85. Timnit Gebru Looks at Corporate AI and Sees a Lot of Bad Science

Portrait of Timnit Gebru
Reimagining the Internet
Reimagining the Internet
85. Timnit Gebru Looks at Corporate AI and Sees a Lot of Bad Science

Timnit Gebru is not just a pioneering critic of dangerous AI datasets who calls bullshit on bad science pushed by the likes of OpenAI, or a tireless champion of racial, gender, and climate justice in computing. She’s also someone who wants to build something different. This week on Reimagining, we talk to the thrilling, funny Dr. Gebru about how we can build just, useful machine learning tools while saying no to harmful AI.

Timnit Gebru is the founder of the Distributed AI Research Institute and co-founder of Black in AI.

Papers mentioned in this episode:


Ethan Zuckerman:
Hey, everybody, welcome back to Reimagining the Internet. I’m your host, Ethan Zuckerman. I am really thrilled today. We are sitting down with Dr. Timnit Gebru. Timnit is so many different things to the field of AI. Honestly speaking, in talking about AI at this moment, she is the person I most want to be talking to.  

She is the founder and executive director of the Distributed AI Research Institute. She’s the co-founder of Black in AI. She’s written a ton of really influential papers, including a paper called “Gender Shades,” along with Dr. Joy Buolamwini, which helped document how racial bias can come into facial recognition algorithms. She’s responsible for some really groundbreaking work on labeling what is in AI data sets. There’s a paper that she’s a co-author of called “Stochastic Parrots,” which comes up a ton now as people try to figure out the dangers and the limits of large language models.  

She left the private sector in 2020 found Timnit left the private sector in 2020 after sounding alarms at Google about its AI tools. Timnit, I’m so glad to have you here. I know you want to start by telling me that that maybe isn’t the best characterization of the way that you left Google.  

Timnit Gebru:
Yes, I always correct people. I was, you know, unequivocally fired in the middle of my vacation in the middle of a pandemic. I found out after, you know, I was trying to log into the corporate account and was denied access. And so I was wondering what happened. And I saw an email to my personal email. So yes, definitely fired.  

Ethan Zuckerman:
So, so talk about this paper, you had written this paper with two co-authors. People often refer to it as the “Stochastic Parrots” paper. What was so radical about this paper that Google felt compelled to fire you? And let’s just point out, you were really heading up their work on ethics and AI. You were doing incredibly important work within Google around diversity and equity and inclusion. And it was a giant PR fiasco for them, as well as a really challenging situation for you. Why is this such a dangerous piece of research? What’s in the Stochastic Parrots paper?  

Timnit Gebru:
So when we wrote this stochastic parrots paper, we didn’t think it was radical or dangerous. We just thought it was going to be, you know, some kind of a cool and interesting survey paper that some people will not like, but we didn’t think it was, you know, groundbreaking or anything like that.  

And so what happened was that I was seeing this rush to build larger and larger language models, right? So these are models that are trained to output the most likely sequences of words. And they’re trained based on the troves of data on the internet, right?  

And so I was seeing this conversation around, why are we not building the largest models? Why don’t we have the largest ones? Why do they have the largest ones? And all these companies seem to just be like rushing to build larger and larger models and larger in the sense of more—requiring more data, requiring having more parameters, requiring more compute power. I just did not understand the reason that they were just so focused on the size, you know?  

Actually what was interesting was there were some people at Google, even internally to Google, who were asking our team, the ethical AI team, co-lead by Meg Mitchell and I, what kinds of things they should think about with respect to large language models.  

So I messaged Emily Bender, Emily M. Bender who is a linguist and who came up with the terms stochastic parrots. And I said, look, do you have any papers that I can point to? Do you have any? because I’m like, I’m just emailing your tweets right now and it’d be nice to cite a paper. And then she says, you know, which of my tweets are you sending? And I sent it to her.  

Ethan Zuckerman:
Always, always good to check on which tweets you’re forwarding for all of us on Twitter.   

Timnit Gebru:
Yeah. And then the next day, she sends me an outline and the outline has the title, all the sections, the title of the sections. And starting with the environmental impact, the environmental impact of training and using such models, because you have huge compute requirements, like I said, these things are not just trained on like quote unquote, the cloud, right? You have, you know, it’s not just the actual cloud, you have data centers that require to have people working in them that require water, that require electricity, you know, power. And we even, we talked mostly even about environmental racism, not just, you know, the environmental catastrophe, but environmental racism, because when you see the languages and the people who benefit from these systems, they’re often not the people who are paying the cost of the climate catastrophe. So that’s what we started with.  

Then we talked about what kind of issues you have when you just, you have what we call, you know, laissez-faire kind of data collection. And this was a term that my collaborator, Eun Seo Jo, came up with in another paper that we’ll talk about. But we talked about the kinds of issues that you have when you just gather data from the Internet and you don’t curate it.  

And when you feel like you have to create huge models, you’re less and less likely to curate it, right? because you think, oh, it’s too big to do anything, right? And so we talk about the issues that you have when you do such things.  

So for instance, we describe how size doesn’t guarantee diversity. many people think that just because you have so much data, that means that diverse viewpoints are represented. However, we know from the Internet, just starting from who has the internet and who doesn’t, right? Who has access to the internet and who doesn’t, we know that so many people are already excluded. And then when we see representations of people on the Internet, we know, for example, Wikipedia has mostly talks about men, and certain geographic locations.  

So when you analyze the data, you very quickly see that it represents hegemonic views, which is what will be represented in the outputs of the models as well. So you don’t see the things we talked about.  

Ethan Zuckerman:
And this critique, this critique of hegemonic models, is one that you had published on previously. You’d published with our mutual friend, Dr. Joy Buolamwini, a paper that she worked on when she was a master student at MIT called gender shades, which was looking at very large computer vision systems that had systemic biases against women, people of color, and particularly darker skinned women of color.  

How does that work sort of connect into this sort of critique about, you know, big isn’t good enough, it needs to be diverse rather than big? 

Timnit Gebru:
It’s exactly the same issue, right? It’s the same issue in the textual domain. So a lot of these systems are trained using those gender recognition or gender ascription models are trained on images from the Internet. Many of them were trained on images of celebrities, for example, right? And so then you get to see who is considered a celebrity, what their demographics are, and the same exact problem, right? Who is represented on the internet, and how are they represented?  

And other people like Abeba Birhane and Vinay Prabhu, a whole bunch of people have written about, have shown, for instance, the issues in the ImageNet data set, and how people are represented, right? Especially Black women and other groups of people, the ways in which they are described. And so if that’s what we’re using to train any kind of model, how do we expect anything else to come out?  

And what’s interesting is a number of us independently had also written about curating data sets and documenting data sets. My colleagues and I wrote about a paper called “Datasheets for Data Sets,” which was about curating, I mean, documenting your data set, annotation and gathering process. Emily M. Bender and Batya Friedman had written “Data Statements for NLP.” And then later on, Meg Mitchell and I and our colleagues wrote a paper called “Model Cars for Model Reporting,” which was this idea of testing and documenting or testing for models. So these were ideas that were brewing for a while.  

And so it’s so disappointing to see, right? First of all, after all of this work that Dr. Joy Buolamwini and I put in, and Deborah Raji and many other people put in to alert the world about these issues with respect to the visual domain, right? We still see the same issues in the visual domain. Let alone help address them in this other domain that we’re talking about. So it’s the same kind of issue of power, of access of representation that comes up over and over and over and over again.  

Ethan Zuckerman:
What’s so amazing for me about the Stochastic Parrots paper is that it’s so prophetic, right? You’re writing a paper that, you know, as you’re saying, it isn’t necessarily, you know, presenting novel research. It’s really sort of a summary of what’s known in the field.  

We know that these very large language models are extremely expensive to generate. got severe environmental consequences, we know that they end up embedding whatever biases of the data sets that come into them. And really through your work, we know that often these data sets are put together without much care, without much consideration. There’s sort of this desire to stuff in as much data as possible with the idea that bigger is better.  

And then we’ve watched this happen in real time. We’re watching this fight between OpenAI and Bing and Google, using everything from the colossal common crawl to almost anything else that they can get their hands on.  

What’s going into these very large language models? How are they getting built? And why is that so dangerous, Timnit?  

Timnit Gebru:
Well, the first point is that what is, what they’re using to train these models, they’re not transparent about, right? OpenAI doesn’t even tell us the architecture of the actual model or what data they’re using to train it, right? And now they’re starting to say that for our own good, they’re not going to tell us the data.  

However, there are some other models that people have documented. And it’s a bunch of data from the Internet, whether it is from Wikipedia or books, but ranging to all sorts of other dark matter from the Internet.   And then sometimes they try to filter things out and they have some crude ways of filtering things out. But those are also detrimental.  

William Agnew pointed out that some of the filtering keywords that they used to filter out data was actually leaving out all sorts of websites dedicated to queer people, right, and expression. We talk about how, for instance, this was actually a contribution by Mark Diaz, whose name does not appear in the paper because he’s still at Google and he was told that he can’t have his name on this paper. He talks about how in his research, older people in the US talk about age discrimination or other kinds of issues, mostly in blog posts and not on social media or other kinds of, or Reddit or anything like that.  

And that kind of content is underrepresented. we talked about how the static nature of the data, right? That kind of holds society back because even after we have certain really important discussions and phrasings and wordings and websites and things are updated, those are not used as they’re not to train these models, right? And so even if society makes progress, what’s represented in these models is not the progress that we made. And I’ve seen some discussion of that now.  

And so, you know, it’s what’s really interesting to me is that when we were, this whole large language models kind of thing was not initially when we were seeing earlier work, not the ones coming out of OpenAI, but for instance, a precursor, it wasn’t, they weren’t being described as artificial general intelligence or generating content or anything like that. It was kind of being used as a component in a system for something else, right? So it was more about what people called representation learning.  

But what happened with OpenAI is that they started, this is why I’m now, if you pay attention to all the things I’m talking about right now, I’m talking about the whole idea of artificial general intelligence, which I don’t even know what it is. Many of us don’t. It sounds kind of like an all-knowing God that they’re trying to build. And so this idea of trying to build an all-knowing God machine or whatever it is, is what to me has brought many of the issues to the forefront. Because if that weren’t the goal, I really don’t think that they would have tried to build larger and larger language models. I also don’t think that they would have tried to build like an all-knowing chatbot or an all, the text and image generation stuff that they have now.  

I think that it would have been more a component in a system that helps people with various tasks, right? 

Ethan Zuckerman:
Which, so it’s actually a good moment to stop for a moment and talk about what’s in a system like chat GPT. Because there’s actually a whole bunch of different tools going on there. There are some very well-trained algorithms that can summarize a text. It’s leaning on some algorithms that can pull transcripts off a YouTube video or maybe do translation for it. And then it’s all wrapped up within this text generation system, which is trying to figure out statistically probable words to add and sort of put together in sentences.  

And what I hear you saying in some ways is those are all interesting components. They can all become components of different systems. They’re all components we would want to interrogate. Each one of those components has biases built into it based on what it that was trained on, what it was not trained on, somehow in part, because we’ve had this idea for years of the Turing test, this idea that if a computer can have an intelligent conversation with us, if we can’t tell a human or a machine apart, maybe the computer has achieved some sort of intelligence, that feels like the idea that’s kind of swallowed up the rest of those ideas.  

Is there a better way that we should be thinking about systems like Chat GPT as not necessarily a question of whether they’ve achieved general intelligence, but how would you like us to sort of interrogate them and think about them? 

Timnit Gebru:
I think the “stochastic parrots” phrase is a good one, right? If you, when you see a parrot repeating what humans are saying, you’re not thinking that the parrot is understanding what we said, but the parent can repeat what we said back to us.  

And so these systems are kind of a more complex form of that. They have been trained to explicitly look at what’s on the Internet and kind of regurgitate it in certain forms. And when that’s not enough, when that’s not enough for that to be not the only thing, you have all sorts of humans kind of looking, checking the outputs and trying to train it to do something different.  

So for instance, when the outputs of Chat GPT are toxic, extremely toxic, or where people use it to train extremely toxic things—similar to social media platforms where the rest of us, sometimes don’t see some of the worst and goriest stuff—you have content moderators who are just working day and night to filter out that gory and horrible stuff, right?  

So you have all of these people sitting day in and day out kind of looking one by one and saying, “Okay, this is not toxic. “This is another toxic.” We’re giving some very elaborate kind of descriptions or labels.  

Those people then, you know, that output is used to create another filter that they use to then put on top of the outputs to then filter out the ones that are toxic versus not. And so those people who are doing that kind of filtering, I don’t think that they have any illusion that these systems are incredibly intelligent or sentient or anything like that because they’re seeing how the sausage is made and they’re seeing the kinds of inputs from them that is necessary to give the appearance of human-like language or other kinds of behavior of the system.  

Ethan Zuckerman:
And that what we’re getting when we’re interacting with one of these systems is an algorithm, but it’s also then going through poorly paid contractors of SamaSource in Kenya who are editing it and in the process doing annotation to train the next version of this.  

And the stochastic parrot’s critique, which is that this is not intelligence, even if it sounds very much like human speech, makes perfect sense. It’s really emotionally hard for people to get. People really do seem to have an enormous amount invested in the idea that these systems may be sort of approaching some sort of intelligence. One of the things you hear people talking about a lot is emergence and this idea that systems in some cases seem to be gaining abilities to do things that they weren’t actually trained to do. So that a system that’s mostly trained to have a conversation seems to over time develop the ability to do five digit addition, for instance.  

Is that something that sort of gives you pause that this is just parroting information back? There’s something deeper going on there, or people sort of overweighing those stories and emergence?  

Timnit Gebru:
I really just don’t trust any of these claims. And in fact, because there are so many things going on. When people claim that this recently there was some, some manuscript that was put out on our archive saying that chat GPT is passing some exams at MIT or something like that. the professors finally had to put out a statement, but not before a bunch of other people analyzed those claims and debunked them, right? They could have put out that statement way before. Now that somebody put out those claims, they’re trying to, I don’t know, there’s a whole process issue there.  

But again, there is a fundamental issue where, and this is the most basic thing in machine learning, where you don’t test on your training set. You don’t validate on your, you know, and so, and it’s like that seems to be completely broken. People don’t even know what the training data is. They don’t, they have no idea what, what the training data is. And if the training data is all of the internet, so, and then what are you, what are you testing on then? You know, you’re tracking on stuff that you already saw on the Internet. 

Ethan Zuckerman:
Right, so just to, just to unpack this for someone who may not immediately get the implication what you’re saying. When someone is claiming that an AI model is passing the LSAT, it’s possible that there are LSAT test exams online that the model is already trained on. And that testing against data that it’s trained on is always going to give you invalid results. The model has actually learned from that data. It’s seen it before. It’s going to be able but predictably come up with it, the only way to do a fair evaluation is to test it on novel data. We normally do this by withholding some training data and then testing on that clean data.  

But in part, as you’ve been pointing out since very early on in your work on this, we don’t know what is in these models. It’s really hard for anyone from the outside to do a clean test. 

Timnit Gebru:
Right, exactly. Then we also had some Google executives on 60 Minutes talking about how the model learned language on its own without seeing it before. We had a senator tweeting out to his 1 million plus follower saying, this is so scary because chat GPT learned chemistry.  

I just find that the scientific community is being so unscientific and so irresponsible with all of these claims. And I find that terms like emergence, terms like foundation, terms like phase change, You know, phase change is like when a solid becomes a liquid because of, you know, heat or something like that. And just knowing the amount of experimentation and validation and theoretical foundations that go into making those kinds of claims in physics, for instance, right? And trying to make this kind of claim here is just so irresponsible, in my opinion.  

And in terms of this emergence claim, there was just this manuscript by Sanmi Koyejo and his students, right, who also analyzed a whole bunch of issues with all these papers claiming emergence, like for instance, very little few data points on their plot. How do you just plot something and you see something that looks like a line and call it emergence, you know?  

And so I currently don’t see any scientific grounding in any of these claims, so it’s very difficult for me to take any of them seriously.  

However, we know that such claims are incredibly useful for the corporations. Because first of all, look at how our legislators have been distracted by claims of superpower impending apocalypse, impending human extinction, right? That they are not talking about corporations. They are not talking about labor exploitation. They are not talking about data protection or corporations stealing data. They’re not talking about corporations not putting in enough resources to ensure that these models are safe. They’re not talking about transparency and asking them what data they’re using. That’s not any of the conversations that we’re having.  

We’re having conversations about some unknown machine entity destroying the the world. This is extremely useful for these corporations because instead of being regulated as entities, by entities that already know how to regulate corporations—like the FTC enforcing things, for instance—they’re kind of distracting people talking about how there needs to be some new entity that thinks about these existential risks and things like that, right?  

So now we’re not talking about humans being underpaid. We’re not talking about data being stolen. We’re not talking about privacy. We’re not talking about safety. We’re talking about these capabilities of these models.  

And the second way in which this is extremely helpful to the corporations is that it hypes up the capabilities of the model such that people want to– people think that it’s going to solve all their problems, and they want to buy it.  

I was just seeing a tweet by someone who said that their friend who was a doctor made the assumption that the diagnosis output by something like chat GPT must be accurate because it must have been trained to, you know, I don’t know, in some way, to be more accurate like that. And that’s incredibly dangerous.  

And ordinarily, it would fall under, I would imagine that it would fall under deceptive practices by corporations, right? Being deceitful about what their models can and cannot do.  

I was looking at the readme file or is it the API of chat and in Chat GPT, one of them. And I remember I saw it saying that it understands language. It understands code. It understand it didn’t even specify– first of all, even setting aside that it, quote unquote, understands language part. It didn’t even specify which language it, quote unquote, understands which one it doesn’t.  

This is so deceptive, in my opinion, because even if you argue about English or whatever understanding versus knowledge, this, that, we know it doesn’t even work at all in Tigrinya, let’s say my mother’s home tongue, right? 

Ethan Zuckerman:

Timnit Gebru:
That it doesn’t understand language. 

Ethan Zuckerman:
The paper of yours that has most recently just blown my head back and I’ve I’ve passed it on to a couple dozen people at this point is a paper called “The Lessons from Archives: Strategies for Collecting Social Cultural Data and Machine Learning.” You wrote it with Eun Seo Jo in 2020 and if I’m reading it correctly, you’re basically suggesting that one of the best things that we could do for AI right now is to think about the world as if we were archivists.  

What can we learn from the archivists that would help us figure out how to deal with some of the harms of AI?  

Timnit Gebru:
Exactly, and Eun Seo Jo is an archivist. She got her PhD in history, an archivist, she was also working natural language processing. So she could see some of the ways in which that training as a historian could help her in this case.  

And it was sort of what I did with “Data Sheets for Data Sets” where my training is as an electrical engineer. And so when I got into this field, I could see, some of the things I did as an electrical engineer how they could be applied here. So we know and this also goes back to the effect of altruists, right?  

A big thing about deep learning, what they want to do is they talk about end to end a lot. They want to remove, they want to feel like they’re removing humans from decision making. And so they’re like, we have data, training data, and the model, we have a model, and then we have predictions, end to end.  

And I remember in the last 10 years, when you read papers in deep learning, there were many of them were like, “Oh, we’re becoming more end to end. We removed the need to have this labeling. We removed the need to have this expert.” 

And to me, that shows a lack of understanding or maybe not wanting to understand or really or not wanting to realize, what is that data? Right? Where are you getting that data? And your model is basing things on that data that you’re giving it. And so what is that data representing? It’s representing people’s points of views, right? We say, you know, history is written by the victors. Of course, data is collected by the victors and the gaze and the people who are collecting data on behalf of, you know, something, someone.  

So in archives, from what I learned from Eun Seo Jo—right, this is all of the stuff that I learned from her—you have a curator, you have an institution curating things, and you are under no, you know, you don’t think that it’s quote unquote neutral. You’re under no illusion that this data is going to be neutral because you will see which institution is collecting it, which archivist is doing what their politics are. And they have a procedure from macro to micro by which they collect their data, right?  

And so she came up with this line from, you know, laissez-faire to interventionist data collection. And interventionist means you are, as a human intervening, you’re curating in the data collection process. And so, laissez-faire means you’re just taking what’s out there and making an assumption that it represents some sort of mutual thing about the world, which obviously we know that’s not true.  

And so then we also touch on, okay, now when you have people collecting data about other people, we have all of those issues. We have the colonial issues, we have the racism. We have the, you know, who is collecting data on whom and what kind of views are they representing, right? So all of those issues have been contested for many years. So we can see what kind of contestations arose and we can see how they try to address them.  

So for instance, we have an example, which is the Mukurtu platform, of indigenous people who are sick of their representation being represented by archivists in the way, not in the way that they would represent for themselves, right? Or maybe they don’t want to hand over data to various people. Maybe they want to annotate themselves in the way that they want to be represented.  

We’re gathering data. And data is about humans and humans archiving. That’s really what it is, because they want to represent something. They have an idea of what they want to archive. And so we can’t just pretend that this process does not exist, right? That the data is like some sort of, just represents some quote unquote objective reality. 

Ethan Zuckerman:
There’s two things that I really love about this paper. The first is this sort of observation that if you were putting together an archive, you wouldn’t just stuff everything into it. You would think really seriously about where it was coming from who was represented by it. The second is the fact that it opens AI discourse to discourse that’s been going on around museums and collections and history for decades.  

When I was working at the Berkman Center at Harvard, I used to go over to the Peabody Museum of Archeology and Ethnography, and this is a museum that’s been around forever. It was founded in 1866, and, you know, 20 years ago when I used to go to it, I would go and look at artifacts and the labels on the artifacts would not tell you anything about the people who’d created them. The label would tell you what white American or British colonial collector had collected the item. That was the thing that was most important was which explorer in a pith helmet had gone out and brought this item back. And obviously, that’s not okay.  

And we now sort of understand in the museum world, like that’s not okay. It’s probably also not okay to have that item. We probably have to have a conversation about who that item belongs to, where it goes back to. But we’re at least starting to have that conversation. We can’t have that conversation around AI, around large language models, until we do the work that you’ve been advocating, that we label, that we understand what’s going into these collections.  

Somehow though, you have managed not to, say, burn it all down. You’re generally not one of those researchers who says, just stop all this AI stuff. You often get associated with what people refer to as the harm reduction camp. Can we make these systems less broken, less biased, more transparent? What is it that’s giving you hope within this? what is it that makes you continue to be interested in AI, despite just frankly how badly we’re doing at this stage of the game? 

Timnit Gebru:
I think it’s because I see AI as a, I don’t even call it AI, I never, I guess I was in the AI lab, but my specialty was computer vision. And so I think of tools and expertise. And so I think of, you know, what tools to build and how to build them. And I don’t see it as inevitable that any tool has to exist, right? I put it in the context of the society that it’s in.  

So, the history of statistics is absolutely awful. It was based on eugenics, right? There was eugenicists who started it and that’s why they wanted to work on it. And so a lot of the ways in which they do things evolved with eugenics in mind. And so if you understand that, you can do something else instead and see if there are ways in which the tools of statistics can help. And there are things that are really essential in that field, right? That’s how I see AI.  

I think that there are some things that just should not exist and I don’t want them to exist. So in that sense, I would burn it down, right? Like I don’t understand why there’s certain ways, things that I don’t understand why they should exist, but there are other things that could actually be helpful to people, speech transcription, or there’s a number of other spell checkers, or just things, automated machine translation systems could be helpful, but we have to think about how they’re built, who’s building them, who’s going to be using them.  

So for me, it’s the knowledge that what makes me hopeful is the knowledge of understanding that it’s human beings who make these decisions in the first place. And human beings can say no, and we can have different processes for the tools that we build. And if we don’t want to build certain tools, we can just say no, and we can just not build them. And I don’t believe in any sort of inevitability, whether it’s technology or anything else. Some of your most inspiring work in my opinion is really focused on this question of who gets to build these machine learning systems.  

Ethan Zuckerman:
You’re the co-founder of Black in AI, which I know has been based on your own experience of being Black in AI. You’ve started this really remarkable project called the Distributed AI Research Institute, which I understand from our previous conversations, has something to do with trying to make sure that people other than people who are in Cambridge, Massachusetts or Silicon Valley get to be involved with this. Why is it so important who’s building these systems and what can we do to make machine learning algorithmic systems look more like the world? 

Timnit Gebru:
If we go back to our conversation about archives and archivists, right, and then it becomes super clear, who is collecting the data, who is, you know, And who are you asking about the data collection? Some people might say it’s stealing. And other the colonials that we were discussing might just think it’s not stealing, right? But the people who are stolen from will say this is stealing. So if you don’t talk to those people, and if they’re not at the helm, there’s no way to change things. There’s no way to have anything that’s quote unquote ethical.  

However, if those are the people at the helm, then they know what they need, and they know what they don’t want. And there can be tools that are created based on their needs. And so for me, it’s as simple as that. If you start with the most privileged groups of people in the world who don’t think that billionaires acquired their power unfairly, they’re just going to continue to talk about existential risks and some abstract thing about impending doom. However, if you talk to people who have lived experiences on the harms of AI systems, but have never had the opportunity to work on tools that would actually help them, then you create something different.  

So to me, that’s the only way to create something different. You have to start from the foundation of who is creating these systems and what kind of incentive structures they have, right? And what kind of systems they’re working under and try to model, I mean, I don’t think I’m going to, you know, save the world, but at least I can model a different way of doing things that other people can then replicate if they want to. 

Ethan Zuckerman:
She is Dr. Timnit Gebru. That is, I have to say, maybe the most hopeful I have felt about the field of artificial intelligence having talked to you, just because you’ve helped put things in such perspective and offer hope for ways that people can actually work to make these systems better. Timnit, what a pleasure. Thank you so much. 

Timnit Gebru:
Thank you so much for having me. This was a lot of fun.