is it's still very it's going to be very important to keep your HR professionals on staff, making sure there are still human checks over any decisions a computer is making in your organization.

[00:25:18] So that is going to be because I know a lot of employees are really worried, is AI going to replace my job? I think a lot of HR professionals are worried about that, but it's going to be essential

[00:25:28] to keep them on staff because of these biases that Sarah was already talking about was like these computers are going to have initial biases. And until we get those biases out, which we may never we may never be able to, we still need to have the human check.

[00:25:45] Absolutely. As a person on the call, I will say I'm not coming for your job. It would be your partner. Right. We want to make your job easier, not getting rid of humans in the loop anytime soon in this field. It's incredibly important to keep that judgment.

[00:25:58] Yeah, I absolutely love such a couple of good points there. I mean, Lulu, great point on the human interaction has to still be there. And Sarah, you just nailed it again there. But also earlier you mentioned that AI, these are tools to empower humans to get the

[00:26:14] information they need to still make the decision. Right. I think that's so critical. And we were, as Sarah mentioned, we're looking at a lot of different opportunities here at PayScale to use generative AI.

[00:26:22] I was literally just on a call yesterday with our legal team as we think about how to roll this technology out in a responsible way, really looking at a lot of the laws that Lulu is talking about that are coming from governments around the world around this.

[00:26:35] And it really comes down to we have to make sure that wherever we use this tool, we have to be explicit, right? Help the user understand where we're using it and always give the option for the user

[00:26:45] to to see as transparently as possible what's actually happening and to make the final call. That is so important because as you guys just did a great job pointing out, there's so many biases that can come into this, whether it's based on legacy data, whether it's

[00:26:59] just over time, it relearns it for whatever reason, like the vehicle ownership, perhaps example. So so much great stuff there. Thank you all for all this great inputs and the great discussion and so much great knowledge today. That's what we think about what's happening around

[00:27:16] here with with algorithmic pay and generative AI and how this all might come together. What do you guys think? Let us know. Send us an email to coffee at Payscale dot com or reach out to us on Twitter over at

[00:27:28] Payscale. Thanks again for listening to another comp and coffee. Comp and Coffee is a Payscale production dedicated to the compensation community. We welcome your feedback. Send email to coffee at Payscale dot com, tweet us at Payscale or share your thoughts in our Payscale Connect community.

[00:28:00] Until next time, keep your coffee hot and your data fresh.