Why we can't fix bias with more AI w/ Patrick Lin
34:00 Share

My morning routine goes something like this: I wake up, hit snooze a couple times, finally get out of bed, and make myself a cup of tea. The tea is probably the most important part. And then I sit down with my tea and look at Twitter to see what everyone's talking about. And yes, I'm still calling it Twitter. Now in Feb of this year, I got out of bed, brewed my tea, and started scrolling. And that's when I saw that Google had launched an AI image generator inside their chatbot called Gemini.

And Twitter was on fire. People were calling Gemini all kinds of things. Racist, woke, biased, and everything in between. And the thing was, everyone seemed to have a completely different take on what was happening. So what actually happened? Well, Google launched their image generator inside Gemini and folks started using it.

One of the initial tweets from the user @endwokeness showed screenshots for Gemini's response to a bunch of different prompts, ranging from "The Founding Fathers of America" to "Viking" to "Pope." While you might expect that each of these prompts would yield pictures of mostly men, and mostly white men at that, Gemini would prove you wrong.

The founding fathers, Vikings, and popes Gemini generated were people of color, and in the instance of the pope, some of them were women. Okay, so this is obviously a pretty significant problem of historical inaccuracy. Other users shared screenshots of Gemini's responses to their own prompts, and some of them were pretty bad actually.

Among the most egregious? Pictures of people of color in Nazi uniforms. Definitely not the kind of diversity we're looking for. Then Google weighed in. Prabhakar Raghavan, a senior vice president at the company, wrote in a blog post that the Google team tried to get ahead of, quote, "...some of the traps we've seen in the past with image generation technology, such as creating violent or sexually explicit images." In other words, they were trying to correct for AI bias.

which had become a major topic of conversation in AI ethics circles. But this resulted in pissing everyone off. To some on the left, it was advancing a kind of colorblind identity politics that glossed over the history of oppression. To some on the right, it was over-representing minority groups and advancing some kind of conspiratorial big tech woke agenda. In fact, Elon Musk, always the provocateur, called Gemini both woke and racist. Go figure.

Everyone was shouting about the bias baked into the system. But in doing so, they ended up revealing their own biases. In other words, the lens through which they see the outputs of these AI systems. I'm Bilawal Sadu, and this is the TED AI Show. And on this episode, we're tackling one of the thorniest issues out there, bias in AI. Want a website with unmatched power, speed, and

and control? Try Bluehost Cloud, the new web hosting plan from Bluehost. Built for WordPress creators by WordPress experts. With 100% uptime, incredible load times, and 24-7 WordPress priority support, your sites will be lightning fast with global reach. And with Bluehost Cloud, your sites can handle surges in traffic no matter how big. Plus, you automatically get daily backups and world-class security. Get started now at Bluehost.com.

So look, some of these images that Gemini generated are really bad. I'm not excusing those. But I get what they were trying to do. Essentially, they ended up overcorrecting for some of the major misses that AI image generators like DALI, Midjourney, and Stable Diffusion have stumbled through in the past. Like in late 2023, a group of universities put out a joint study about their findings on text-to-image generators.

They found that when they prompted for pictures of surgeons and surgical trainees, they got some troubling results. The vast majority of images were white men in surgical gear, which isn't actually reflective of the demographic breakdowns of surgeons today.

Another investigation by the Washington Post on bias in AI-generated images showed similarly troubling trends. The prompt "Muslim people" revealed men in turbans and other head coverings. The prompt "attractive people" yielded pictures of only young light-skinned folks. And the prompt "productive person" generated pictures of men sitting at desks, most of them white.

Now, there's been so much talk about bias in AI, from flaws in the outputs to flaws in the training data. In cases like the Gemini scandal, we risk historical inaccuracy and miseducation, undeniably real problems. And in other cases like the study and the investigation I just mentioned, bias in AI can keep perpetuating harmful stereotypes that don't actually reflect reality. So if everyone agrees bias in AI is a problem, then why isn't anyone fixing it?

To help us unpack this, I sat down with Patrick Lin, a professor of philosophy at California Polytechnic State University and the director of the university's Ethics and Emerging Sciences group, which tackles topics like AI and predictive policing, the future of autonomous vehicles, and cybersecurity in space. He's been examining the ethics of technology for a long time and is thinking a lot about bias in AI, where it comes from, and how it impacts us on a daily basis.

So when we say the ethics of AI, we're talking about a huge, almost never-ending topic, right? Can you explain for our audience why AI ethics is such a vast topic?

So when we're talking about human ethics, you know, your ethics and my ethics, we could do that because we like to think we have free will and we can make choices. And some of the choices we make are ethical or unethical or neither. But when you talk about machines, some people rightly point out, hey, they're not moral agents. They don't know what they're doing. So how could they be held to any kind of ethical standard? Right.

So quickly, right off the bat, I would say that's a wrong interpretation, a wrong understanding of technology ethics or AI ethics.

Ethics could be about, you know, not so much about the technology as an agent, but about how the technology is designed. It could be about the ethics of the technology developers. It could also be the ethics of the technology users. It's about a whole ecosystem of developers, users, stakeholders, unintentional stakeholders, environmental interest, you know, and so on.

When it comes to AI, now, AI, I mean, you know, at a very high level, you could think of AI as, you know, an automation of a decision-making process, right? So AI decides, well, what is this image I'm looking at? What is this text I'm seeing? And it makes a decision on predicting the next words, right? So it's a decision engine of sorts.

And because it's a decision engine, it could be used to replace decision makers. If AI can be integrated in society and a lot of these decision-making roles, then that already implicates countless domains, right?

Right. AI in agriculture, in chemistry, in education, in warfare. It's hard to imagine a single domain where AI cannot be applied to. This means that, you know, you're really looking at the entire universe of ethical issues potentially for AI ethics. That's a great point, especially as AI permeates all these verticals and domains. As you say, the surface area for this bias to manifest itself is also very, very broad. Right.

And so today you and I are going to talk about bias in AI. And there's a bunch of interesting examples in there, just a few that we've come across. One is racial bias and facial recognition, right? Some facial recognition systems have been shown to have higher error rates for people with darker skin tones, potentially leading to false identifications, right?

Amazon scrapped an AI recruiting tool after discovering that it was penalizing resumes that contain the word women's after downgrading graduates of all women's colleges, reflecting gender bias in the training data. What are some concrete examples of bias that you've encountered in the AI space?

If we start out with the, you know, technology du jour, which is LLMs, you know, like ChatGPT and, you know, AI writers or chatbots, we could already see bias in their outputs. I'm thinking about Google Gemini's recent debacle where it wants to diversify the ethnicity of Nazis and the founding fathers of the United States, right? Yeah.

Who are all white? But if you think that, wait a minute, anti-bias, anti-discrimination means you got to mix it up with the ethnicities and the genders, then that's how you get some false negatives.

But I think the big stake examples are still related to AI bias in hiring, which you mentioned, but also AI bias in bank lending and criminal sentencing. And I would even include things like AI policing. So these are potentially life and death decisions.

Even a bank loan decision could be a life and death decision. If you're denied a loan for a mortgage, then that could mean you lose your house. You could become homeless. And, you know, homeless, unhoused folks tend to have a shorter lifespan than other folks. Right. So these are big, serious decisions. And even if the AI doesn't look specifically for gender issues,

ethnicity, age, it can still deduce a lot of this information from other data. So, for instance, a banking AI, you know, making a loan decision could be programmed or trained to ignore ethnicity, right? Ignore race. But

it could still discriminate in its outputs, in its impact. So for instance, it might say, it might, you know, given this trading data, it could say, oh, you know what? Borrowers from a certain zip code, you know, have a high rate of default. So we're going to just not give loans to people in the zip code. But guess what? It turns out that zip code is full of minority neighborhoods. All right. So it is a proxy for race or ethnicity, which is discriminatory.

For almost any given AI application, you could probably come up with some kind of, you know, some kind of weird case of bias. I mean, we talk about healthcare AI. If a medical AI is trained primarily on, say, white patients, then it might misdiagnose someone who, you know, of African descent or Asian descent or Jewish descent, right?

And by the way, this isn't just a white versus other thing. I mean, if you look at facial recognition projects in China, for instance, where they train their AI mainly on Chinese faces, they have a hard time recognizing differentiating white faces. Just because white faces are underrepresented in their data set.

So it's not that AI is inherently racist in one direction. It depends on the training data. Absolutely. It brings up the concept of implicit bias also and how it might surface in AI in ways that we don't expect. One example that I've been really fascinated by recently is if you type in just the word or token thief into any text to image model,

You're going to get an image that resembles a character from the video game Assassin's Creed or the 2014 video game Thief, rather than, you know, the stereotypical depiction of a thief wearing that mask with a money bag slung over the shoulder or worse, a racist caricature. You get this person in a cape and that's what the model thinks a thief is. Right. And so this seems to reflect biases present in the training data, which in this case over represents video game imagery and

How can we account for and mitigate these sort of implicit biases in AI systems, you know, so we can ensure that we're not embedding or reinforcing problematic stereotypes, let's say, for media or pop culture? Implicit biases, by definition, are hidden. They're under the surface. I mean, you've lived in it for so long, you don't even...

realize it's there. You know, I mean, you must have heard the joke where, you know, there's two fish and one fish asks the other fish, hey, how's the water? And the other fish says, what's water? Right? I mean, it's just so pervasive that you're not even aware it exists. I mean, that is part of the problem, recognizing bias when you see it.

And AI bias is a popular problem because people understand bias. They can imagine that they could be on the wrong end of an AI decision someday. No matter who you are, no matter how privileged you are, that could be you. So that's why AI bias, out of all the various issues in AI ethics, might be the most well-known one.

most widespread one. But the big trick is how do you get rid of AI bias? Bias is such a tricky problem because of, you know, I think how humans are just simply hardwired and constructed.

I mean, think about our brains, right? We're not just flawed machines. We're stereotyping machines. That's what we're built to do. We're built for one-shot learning. We're built to learn very quickly. And I mean, early on in humanity's history, this was critical for survival. So imagine you're the first caveman who's ever come across a

You think, ooh, what is this weird orange thing? I wonder if it's poisonous or not. You nibble it, you eat it, you survive, right?

So it's natural to make a judgment that anything that looks like this is also going to be edible, right? So that's a form of stereotype. It could also go too far, especially when you talk about people. Individuals have so much variation from one person to the next, even inside the same groups, whether you're talking about ethics groups, religious groups, or whatnot. But also another tricky thing about...

Bias and stereotypes is that, you know, it seems that there's some kernel of truth in the stereotypes, right? Otherwise, they wouldn't be stereotypes. But, you know, but to make such a broad judgment and start making decisions based on stereotypes, that seems to cross the line.

I think that's a really good point, right? Like, as you say, bias has existed since time immemorial. It's almost intrinsic to our nature. It's perhaps a simplistic way of, you know, looking at patterns and extrapolating based on that. And so it's going to be really hard to solve this problem

in the AI space, right? And we can't use more AI to solve it because AI doesn't know right from wrong. Like what even is right? What is the truth, right? Since it can't detect what is and isn't biased or racist or misogynist,

You know, the logical fix is to train the AI on less biased data. But where do we find all this less biased data? Because the data is generated by humans that bring their own biases to the party. So this is no small task. Right. And I think you've set up the problem appropriately. So to your mind, how can we even start to fix this bias problem in artificial intelligence? Yeah.

We don't understand bias well enough. We do have an intuitive, superficial understanding of bias. You know, you might think of bias or discrimination as just treating people

People differently because they're different, because of their different gender or ethnicity or religion. And these are generally legally protected categories. That's the usual understanding of what bias is. But if that's all you have, you're going to get it wrong. We need a deeper, more nuanced understanding of bias if we're going to truly tackle the problem.

Support for the show comes from LinkedIn. LinkedIn ads allow you to build the right relationships, drive results, and reach your customers in a respectful environment. They are not getting barraged. This is very targeted. You will have direct access to and build relationships with

With a billion members, 180 million senior-level executives, everyone's on a LinkedIn, it seems like, and 10 million C-level executives, you'll be able to drive results with targeting and measurement tools built specifically for B2B. In technology, LinkedIn generated two to five times higher return on ad spend than

than other social media platforms. You'll work with a partner who respects the B2B world you operate in. 79% of B2B content makers said LinkedIn produces the best results for paid media. Start converting your B2B audience into high quality leads today. We'll even give you a $100 credit on your next campaign. Go to linkedin.com slash TED audio to claim your credit. That's linkedin.com slash TED audio. Terms and conditions apply. LinkedIn, the place to be, to be.

What are problems that arise when our definition of bias in AI isn't sufficiently nuanced? One example would be this. If you think it's inappropriate, if you think it's discriminatory and biased to treat people differently because of age or gender, you know, if that's all you think bias is, it's going to give you a lot of false positives, right? So here's an example that shows that it might be okay to discriminate on age and

and on gender and on ethnicity at the same time. So imagine I'm a filmmaker and I'm interviewing actors for the title role of Martin Luther King Jr. I'm going to reject every single Asian teenage girl who auditions for that role.

And I'm rejecting them for a job precisely on the basis of their age, their gender, and their ethnicity. But that seems okay. At least if I'm trying to make a historical, accurate biopic, right? It seems legitimate that I could filter out applicants based on their profile if they don't match the age, ethnicity, and gender of what I'm aiming for.

You can't simplistically say thou shalt not discriminate based on protected categories and hope for the best, right? So clearly there are a lot of problems. What role do you think subjects like ethics, philosophy, social sciences play in the training of these AI researchers and developers that are building these next generation systems?

Oh, I think it's huge. I have great respect for science and technologists. I wanted to be one of them when I was growing up before I accidentally found philosophy. They fundamentally are curious. They want to know how things work. But more than that, they want to change the world for the better.

But here's the problem. You also got to understand the world in order to make those kinds of interventions. Technology doesn't really do a great job in solving social human problems. Only humans can really solve social problems. Technology and AI, they're tools. They could do some things. They could alleviate some of the symptoms of these problems.

But they have a hard time getting at the root of the problem. Take the human, very human problem of drunk driving. All right. Now, it's hard to change culture. It's even harder to change drinking culture in America. But one thing we can do is make cars safer. All right. So we can make cars more survivable if you get in an accident. They're saving more lives. But are they doing anything differently?

to the drunk driving problem? You know, are they making any progress in rolling back drinking culture? I would say no, right? In fact, it might be worse. They could be encouraging more drinking and more drunk driving. If you know your car is safer and you're more likely to get home in one piece and you're less likely to kill random pedestrians or other drivers, then that's an incentive to drink more because you know things will be okay.

So we've talked about the problems AI creates. I want to switch gears a bit and talk about solutions. I'm wondering if you see folks or companies or organizations out there doing anything to fix the problem of bias in AI.

Well, I mean, I do see a lot of organizations say they're working on things to fix AI. It's not entirely clear how they're doing it, some of this proprietary information. But still, you know, again, I would still be skeptical that their solutions are going to do a whole lot in solving the problem. The one move they're thinking of is to throw more AI at it.

And this is the, you know, this is exactly the problem of if all you have is a hammer, then everything looks like a nail, right? If you're an AI, you know, programmer or developer, of course you're going to think AI is going to be able to solve that problem. And that's what you're going to try. But I think with bias, it's a different kind of challenge. For one reason, bias is,

is a social construct. So the challenge facing developers and making AI that can detect bias is the same kind of challenge with developers who

Think they can make AI that can detect pornography or can detect an unethical situation, right? Pornography, ethics, they're social constructs too. They're very squishy. They're very hard to define. They resist definition. The U.S. Supreme Court had famously concluded, we might not be able to define pornography, but we know it when we see it.

And I think humans are like that with bias, too. We can recognize bias when we see it most of the time. In my Martin Luther King film example, you can recognize that I'm not being malicious. I'm not doing anything inappropriate. But a machine might not be able to. So machines are not great with ambiguity.

There's no law of nature that says technology will solve all your problems, right? I mean, it's made life easier in a lot of ways. It's made us more secure in a lot of ways. But it still hasn't solved hunger, right?

racism. Think about all the societal ills. If it could, why aren't we working on that? I mean, all we have now are apps that make life more convenient. Here's an app that could find me a ride. Here's an app that could find me a place to crash at night. Here's an app where someone will do his chore for five bucks.

We're putting AI and all our best minds on to these projects to do things that someone else's mom will do for you. They're not like these great world-shaking applications.

So back to bias, I think the temptation here is just to throw more data at it. A couple of ways we can go here. Yes, you could curate your data sets to ensure that AI is being trained on examples where there's no bias. But when you're talking about millions and billions of, you know, examples in a large training set, that's not a very feasible solution. It's definitely not scalable. So, yeah,

So if you want a scalable solution, it seems that you would need to create an AI, train it, program it to identify bias when it sees it. But to do that, it needs to be crystal clear on what bias is, what discrimination is. If you try to look up the definition of bias, you're not going to find a really good one.

They say things like discrimination is the unfair treatment of people. Now you have to define what fairness means or unfairness means. Right. But I think so that work hasn't been done. And if if developers don't understand the nature of bias, then they're going to have a hard time fixing the problem.

You know, I'm trying to imagine an AI that can understand all the nuances of every situation here, you know, of whether this is relevant or this factor is not relevant. And I have a hard time imagining that that could be done. It sounds like what you're saying is we're not going to have one model to rule them all, right? It seems like it's a lot about context. I want to spend a little time on that. Yeah.

There's a lot of discussion about national AI models that are tailored to cultural contexts that vary from region to region, right? The concept being that

that bias in AI isn't a one-size-fits-all solution. And what constitutes bias can vary significantly based across, you know, across different cultures, regions, and nations. What's considered acceptable or unacceptable, offensive or benign can differ based on local norms, values, histories, and sensitivities. What do you think about, so do you think we should have a

plurality of models that are tuned to various cultural contexts. And could that be one near-term solution to address this rather thorny issue of bias? As a general approach, I think it makes sense because there's no one set of values to rule them all, right? There's no one ethical theory to rule them all. And there are variations in ethics that

uh, from culture to culture. And many of them are reasonable variations, right? I mean, others are not reasonable. Others are just plain offensive, right? So if a culture that doesn't, uh, want women and children to be educated and thinks it's okay to throw acid on them, uh, to prevent them from going to school, that's bad. I, I don't see any reason to respect that kind of, um, you know, those kinds of values, but other, other differences could be reasonable. So for instance, uh,

You know, in Asian cultures, elderly tend to be more valued than kids, right?

So, for instance, let's imagine an AI app that does triage in a hospital, right? That does hospital admissions and these hospitals in these big cities are generally overworked. So it's got to figure out a priority list for the patients to be seen. In one culture, let's say in Asia, it might give bonus points if you're older, if you're elderly. It might move you up the priority list.

And that seems okay, neither here nor there. And other cultures, they might have the opposite value. They might treat their children as kings and queens, and they put a premium on them, in which case a hospital AI or triage AI in that culture would move younger people up on the priority list. I think we would want to respect that.

diversity and these variations, especially since there's no one ethics, no one culture to rule them all. I certainly don't think we have it right here in America. Same with just about every other culture. But if I were an AI company looking to roll out products worldwide,

Now I'm thinking, wait, I got to localize my products. You know, my AI in these products, that means I got to train my AI from data that comes from those geographies and cultures for every market I want to play in. Right. And that...

Sounds like a lot of work. I mean, it could be a deal breaker. You know, the diversity model makes sense in theory, but in practice, how do you implement that? I don't know. So I have to say, do you think we're going to be stuck in the same whack-a-mole loop where the problem grows and multiplies exponentially, especially as all these major labs chase this current paradigm we're on, which is let's throw more data and more compute at it, do even bigger training runs,

Or is there a chance that we could get this under control? I do see potential fixes, but they're hard fixes and people don't want to hear about them because they're about human labor, things that human beings need to do, the work we got to put in to solve this problem. What should individuals do to address bias in AI, right? Look, this...

This is a hard problem because it's a social problem. It's a human problem. And I think it would be a mistake to put all your eggs in one basket and hope that technology can solve this. Unfortunately, I think it's going to take a lot of hard societal level work. It's worth trying out. Even if only technology can paper over the symptoms...

That might be okay for now. Just as safer cars aren't fixing the problem of drunk driving, but they're saving lives, that might be enough.

So, you know, I would say good luck to the developers. But I think, you know, if you really are serious about tackling bias, you got to understand what it is. Thank you so much, Patrick. This is clearly a complex problem. And I really appreciate you taking the time to break it down and explain to us why it isn't a very simple one size fits all solution. So thank you for your time. And yeah, we really appreciate it. You're welcome. Thanks for having me on.

So Patrick said something that I think is really important for us to recognize. In order to solve the bias in AI, we have to solve the bias in ourselves, which is a pretty tall order, right? Especially when, as he says, bias is pretty much implicit to human nature.

And I tend to agree with Patrick that throwing more AI at already flawed AI systems isn't necessarily going to solve the problem for us. Because here's the thing, AI is a reflection of who we are. After all, it's trained on us, our art, our memes, movies, jokes, history, music, math, science, philosophy. It's complicated because we're complicated. It's flawed because we're flawed. It's biased because we're biased.

But I also don't want to throw up my hands and say, "We can't fix this," or "There's nothing we can do," because I think there are some things that we are actually in control of when it comes to bias in AI, particularly our responses to this nascent technology.

Now, much has been said about large companies creating transparency around their training data, and that's a welcome step. But even training data transparency presents its own challenges. For starters, these data sets are enormous. We're talking billions of images and trillions of words.

it will be a massive effort to comb through it all, find all the flaws, and make the necessary changes. So it's a big, big problem with elusive solutions. But I want to offer up a couple of solutions that I think could, at the very least, help. The first is something I mentioned in my interview with Patrick. The idea that we could create more nuance in our AI models by keeping it regional. Like, generative AI systems in Singapore probably should not behave identically to generative AI systems in California.

In fact, the Singaporean government called for AI sovereignty, addressing the fact that, quote, Singapore and the region's local and regional cultures, values, and norms differ from those of Western countries where most large language models originate, end quote. I believe that AI sovereignty can help preserve our diversity, whether that's state-to-state or country-to-country.

The second might seem at first glance to be a little too obvious. You might not like it at first, but hear me out. What if we gave ourselves a bit more agency in this issue and committed to getting better at using tools like ChatGPT, Midjourney, and Gemini?

Let me give you an example. I want to share a process I go through in my mind every time I prompt an image generator or a chatbot or what have you. First, I remember that I'm using a flawed tool. Just because it's AI and it's supposed to be really smart, it's not always going to give me the most accurate results. Second, once I get a response to my prompt, I scrutinize it. In the same way I scrutinize the news I read, I'm skeptical about the source of the results.

knowing that these AI systems are trained on imperfect data. Third, I take a beat. Before I'm even tempted to jump on Twitter and post a spicy screenshot of this messed up response to my prompt, I pause and I think about what I need to do to get a better response, and I revise it accordingly. Maybe I'll respond to ChatGPT with something like, hey, not all nurses are women. Can you show me some images of nurses that are more reflective of the actual demographics of the nursing field?

Because if these generative AI tools are only as good as the data we feed them, they're also only as good as the prompts we give them. And yes, it is up to these big tech companies to make better products and give us more and more transparency about how they're trained.

They should also give us more transparency into how and when they're trying to solve for bias within their systems. For example, if Google had told us they're trying to address some of the bias in Gemini, it may not have solved the problems in the images generated by their system, but it at least would have helped perhaps to know why the AI system was generating those images in the first place.

But also, we need to be educated consumers and users of these tools and know that the better we are at identifying their flaws, the better we will be at prompting them for better responses. That's why developing AI literacy is so crucial. We need to understand how these systems work, how they learn, and how they can go wrong, sometimes horribly wrong.

and we need to stop taking their outputs as gospel or a factual reflection of reality. It's as crucial as being literate about our own biases. If what an AI system generates is not consistent with our values, we can absolutely take control and shape it for the better. The TED AI Show is a part of the TED Audio Collective and is produced by TED with Cosmic Standard.

Our producers are Ella Fetter and Sarah McRae. Our editors are Ben Benshang and Alejandra Salazar. Our showrunner is Ivana Tucker, and our associate producer is Ben Montoya. Our engineer is Asia Pilar Simpson. Our technical director is Jacob Winnick, and our executive producer is Eliza Smith. Our fact checker is Julia Dickerson. And I'm your host, Bilal Siddoo. See y'all in the next one.