Artificial intelligence (AI) technologies are not new, but they have made rapid advances in recent years and attracted the attention of policymakers and observers from all points on the political spectrum. These advances have intensified concerns about AI’s potential to discriminate against select groups of Americans or to import human bias against particular ideologies.
AI programs like ChatGPT learn from internet content and are liable to present opinions – specifically dominant cultural opinions – as facts. Is it inevitable that these programs will acquire and reproduce the discriminatory ideas or biases we see in humans? Because AI learns by detecting patterns in real world data, are disparate impacts unavoidable in AI systems used for hiring, lending decisions, or bail determinations? If so, how does this compare to the bias of human decision-making unaided by AI?
Increasingly, laws and regulations are being proposed to address these bias concerns. But do we need new laws or are the anti-discrimination laws that already govern human decision-makers sufficient? Please join us as an expert panel discusses these questions and more.
Curt Levey, President, Committee for Justice
Keith Sonderling, Commissioner, Equal Employment Opportunity Commission
Dr. Gary Marcus, Professor Emeritus, New York University & Founder, Geometric Intelligence
[Moderator] Ken Marcus, Founder and Chairman, Louis D. Brandeis Center for Human Rights Under Law
To register, click the link above.
As always, the Federalist Society takes no position on particular legal or public policy issues; all expressions of opinion are those of the speaker.
Sam Fendler: Hello everyone, and welcome to this Federalist Society virtual event. My name is Sam Fendler, and I'm an Assistant Director of Practice Groups with The Federalist Society. Today, we're excited to host "Artificial Intelligence, Anti-Discrimination & Bias." Our guests today are EEOC Commissioner Keith Sonderling, Dr. Gary Marcus, and Kurt Levey. Our moderator today is Ken Marcus. Ken is the founder and chairman of the Brandeis Center for Human Rights Under Law and former Assistant U.S. Secretary of Education for Civil Rights. Ken also serves as the Executive Committee Chairman of The Federalist Society's Civil Rights Practice Group.
If you'd like to learn more about today's speakers, their full bios can be viewed on our website, fedsoc.org. After our speakers give their opening remarks, we will turn to you, the audience, for questions. If you have a question, please enter it into the Q&A function at the bottom of your Zoom window, and we'll do our best to answer as many questions as we can. Finally, I'll note that, as always, all expressions of opinion today are those of our guest speakers and not The Federalist Society. With that, Ken, thank you very much for joining us today. And the floor is yours.
Kenneth L. Marcus: Thank you, Sam. And welcome, everybody. We've got a great audience today. I see from the records that I have some friends and colleagues and some very distinguished lawyers and others who are joining us on the call. But it's our participants here who are really terrific. The topic today is on "Artificial Intelligence, Bias & Anti-Discrimination." Really an interesting topic that we haven't drilled into nearly as much as I would like. Lots of questions are coming from different sorts of people on the left and on the right.
Some people might have thought that with artificial intelligence we would no longer have to worry about different forms of bias or discrimination. And yet there have been arguments from people of different perspectives that, in fact, artificial intelligence may just bake different sorts of bias into the system, whether that bias might be a bias against racial or ethnic minorities or women, whether it's a bias against certain intellectual or political perspectives, or other sorts of bias that may be troublesome either from the point of view of accuracy and just general correctness or from the point of view of wanting to have an equitable or even-handed approach to the data.
We have some really interesting thinkers who are with us today who can help us sort this out. Some of you may be familiar with Kurt Levey, a member of our practice group, who has, as an attorney and as a distinguished attorney as president of the Committee for Justice, spoken on various legal topics. But he also has a background and two different degrees from Brown in artificial intelligence. So this is a specific topic of his expertise.
Commissioner Keith Sonderling is a sitting member of the Equal Employment Opportunity Commission and an expert on the legal and regulatory aspects. There's really no one who is better suited to discuss the sort of Washington angle, the regulatory angle on the employment and workplace issues that are raised by AI. He's an attorney, a legal background; he's also a professorial lecturer in law at George Washington University.
Dr. Gary Marcus, since this is partly on bias, I'm not sure whether I should say whether he and I are or are not related. We might take a survey later in the chatroom on what you think. At any rate, Dr. Gary Marcus, I'm glad to meet him. He is a distinguished psychologist and an expert on AI who has written prolifically. He is emeritus from NYU, published many books. And, on the issue of AI, he also has an entrepreneurial background as a founder of companies. But he is one of the most outspoken people on the subject. So, an excellent person to help us to see different sorts of perspectives on the issue.
We'll start with, perhaps, some opening thoughts by each of the speakers, starting with Kurt Levey.
Curt Levey: Thank you all for inviting me. As Ken mentioned, I come at this with two hats on: one from my previous career in AI, and the other from my current career as an attorney, in which I've done a lot of work on civil rights law. And both perspectives lead me to conclude that the problem of bias in AI is at least somewhat exaggerated. What's really interesting, I think, about this issue is that both the left and right are complaining about the danger of bias in AI, but they're focusing on different things.
The left is focused mainly on discrimination against various minority groups in any AI systems that make decisions about important things like hiring and lending, bail amounts, facial recognition. And the right is focused mainly on bias against conservative viewpoints and people in large language models like ChatGPT and other forms of AI-powered search. I'll start with the latter.
I think the study that most got my attention was from the Brookings Institution, which, certainly, we do not hear a lot of complaints about conservative bias coming from. And they gave prompts to ChatGPT to answer yes or no for whether it supported various political statements: "Are you in favor of abortion? Are you against abortion? Should we have more gun control or should we have less gun control," etc.
And for each of many social issues where ChatGPT was consistent — and it wasn't always — it supported the liberal position and rejected the conservative position. There's also a German academic study that had similar results. And then various reporters have reported that ChatGPT refused to write poems to commemorate Donald Trump or Ted Cruz but were happy to do so for Fidel Castro and Representative Ilhan Omar. And then Politico's tech reporter couldn't get ChatGPT to write a legal brief in favor of overturning the Supreme Court's gay marriage decision.
But all that doesn't necessarily mean that these systems are designed, that large language models are designed to be biased. So, what could be the source of the bias? Could it be the scientists that are building these models who make choices about what data to use? I suspect that's not a major source of bias. But then, one of the steps in building these models is reinforcement learning with human feedback, which OpenAI, which creates ChatGPT, says helps to align the model's output with human values, whatever that means. And there's a lot of room for bias there when we're talking about human values. And probably all the more so because OpenAI is based in San Francisco.
Then, it could be, the bias could come from the platforms. Microsoft, Google — now Meta, with its large language model — because they designed the interface between the user and the model. And I think it's fair to ask if there's bias at that stage, given that there's documented, if often exaggerated, anti-conservative bias in search results and content moderation. And it reminds me of Elon Musk's comment that these companies are teaching AI to lie. That may go too far, but they're certainly trying to design the interface, if nothing else, to make sure that it doesn't say things that are politically incorrect.
And the Obergefell brief may well be an example of platform censorship. I don't know if that was actually built into the actual model. But I suspect the biggest source of bias is likely just the collective bias of the contents of the internet, because that's mostly what these are trained on: about 85 percent internet, about 15 percent books. The cultural elites in the media and academia contribute much more internet comments than the average American. And among those elites, Castro is more popular than Donald Trump or Ted Cruz, whether you like that or not.
So, again, this may be just reflecting the reality of what's out there. I think there's probably also a recency factor, a sort of pro-wokeness bias that we've seen in these models. It's probably due to the fact that there's just a lot more on the internet from 2023 than there is from, say, 2013. So these models learn that virtually everyone agrees, say, about how wonderful the gay marriage decision was.
Let me turn — because our time is limited — to the other type of bias that the left worries about, which is bias in decision-making, hiring, lending, etc. And, as with these large language models, I think the bias could stem from bias in data. Unlike the contents of the internet, which no one can control, I would be less forgiving if there is bias in the data, because model-builders should be able to create data sets that are unbiased. For example, if it's a lending model, you should be able to get all of the banks' loans. And that should accurately reflect the results of those loans.
If you're careful like that, AI, I think will actually tend to reduce the amount of discrimination in decision-making, because the alternative is reliance on human decision-making and people come to decisions with a lifetime of irrational prejudices. AI models at least start out as a blank slate without any biases. Everything they know, they learn from the data, consisting of examples where the output — say, whether the loan was paid back — is known. And the model's only goal is to minimize its error rate for predicting, say, loan defaults. There's no room for irrational prejudices to slip in. It's also a lot easier to hide race and gender from an AI model than it is from a human.
So, AI models don't have irrational bias or intentional discrimination. But that doesn't mean they can't have disparate impact. They may discover correlations that are real but have a disparate impact on one identity group or another. And so, the question comes, should relying on that correlation be unlawful? Some would say, "Well, those disparate impacts are basically just correlations that people find uncomfortable." Other people would point to civil rights laws in which, certainly under some circumstances, disparate impacts can be unlawful.
But since these models are using the correlations to increase their accuracy, you will sacrifice some accuracy if you prohibit certain correlations, for example, a parole model. And there are ones out there that try to help judges decide about bail and parole and things like that. It might learn that males are more likely to be repeat offenders, and thus it would improve its performance by taking gender into account. But you'd make the parole model more error-prone if you took away its ability to use that correlation.
And, again, to be fair, these models maximize their accuracy on the training data they're given. So disparate impact or any other bad performance may be due to bad data. We have to always take that into account. And, lastly, there's a plethora of AI laws and regulatory policies that are being proposed, including to address bias against minorities. And a lot of these proposals include having to test and certify that your AI system has no discriminatory effect. But, again, what does that mean? Does that mean you can't consider the history of violence in a bail or parole model if it results in women getting a better score than men?
My view is that we shouldn't reinvent the wheel, that AI bias can and is being addressed with the same anti-discrimination laws that govern human decision-makers and various non-AI tools that humans use. For example, hiring exams are a tool that has been around forever and are subject to civil rights laws. And the fact that you use an AI-based hiring tool doesn't provide an exemption under any current anti-discrimination law.
So, for now, at least, AI is just another tool. And every new tool does not require a new law. We may need new laws when AI can produce human-like autonomous decision-making that we would call artificial general intelligence. But that's many decades away, at best. And, besides, I don't think politicians and bureaucrats understand AI and its risks nearly well enough yet to effectively regulate. And I'll stop there.
Kenneth L. Marcus: Well, thank you, Curt Levey, both for those remarks, and also for mentioning the workplace and hiring late in your comments, which is a beautiful segue as we are now about to hear from Commissioner Keith Sonderling.
Keith Sonderling: Well, thank you so much. And so much to discuss from your statement there that we'll dive into. But I want to talk now about how AI is being used specifically in employment. And we'll touch on a lot of those concepts and also leave open for discussion some of the regulatory aspects that I know we are going to dive into.
So, we all know now that there's countless new AI technologies on the markets to help employers make employment decisions and all other types of decisions more efficiently, economically, and effectively. But when it comes to using AI in HR, it's different than the use in other types of businesses like reviewing documents, making deliveries faster, making widgets better. AI in HR intersects with some of the most fundamental civil rights we have in this country, which is the ability to enter and thrive in the workforce and provide for your family without being discriminated against because of your race, religion, national origin, sex, color, disability, age, and so on.
Many of the vendors in this space specifically, using it in human resources, promise to advance civil rights protections in the workplace by mitigating human involvement. And if you mitigate human involvement, therefore you're going to mitigate human bias. I believe that, carefully designed and properly used, AI does have the potential to advance civil rights protections in the workplace by mitigating the risk of unlawful discrimination.
But, at the same time, if it's poorly designed and carelessly implemented, AI can discriminate on a scale and magnitude far greater than any individual HR professional. The reason I'm talking about AI, the reason the EEOC and other federal government agencies are looking at AI in the workplace, t's already been involved in many stages of the job lifecycle for employees for years.
There's AI that writes job descriptions, screens resumés, chats with applicants, conducts job interviews. There's AI that predicts if an employee will accept an offer, and how much. There's AI that identifies a current employee or potential employee's current skills or future skills. There's AI that tracks productivity. There's AI that assesses worker sentiment on a day to see if they're happy or if they're going to quit. There's AI that does all the performance reviews for employees. And if an employee falls short of expectations, nowadays there's AI algorithms that may even send them a message saying they're fired.
And this isn't my forward-looking prediction. For each and every one of these tasks, you can find a commercially available product ready for purchase right now. So, given the current state of the market and just AI's implementation and adaptation across the board, deciding whether to adopt AI in HR is no longer a hard choice. Deciding which technology to adopt and for what purpose and how to comply with longstanding civil rights laws is a very hard choice indeed.
And this is really the challenge for all of us who are listening today, and for this panel, who care about how AI will impact civil rights in the workplace. How do we define those terms, saying, "What is carefully designed? What is properly implemented, in terminology that lawyers and policymakers can understand, without having PhDs in computer science or really understanding the underlying AI technology?" An AI-enabled program that conducts preliminary screening interviews, for instance, can be engineered to disregard factors such as age, sex, race, disability, pregnancy. It can even disregard variables that may even suggest a candidate's membership in a protected class, including foreign or regional accents or speech impairments.
And, in some cases, replacing an interviewer with a bot completely eliminates the opportunity for intentional discrimination at the screening phase of the hiring process. When you come to an interview, or Zoom turns on for an interview, you see a lot of things about that person you're interviewing. And it's not unheard of for an interviewer to meet a highly qualified candidate who's visibly pregnant, disabled, or religiously observant, and think to themselves, "This person is going to cost me. Accommodations will cost me. Leave will cost me. Health care will cost me. It just isn't worth it. So I'll go with someone else with similar skills who won't make these requests."
And although this is a very highly illegal example, it's one of many instances of bias that AI, properly designed, can mitigate from the outset. And those individuals who were not able enter the workforce because of those biases can then move to the next step. But for all the good, there's potential harm. So, at the same time, again, if it's poorly designed and carelessly implemented, it can discriminate on a scale and magnitude greater than one individual person.
And, like everything else, a bad HR decision impacting civil rights in the workplace using AI to do that is going to damage the trust of the organization, can create a toxic work culture, and ultimately damage the profitability, all while increasing the risk of litigation. Because AI is only as good as its purposeful and insightful application by informed organizations with an eye on the actual impact in compliance with the law.
And purposeful application of AI means careful consideration, from everything as you just heard before about the quality of data being used to the continuous monitoring of the algorithm after it has been deployed. That's because AI is just making predictions, in my case, here, what we oversee about applicants and their skills. It's only as good as the training data on which the AI is relying. So, in simplest terms, an algorithm that relies on the characteristics of a company's current workforce to monitor the attributes of the ideal job applicant may unintentionally replicate that status quo.
If the current workforce is made up primarily of employees of one race, one gender, or one age group, the algorithm then is automatically going to screen out applicants who do not share those same characteristics. And there's a legendary example, at this point, that one of these companies went to an AI recruiting program that promised diversity, equity, and inclusion, used all the buzzwords. And a resume screening company went through and used machine learning and AI to look at what makes employees successful at this one company, and it came back and said the most likely indicators of success at your company is being named Jared and having played high school lacrosse. And this was a program meant to diversify the workforce.
So, as an attorney, I've dedicated my career to labor and employment law. I do want to see AI reach its full potential. As an EEOC commissioner, I'm committed to helping workers and employers understand their rights and obligations when it comes to these technologies, because our mission here is to prevent and remedy unlawful employment discrimination and advance equal opportunity for all in the workforce. So, many of the laws that we enforce predate by over half a century the AI technologies we're discussing.
Nevertheless, federal anti-discrimination law is as applicable to employers who use AI to make employment decisions as they are to employers who exclusively rely on the brains of HR professionals. And, as you heard before — and we'll talk about the legalities of this — what's different about labor and employment law than other laws — because you're dealing with civil rights protections — is that employers are generally going to be liable for discrimination, whether or not they intended to discriminate.
And that is an absolutely crucial thing to bear in mind as companies start going all in on AI HR technologies, because here the data is everything. Data makes a difference between a good hire and a bad hire, a good promotion, and a bad promotion. And what I'm trying to raise awareness of is the difference between a lawful and unlawful employment decision.
For example, like you heard before, again, if that training data skews heavily towards people of one race, sex, religion, or national origin, these protected characteristics may come to play an improper role in the algorithm's training predictions about members of that target audience. It may downgrade proxies for race or gender, such as the names of historically black colleges or the names of womens' sports teams. Again, not because the computer is intentionally targeting and discriminating against these people. It's because they simply were not represented in the training data which the AI was given.
And, outside the realm of science fiction, AI still does not have motives or intentions of its own. It only has those algorithms that help enable it to correlate data to make those predictions. And that, according to industry experts, is why employers are flocking to it when it comes to AI in HR, because its reliance on hard data creates a potential to eliminate discrimination by removing the human from that decision-making process.
And, look, if it's designed in a clear and explainable way, it does eliminate one of the biggest challenges to effective human resources management. That's the human. That's the human bias that can be baked into some of those decisions. Because when employment decisions are made with bias, whether intentionally or unintentionally, such decisions are difficult to prove, because the human mind is difficult to understand. And, absent communication — or, in our cases, subpoenas, depositions — we never really know what a person's true motives are. And that's where AI can really come in as a positive to correct for that black box problem.
It can not only mask for certain protected characteristics like race, sex, age, disability, it can mask for proxy terms. It can really help employers take that skills-based approach to hiring, using documented skills and characteristics, which employment decisions should be based on, all while stripping out human bias. But, again, it can replicate and amplify existing bias when it's not properly designed. And, in conclusion here, my fear is that the apparently objective nature of algorithmic decision-making can result in technological bias on the part of the user. That is an overreliance, if not blind trust, that the AI is going to get it right.
And you can't lose sight that AI is self-reinforcing and requires close monitoring. Because inaccurate, incomplete, or unrepresentative data will only amplify, rather than minimize bias in employment decision-making. So, we can talk about it in our panel, but that's really where the testing, evaluation comes in, which is really, really important.
But, in conclusion, it presents, right now -- in my space, for HR professionals and businesses, it's finding the right division of labor between algorithms and HR personnel, between using AI to improve human decision-making and completely delegating decision-making entirely to an algorithm. Because, again, in my case, you're dealing with civil rights in the workplace. And while we're seeing AI continue to become mainstream technology in the workplace, discrimination by algorithms can't. So, thank you for having me on this panel. And I'm really looking forward to Dr. Marcus's comments and our discussion.
Kenneth L. Marcus: Thank you, Commissioner Sonderling. I noticed during your remarks that some members of our audience changed their screen names to Jared. I don't know why that is, but I wish them well in their pursuits. And now, we'll turn to Dr. Gary Marcus.
Dr. Gary Marcus: To clear up any uncertainty and resolve mysteries, I am not, so far as I know, related to Ken Marcus. But he does look, suspiciously, like my father. And I actually want to mention my father, my later father, in part, because he worked at the Maryland Human Rights Commission for almost a decade on discrimination law. And my first paid programming gig was working for my father when he was at the Maryland Human Rights Commission doing statistical analyses of the following sort: if you see four people on a panel and they're all male, what is the statistical probability of that happening by chance?
Or if you have 45 people working for the Maryland Bus Transportation Authority and there are no obese people, what's the probability of that, given the baseline? So, I've been thinking about these things since I was a little child. I'm not really entirely up on the bias in AI literature. I know some of it. But this runs deep in my blood. And I think my father would be deeply proud that I'm here on this panel with Commissioner Sonderling, whose remarks I loved. So, kind of a special moment for me.
One of the most important words in Commissioner Sonderling's remarks was the word "proxy," which he glossed over fairly quickly. But it's really the name of this game. The Jared stuff is an example. You have something that's actually standing, in some indirect way, for something else, like "rich white guy." We probably all saw that stuff about very, very, rich people getting into colleges at very high rates. It was in the New York Times just a couple days ago. These are proxy variables. Another one is ballet dancers. If you're a ballet dancer, it's very hard to get an AI system to not say that you're not fit for computer programming. "Ballet dancer" is historically associated with being female. And females have been less likely to go into computer programming or computer science.
And so, these systems will tend to use things like "ballet dancer" as a proxy, even if you try to correct them explicitly for gender bias. And so, for example, Amazon had something like a seven-year effort to make non-discriminatory job-hiring software that they eventually abandoned. And it was really because of this proxy variable issue, which nobody really has a great solution for. And something that Commissioner Sonderling said that's very important here is that employers are liable, whether or not intended.
And so, if you use this software and it gets stuck on these proxy variables, even if the manufacturers of the software were kind of full of good intentions, it's still a problem. It's still a problem in society. I think that Mr. Levey is correct that a lot of the bias does come from the data sets. But I think he's underestimating that problem in a couple of different ways. One is that the data sets themselves represent a history of bias that we don't want to perpetuate. That's why we pass these laws, is to not perpetuate that past history.
So, we have a history, of course, of slavery, redlining, and so forth, that gets led into the data set in various kinds of ways. So there's a very strong history of discrimination — a very unfortunate one — in our nation. And some of that gets put in the data. And we want to have machines that actually have the values that are represented in things like equal opportunity law but aren't really represented in current AI systems. And so, I think we need to really be aware of that.
A second issue there is that the systems are incredibly greedy, in terms of data. So, currently, the systems are, roughly speaking, using the entire internet to train on. And that means you can't afford to be choosy, or at least it's very difficult to be choosy. The smaller the data set, the worse the performance is. The larger the data set, the better it is. Even with internet-scale data, these systems still have a lot of problems. They make stuff up. They're inaccurate. They're unreliable. And so, everybody's pushing to have as much data as possible.
The consequence of that is they use a lot of garbage. So, they'll, for example, use data from fan fiction on Reddit. And it's probably not the best source for making an informed decision about who should be employed. But the way the systems are set up, that's the case. And then there's a third sort of hidden gotcha, which is, these things are not like databases. So, a database or spreadsheet, you can be like, "Well, what if I only look at people from this zip code, or only at this year or age or whatever?" You can do these kinds of what-if scenarios.
It's so expensive to retrain a large language model. It can be on the order of tens of millions of dollars. It can take the amount of energy that might be consumed by a number of cross-country trips on jumbo jets and so forth. So people don't constantly retrain. They don't do what-ifs in that format. So you're sort of stuck with, "This is GPT-4. It was trained in 2021, and we are going to use it. We can't retrain a version of it for some economical price without 'Let's just see what would happen if we dropped the fan fiction.'"
We can't do that. Maybe someday there will be some future technology that allows you to be much more selective about what data you use. But right now, it's one-size-fits-all. And because you need so much data, you wind up bringing a lot of discrimination along the way. Another super-important term that came up is "black box." Now I would like to talk about large language models, because they're the biggest black boxes and the most incorrigible black boxes that we know.
So there's a lot of existing software that works in employment. Commissioner Sonderling gave a beautiful summary of those. A lot of them use multiple regression where we can actually interpret the outputs. We can understand, what variables are they using to make their decisions? With a large language model, we simply can't do that. So, for example, some people may be taking in job applications, feeding them into ChatGPT and saying, "What do you think?" ChatGPT can't give you a documented trace. It is a black box vel non. It can't give you a trace of how it got to its decision.
And that actually brings me to my last thing. My understanding is people are, in fact, doing that; it seems like. Very problematic, from the employment law perspective. And I wonder what authority Commissioner Sonderling has to be able to subpoena the logs from OpenAI, for example, or from Microsoft, and say, "Are people using these for job discrimination?" So, the historical way that these things work — that I know of, through my father — is that somebody comes in and says, "I was fired," or "I didn't get the job. Seems to be discrimination here. I notice that everybody they hired is a white male, and I'm not. And this is not fair."
So they tend to be, let's say, plaintiff-centered, if that's the right way to put it. I would like to know, as an AI researcher, whether there is a lot of discriminatory use of OpenAI's products, of GPT-4, and I would like to know that without having a plaintiff. Maybe a plaintiff could come in and subpoena all that. But I think it would be really great if Commissioner Sonderling had that authority. If he doesn't, then that's an example, to me, of where we do need a new law, if the existing law doesn't cover that. And I'll just stop there.
Kenneth L. Marcus: Well, that's fantastic. So, in a few minutes, I'll give each of the speakers an opportunity to say a few more words, perhaps in response to what others said. But first, I'm going to abuse the privilege by asking some of my own questions. So, Curt Levey, I had a question similar to my newly adoptive son, Dr. Gary Marcus's, which is to say that when you indicated earlier in your remarks that if there is opinion bias it might simply be a function of what's out there.
And that wasn't necessarily reassuring to some of us. And so, I guess my question — and this may be similar to what yours is — should it be reassuring? Should it be non-reassuring? If there is this sort of bias, is it something that we should be worried about and perhaps do something about?
Curt Levey: Well, I didn't mean to imply, certainly, that it should be reassuring. I guess it's a little bit better than intentional bias. Intentional bias would worry me a little bit more. The question is how do you deal with it? Certainly, the people who are collecting the data could go out of their way to try to find more balance, perhaps rely on the internet less. They don't use literally everything on the internet anyway. So that would provide some room for bias. But I guess what I really think is that this problem of bias and large language models is downstream from basically the bias of cultural elites in our society.
And I think, often, when people attack the platforms or the big tech companies for having a "woke" bias, again, it's really downstream of a larger cultural issue. And I think conservatives are probably not going to be happy with any model that's built in a society where the cultural elites just have a lot more to say than the average American, where the people who like Donald Trump and Ted Cruz just don't have the cultural influence that other people have. And I guess that's not so much reassuring as it is a reply to people who would say that the platforms, themselves, are being biased.
Kenneth L. Marcus: Thank you. Commissioner Sonderling, I got this sense during your remarks that it is your view that bias in AI is a real and perhaps significant problem in hiring today. Was I overreading? It was hard to tell, at some points, what you --
Keith Sonderling: It can be. I mean, that's the difficulty. And, to get into Gary's question, I can answer both at the same time. The larger issue is that we don't know. The reason the EEOC exists is because there's bias in the workforce. That's why we've been around since 1960. That's why, in the last two years, we've collected a billion dollars from employers, both private sector and federal government, for violating these anti-discrimination laws. So, there is bias in employment decision-making, or my agency wouldn't be here to enforce these laws.
Specifically related to artificial intelligence, it's very difficult for us to tell right now, for a whole host of reasons. First and foremost is that employees do not generally have an awareness that they're being subject to an AI tool. So, you do an interview like this on Zoom, you have no idea what's running in the background. You generally don't know if your voice is being recorded and being analyzed on the patterns of how you're responding or what words you're using in what particular order.
There's some programs that use facial affect technology to see how many times you're blinking in front of the camera or if you're smiling, and rate your application based upon some of those characteristics, which may be or may not be indicative of if you can do the job, or may be discriminatory. So, because the EEOC operates on a charge-based system — meaning that employees have to know they were discriminated against or have a hunch they were discriminated against and file a charge with the EEOC, which then initiates our investigative authority — under very, very, limited circumstances can we start our own investigation.
So, if employees are not aware that they're being subject to this technology and that an algorithm —based on some of the statistical issues that are bad — made that decision, if they just get rejected by a job and say, "we went with a more qualified candidate," they may say, "Oh, I wasn't qualified," or "I didn't interview well." They have no idea that may have been a discriminatory algorithm or a person putting in their own bias and scaling discrimination through one of these programs.
So that lack of awareness prevents some of these cases, potentially, from coming in. And even if they do come in, it's not coming in as -- we don't have a cause of action saying, "AI discrimination." What is it? It's just discrimination. So it could be race discrimination. It could be national origin discrimination. And it takes time for us to investigate to see if it was based upon somebody intentionally discriminating: "I don't want to hire females." Or was it based upon an algorithm that lowered the scores of females because they were not included in that data set of what they were looking for? So, you can see it gets very complicated and can take us years down the road.
Now, to answer Gary's question about our subpoena authority and our authority just to open an investigation. We really don't have that authority to proactively go and test or ask some of these large tech companies or even anyone using the software what their data sets are in advance, to see if those data sets are going to cause that bias and cause those discriminatory results. So, the limit to our enforcement is really based upon employees knowing or believing they were discriminating and starting a federal investigation from there. But, even so, it's then only based upon those national origin discrimination, religious discrimination. We have to backtrack it through there.
Kenneth L. Marcus: Excellent. Thank you. Dr. Marcus, just because it's interesting, I wanted to explore with you a little bit more your comment about the New York Times article describing a recent report by Harvard-based researchers showing that the top one percent, economically, are doing much better in college admissions than the middle class or working class. The underlying report suggested that the bulk of this was attributable to legacy preferences, but that there were two other factors that they found that seemed to explain the rest. One of them was athletic preferences, and the other was personal characteristics. I imagine you might suggest that if there was that sort of bias, it would be in the personal characteristics part.
Dr. Gary Marcus: Well, we've already heard about lacrosse. Let's not forget that.
Kenneth L. Marcus: Good, yes. Are there other aspects of it that you see as part of it? Is it any sort of invidious preference to the sorts of things that rich people like that would be a common factor?
Dr. Gary Marcus: I bet there are. It's a little bit out of my expertise. And I think it's a subtle thing. One tends to get these proxies with variables in all sorts of places. But if you assume that a culture has kind of marks of high-brow versus low-brow, the high-brow stuff, like listening to opera, is probably, to some people, like, "Oh, well, they must be high-brow." And it's probably also a signifier of wealth. So, I would not be surprised if there's quite a lot of those. Some of them are pretty subtle. They're not seen on first inspection.
And the whole personal characteristic stuff for college admissions is going to get wild, in light of recent Supreme Court decisions. But I would not be surprised at all if there were indicators like that that maybe people don't even recognize at first, but that play a role.
Kenneth L. Marcus: Interesting. Why don't I go around one more time, in case any of you had any other brief comments to add? Curt Levey?
Curt Levey: I felt like there was some implication, in what my fellow speakers said, of intentional discrimination. For example, no doubt that Jared is a proxy for being white, but you still then have to ask, "Why does the system think that white people make better employees?" It's certainly not because the system has some irrational bias in favor of white people, like humans might very well have. It would only make that correlation between white and better performance as an employee if the data indicated that, on average, white people had performed better, or whatever reason, whether it's intrinsic ability or discrimination in the workplace, who knows, but only if the data indicated that white people performed better, if there was a real correlation there.
Again, it's like what I said about bail and parole models. Even if you were to eliminate gender, the system could probably still figure out that you were male. But if it recommended against parole more often for males, it wouldn't be because it hates males. It's because there's some correlation between being male and having a less favorable bail or parole outcome. There was also talk about the fact that the data reflects societal discrimination. Perhaps the data shows that black people don't perform as well on loans because black people have been denied opportunity in our society. The question is, should these models be dealing with that?
And that comes down to the societal discrimination versus direct discrimination debate that we just saw in the affirmative action cases in the Supreme Court where if schools are remedying their own discrimination, they can take race into account. But if what they're trying to do is remedy societal discrimination and say that, yes, certain minorities don't perform as well on test scores because of what happened long ago in our history, the Supreme Court's been pretty clear that you can't do that. And I think we have to do the same thing here. We can't ask these AI models to sort of cure all of society's ills. We just have to ask that they not, themselves, be discriminatory.
Dr. Gary Marcus: I think we can ask for more and should ask for more. It so happens that I met Condi Rice over the last few days, talked to her a bunch of times. And so, her example is firmly in my mind. She was the first, I think, African American woman national security advisor, and I think the first African American woman secretary of state. Prior to her appointments, if you had run any of this software, undoubtedly, on a bunch of the proxy variables — and, depending how they're set up, simply under non-proxy variables — would have said "That's not possible. You can't have a black female national security advisor." I'm sure she was a damn good one.
The systems, if they only look at past data, are going to have trouble with this. They do not represent the values that we want. And I will go back to saying that we make these laws because sometimes the practice — and Commissioner Sonderling made this point in a different way — we make these laws because sometimes the practice in our society is not where we want to be. The practice in our society means there's all kinds of bias in this data. And the laws are there because we have aspirations. These systems cannot reason about these aspirations. And so, we do have a problem.
Curt Levey: But I disagree that the system wouldn't have picked Condi Rice. She had incredible credentials. So, unless you went out of your way to program that system to discriminate against black people or women, it very much would have picked Condi Rice. Now, I understand your point [inaudible 00:45:57] --
Dr. Gary Marcus: Look just at the name variable. Like, there are not a lot of high-ranking U.S. employees named Condoleezza before Condoleezza Rice. I think you underestimate how hard it is to root this stuff out of these systems empirically.
Curt Levey: And you may be right. But it also scares me when people talk about putting values into it, because, again, we see what happens with those large language models and their reinforcement learning when we train them based on "promoting human values." Those human values have their own biases, including the cultural biases that I was just talking about. The human values of people in San Francisco, or the human values of the elite are not the same values of the average American.
Dr. Gary Marcus: There are very thorny questions there. But, in some cases, we have, ratified into our national laws, very clear answers to those. So, in particular, you may not discriminate on the basis of age, race, religion, etc., etc. So those values we do actually have agreement on, we have laws on. And the right software would actually respect them. But we don't know how to make that correct software yet, especially not in a large language model context. Multiple regressions systems, maybe. In the large language context, certainly not.
Kenneth L. Marcus: With less than ten minutes to go and a few questions in the queue, let me see if Keith Sonderling had anything he wanted to add.
Keith Sonderling: Yeah. I think it was briefly mentioned in some of the prior discussions. And with all the debate here in Washington D.C., with all the testimony in Congress, and do we need new AI laws? Do we need a new AI commission or a new government body? For me, I think, in a way, that's a distraction for companies who are already instituting and using these products and need guidance on how to comply with them. So, from my position here in the executive branch, all we can do is continue to enforce the laws in the books.
And you heard the laws here that we enforce are from the 1960s, but apply to employment decisions, apply to the results. And my agency is never going to be able to regulate the technology without additional funding and an additional mandate. We just are able to regulate and see the results. And, here, when we see employment decisions, we have results. And how did we get there? Was it somebody intentionally discriminating? Or was it because of a non-discriminatory factor that had the adverse impact of discriminating against certain groups?
So that's how we're left now, regardless, if the decisions are made by technology. But, in a sense, from a law enforcement purpose, this technology can make it even more transparent. So, for all these discussions we're having about was it the model that was discriminatory because you only had individuals with certain protected characteristics, or did you have the most perfect, diverse workforce in your data set and somebody programmed the skills to then intentionally pick out the factors that would discriminate against certain groups, hypothetically.
But, with all of that, there is a data set. There's cliques. If it's designed transparently, we would be able to see that, opposed to now, when we're dealing with depositions. Did you discriminate against that person? Nobody ever admits that. They say "no." So, rebuilding how that occurred, AI can make that process a lot more transparent in doing our job, from a law enforcement agency perspective.
And that's things that employers can be doing right now, with their own internal audits, using a lot of this longstanding existing framework to ultimately know you're making an employment decision. Was that employment decision biased, based upon the data, or the inputs, however you're looking at it? And there's nothing preventing employers from doing that now, before these tools are actually used at large scale.
Kenneth L. Marcus: Excellent. A few minutes left. The first question, I think, reflects very well on our excellent speakers. First question, "Where will the recording be available?" So, the recordings are typically provided on The Federalist Society website. I would invite Sam Fendler, if he cares to, to provide any additional details in the chat or in his closing remarks in a few minutes.
Dr. Gary Marcus: And I have to apologize. I have to go three minutes before the end, so I can only maybe take one question before I go.
Kenneth L. Marcus: Okay. So, I don't know if this would be for you or not. It is on the workplace. Will the recent Supreme Court decision on Harvard and North Carolina impact the ability of employers to use what the questioner calls "woke" criteria?
Dr. Gary Marcus: I don't know how to answer that. But I will pass and say that that's outside of my expertise, not doing college admissions or judging woke criteria. I'm sorry that I have to go because of another event. I sent advance notice, but it may not have percolated. I really enjoyed being part of this panel. Thank you all for having me.
Keith Sonderling: Thank you.
Kenneth L. Marcus: Pleasure having you. Now, I don't think that either of our other speakers will be able to dodge that question on those grounds, since I think it may be within their expertise. So, I don't know.
Keith Sonderling: Since Title VII was enacted, you are not allowed to make any kind of decisions -- just like you can't fire somebody because of race, you can't hire somebody because of race or national origin or religion. It goes both ways. So, certainly, nothing has changed in the employment context or any kind of decision-making that gives a positive or negative factor to any of the protected characteristics in employment that is and has been and will continue to be unlawful.
Curt Levey: The Harvard and UNC decisions only apply to university admissions. So, anything we say about other areas of affirmative action is speculative. Now, while it's certainly true what Commissioner Sonderling said that discrimination based on race has been unlawful since 1964, a lot of corporate America hasn't acted that way. And it's a strange situation in that we have these old precedents now from the '70s and '80s that give corporations a little bit of wiggle room to discriminate, to correct really gross disparities.
In my own opinion, what corporations are doing today, in terms of favoring certain minority groups, goes way beyond even that precedent from a much more liberal court. So, I think, if the Supreme Court were to rule on it again today, then I think what many of these corporations are doing would be found to be unlawful. But it hasn't reached the Supreme Court in 30 years. And, when it did, civil rights groups raised money and settled with Sharon Taxman to prevent a Supreme Court ruling, because they didn't think it would go their way. And so, I don't know when we're going to get another Supreme Court ruling on this.
Keith Sonderling: It certainly increased attention post the Harvard case, even though that was a Title IX case.
Kenneth L. Marcus: I think we can get one more question in within the hour. "How can a job applicant afford to litigate a Title VII case where he or she claims that the algorithm discriminates on the basis of race or gender?"
Keith Sonderling: Well, first of all, no matter what, before you could sue your employer in the United States, whether you work for a private company, state or local government, or the federal government, you have to come to the EEOC. So we see every single case of discrimination. And if it is a systemic case, a large-scale discrimination of AI bias, if it can be proved, I would imagine that's something, certainly, the agency would use federal resources towards. But that's very much dependent on where it's filed and the local offices. But, outside of that, especially, this is going to be a prime area for class action lawyers at some point.
And once they understand and realize how this is being used, and making discriminatory allegations that it's being used improperly, whether or not it is, I think we're going to see a lot of, not just federal government enforcement but, potentially, plaintiffs and class action lawyers having a lot of interest in this space, just because of the scale of decisions it's making, and how one decision, one data set, can essentially impact thousands or even millions of potential applicants. So I can imagine there's going to be a good amount of private actions here, as well.
And, Curt, I think you could probably chime in on that. Once that general awareness of the technology is being used -- right now, most employees -- again, there was a Pew survey done. The vast majority of Americans don't even know this technology exists or it's being used in the workplace. So I think what you're going to see, like in New York City with that consent requirement and that telling you you're being subject to this, then you'll see class action lawyers follow.
Curt Levey: I think the good news is that you have the same protections that you've always had. Like I said earlier, there's no exemption from our civil rights laws or from the processes at the EEOC just because somebody used an AI tool. And perhaps even better news is what Keith implied, which is that because there's so much buzz about AI, you'll probably find it easier to find a lawyer who's willing to take your case, either on an individual basis or especially on a class action.
Keith Sonderling: And I'll follow up on that and answer. It's one of the questions in the chat, which I didn’t get a chance to get to, which is so critical: the question, "Is the programmer or employer liable?" In the United States, employers are liable for the employment decision, whether it's made by a robot, or whether it's made by a human. Only the employer can make an employment action. So, for our jurisdiction at the EEOC, we have jurisdiction over three groups: employers, unions, and staffing agencies. And that's our world.
So, the vendors — these tools, if the vendor completely designs a system, deploys the system, uses their system — you can't turn to them and say, "Well, they're the ones that made the discriminatory hire." Liability is going to be 100 percent on the corporation. And, again, that's different than other areas of the law where there may be some joint liability, whether contractual or otherwise. But here, the employers are liable for the employment decision, going back to the enforcement perspective or the class action perspective, the bias perspective. That means the corporation's using them. They're the target when it's going to be coming to these lawsuits. You don't need to know who their vendor is, who they used to make the decision, because they're liable.
Kenneth L. Marcus: Excellent. And we are at the end of our hour. But we did not get all the way down to the bottom of the chat. So I think it's a tribute to our great speakers that it generated so many ideas on behalf of the Civil Rights Practice Group. I want to thank our speakers. I think this was an excellent discussion and invite, now, Sam Fendler to provide any additional remarks.
Sam Fendler: Excellent. Well, thank you very much, Ken. I'll take a minute at the outset to note that Ken was correct. A recording of today's webinar will be available on our website, fedsoc.org, in the coming days. You'll also be able to find the video on our YouTube channel, and the audio will be published in podcast form on Spotify, Apple Podcasts, Google Podcasts, and some more. But all of those links will be available from our website, again, fedsoc.org.
With that said, and on behalf of The Federalist Society, I want to thank our panelists and our moderator for the benefit of their time and expertise today. All told, it was a great panel, and of course, that led to a great conversation. So, thank you very much. I want to thank, also, our audience for joining us. We greatly appreciate your participation. Please check out our website, again, fedsoc.org. Or you can follow us on all major social media platforms at FedSoc to stay up to date with announcements and upcoming webinars. Thank you all once more for tuning in. And we are adjourned.