*This interview was conducted before Katherine Maher was announced as the incoming Chief Executive Officer of NPR.
KATHERINE MAHER:
I think what I would want to be focused on if I had the resources and capacity, and sort of a whole secondary set of careers, is really around how do you build out the infrastructure to allow for these spaces of discourse that are highly constructive and start with sort of a common set of facts and then build and construct better understandings or better truths around them. And then when you have that, what can you do to then bring that into, into public discourse in such a way that you’re perpetually expanding that circle outward.
REID:
Hi, I’m Reid Hoffman.
ARIA:
And I’m Aria Finger.
REID:
We want to know what happens if, in the future, everything breaks humanity’s way.
ARIA:
We’re speaking with visionaries in many fields, from art to geopolitics, and from healthcare to education.
REID:
These conversations showcase another kind of guest. Whether it’s Inflection’s Pi or OpenAI’s GPT-4, each episode we use AI to enhance and advance our discussion.
ARIA:
In each episode, we seek out the brightest version of the future and learn what it’ll take to get there.
REID:
This is Possible.
REID:
The internet, as we know it, is on the verge of a complete transformation thanks to humans equipped with AI. Historically, most of the information we have shared and consumed online is made purely by humans. But someday soon it’s possible that the majority of online content will be created by humans with generative AI. And then AI generated material will start to be trained and generated based on other AI generated material. How do we keep the internet as a reflection and amplification of the best of us — and steer clear of an internet devoid of human touch and nuance at best, and a swirl of misinformation at worst.
ARIA:
And as we head into an election, access to trustworthy information is particularly important. How do we prevent bad actors from weaponizing AI to influence elections or more effectively spread false narratives? How do we make sure the information contributed by AI is in fact true? And how do we help people trust that they’re true? Trustworthy information at scale is what we’re talking about today.
REID:
We usually have a hard time talking about how AI might impact society because there’s nothing else like it. But when it comes to information, specifically, there actually is an apt comparison: Wikipedia. You’ve probably used it and you probably more or less trust it. It’s important to talk about trust in this context. Can we trust an open source system? And if so, in what ways? The topic of whether AI models should be open sourced is still up for debate. And LLMs stand to be more complex engines than Wikipedia.
ARIA:
We want to dig into all of these questions, and that’s why we’re talking to today’s guest. The former CEO of the Wikimedia Foundation, Katherine Maher, led the team behind a resource that I think we all know pretty well — Wikipedia. Today, Katherine is CEO of Web Summit. She’s a longtime advocate for free and open societies, and has worked around the world leading the integration of technology and innovation in human rights, good governance, and international development. Now here’s our conversation with Katherine Maher.
ARIA:
Katherine, I’m so excited to have you here today. So I just found out that your mom, Ceci Maher, is in the Connecticut State Senate, which is just so awesome. And she is on the Energy and Technology Committee and that, you know, your career has been at the center of technology and advocacy and government. So, you know, what lessons do you take from your mom? Or do you see yourselves working on a project together in the future? I love that it’s all in the family.
KATHERINE MAHER:
Oh, yes, Ceci Maher is my hero. When I was 13 years old, she went back to school, got her master’s degree, started an entirely new career, ran a number of social service agencies, retired, and then was bored and decided that she wasn’t done working and decided to run for state senate. So I have a lot to learn and a lot to look up to. I think the lesson that I’ve learned from her over the years is that, A) reinvention is continuous and possible at all ages and at all times. You know, she once told me that when she married my father, her highest ambition was to be able to take the kids to the beach all summer long. And now [laugh], you know, she’s gone through this incredible transformation—which I think is sort of synonymous with the increased access to opportunity that women of her generation went through, the changing roles.
KATHERINE MAHER:
And I think that she offers a real role model figure in my life. The, as far as collaboration, I certainly give her a lot of unsolicited advice on issues such as child-safety bills, or really thinking about privacy and privacy regulation. We talk a lot about economic competitiveness and how one might drive that through sort of encouraging the tech sector. But no, nothing approaching lobbying. Really she’s an independent person and makes her own determination [laugh]. I think I’d love to, yeah, I’d love to collaborate if the opportunity arose, but I think the joke in our family is that it’s really more about me following in her footsteps.
REID:
Knowing some of your footsteps, that is an amazing recommendation. You recently met with Secretary of State Anthony Blinken to talk about AI and governance. Can you tell us a bit about that meeting? And, in your estimation, what are the top things for legislators and public officials to be actively considering as they take steps around writing and implementing AI-related policies?
KATHERINE MAHER:
Yeah, so I have the honor of serving on the Secretary’s advisory board, and we meet a couple times a year to talk about the future of technology and geopolitics. This most recent session, we are very focused on AI governance. I think the perspective that I brought to the table was really around: How do we ensure that we’re being competitive in a geopolitical landscape that is very much dominated by competition for regulatory primacy, or competition for, not exactly standards primacy, but perhaps model primacy? We, I think that it’s no secret to anybody who’s sort of paying attention to this: there’s a great interest, or emerging sort of rivalry, between China and the United States around dominance not only in AI, but in internet governance writ large. And that plays out in really interesting ways around Chinese—China’s policies around technology transfer for really fundamental infrastructure projects.
KATHERINE MAHER:
It plays out in ways in which China’s contesting the system of multi-stakeholder governance that’s been essential to underpinning internet governance at large. Some of the standards bodies that you probably haven’t heard referred to in years, but are incredibly important in terms of internet engineering, in terms of defining and evolving standards. China’s been a very active player in this space as well while in the West, we’ve perhaps not been as vigorously attentive as I would’ve hoped that we could be. And the reason that matters is—this isn’t just a question of, of engineering standards. This is really a question of dominant models for how the internet should work when it comes to things like privacy, when it comes to things like openness, interconnectivity, data transfers and data flows. And the model by which we’ve evolved the internet over the course of the last three decades or so is one that really reflects our own governance models in the West.
KATHERINE MAHER:
It’s open, it’s democratic, can be a little bit slow, a little bit frustrating. But when it works at its best, it’s bringing all the different stakeholders together from the private sector, from civil society, from nation states that are in, at their best, reflections of democratically-elected leaders. And if we don’t continue to invest in that, not only are we risking harm to the model, to the internet itself. We’re also—I think from a moral perspective or from a values perspective—receding ground away from this idea that democratic participation is in fact the best way to govern, even when it might be frustrating or slow. So, what should legislators be focused on? Well, I would love to see comprehensive data privacy regulation here in the United States. I think that that would be incredibly important in terms of clarifying what our, what our domestic technology policies are. And would actually allow us to have much more coherent policies when it comes to trade negotiations, when it comes to foreign policy conversations around tech policy, allows us to enter into these spaces with greater clarity about how and how we do not regulate technology. That, to me, would be the biggest possible win. So I’ll just leave it with my number one.
REID:
Well, quick follow-up, because, you know, obviously AI is on top of everyone’s mind, including mine. And I haven’t talked to Blinken.
KATHERINE MAHER:
No, you don’t say? [laugh].
REID:
Yeah. But, you know, probably should. I’ve talked to, you know, Gina Raimondo and Jake Sullivan and, you know, even Joe Biden and Kamala Harris about all this stuff. What do you think is the key thing in kind of AI policy regulation that you would add in that’s not kind of part of the common debate? You know, people worried about uncertainty, worried about economic competition between, you know, China and, and the US and chips and so forth. What would you add in kind of saying, “here’s another question or another thought that’s really important to be part of the dialogue”?
KATHERINE MAHER:
Absolutely all the things you just said. I think there’s something about false dichotomies as well around open, closed, risk, existential, et cetera. But I’m going to say something that people probably really haven’t thought of. The data shows, polling shows, that the vast majority of the world is actually quite optimistic about the role of AI. So when you look at global majority countries—the Global South, emerging markets, however you want to call them—people are extraordinarily optimistic about what this is going to do in terms of closing historic infrastructure gaps, historic access to economic opportunity, healthcare and the like. In the West, and particularly in the United States, we actually have the, on average, the lowest positive sentiment about the role and introduction of AI. I think that if we are not engaging with the sentiment that we’re seeing from the rest of the world on what opportunities this provide, and if we remain in an incredibly defensive posture, we actually risk the opportunity of being part of the transformation, being seen as an ally in this transformation.
KATHERINE MAHER:
And I know I keep bringing it back to foreign policy, but in an increasingly multipolar world, our legitimacy in dialogue and the persuasiveness of the democratic, liberal democratic model, depends very much on also seeing opportunity for growth, advancement in human development. And if we’re not engaged with people’s aspirations around these basic human development indicators like, “I want a job, I want to be educated, I want good healthcare, I want to provide for my family,” we absolutely risk losing the story of why this model of human rights, of democracy, of participatory engagement, is powerful in the world. Because we’re only seeing downside risk, not upside opportunity.
ARIA:
So I really love that perspective. I was just talking to someone yesterday, and I was saying that, you know, I don’t know what the future of AI will bring for the United States, but I am so positive that for three, four, five billion people around the world, the explosion of opportunity is just going to be immense. And I think, you know, you’re helping world leaders think about AI, but you’ve also helped shape how the world sees information. And I think right now everyone is talking about trust. Everyone’s talking about neutrality. You know, you were at the Wikimedia Foundation for seven years, the last five as their CEO. And with all of these hot button issues of the day, Wikipedia, you know, was seen as a place that actually got to neutrality. Is that even possible to be neutral, to be the source of truth? And is that even the goal? Is that what we should be striving for in these, you know, sort of crazy times?
KATHERINE MAHER:
I think it’s all about sort of fit for purpose. I think that Wikipedia plays an incredibly important role trying to present information that represents the best possible consensus about what is known or observable in the world. And so, should we be striving for neutrality? I would argue that those who seek to transform our systems to better humanity probably shouldn’t, right? Neutrality actually is a thing that has a place and a time. And at times we need disruption, challenge change. But I also think that we need to be able to go back to sort of first principles at times, or go back to context. And that’s where a place like Wikipedia can be so tremendously valuable. I often make the differentiation between what is truth and what is known—or what is truth and what is observable fact and feeling. I think that some things can absolutely be true, and they can also be factual, in the sense of: we have information about who was present on a certain date at a certain time. You know, who was in the room.
KATHERINE MAHER:
We also have truth that is highly debatable or in a process of evolution, in the sense of understanding world historic events. We might see that who was in the room from the, and have that as a common basis of fact. But depending on who we’re aligned with of those participants, we may see the truth and outcome is quite different. What I appreciate about what Wikipedia has always done is it started from sort of citable—citable, observable fact, and then built out truth. Now that truth, the advantage of a system like Wikipedia, is it’s continuously contested. And the more representative the contributions are to Wikipedia, the more global they are in scale, the more accessible participation is to a broader and more diverse range of people, the more truth will be in the room, in addition to fact. And it allows us the opportunity to understand that it’s not a dominant truth from, you know, the Global North.
KATHERINE MAHER:
It’s not a dominant truth from the English-language communities. There’s actually multiple truths. And this is, I think, what allows us to get to better empathy and understanding. And in the sciences, this is of course true as well because—true [laugh]. It’s of course also the case that what is observable today may, we may have far greater nuance and insight tomorrow with increased tooling, increased capacity, increased AI support and the like. And so what Wikipedia represents is always a snapshot of what is known and what is agreed upon based on who is in the room. And part of my mission was always to expand who is in the room, so we have a much better understanding of what is known.
REID:
Well, this gives a natural follow-up question—which is one of the things I’ve actually been very curious about and we’ve not had a chance to talk about—which is: What do you see as the intersection between the future mission of Wikipedia and generative AI? There’s obviously a huge amount of human amplification that’s possible here, but also, of course, there’s anything from human agency to also hallucination, and what does the navigation of that look like, possibly?
KATHERINE MAHER:
Yeah. Yeah, who knows, right? But [laugh], a couple different thoughts. I mean, one is: we’re very fortunate Wikipedia exists as part of the training data set [laugh] of AI, gen AI. I think we’ve known for years that Wikipedia is utilized for this purpose. In fact, we, I think the earliest use case that was most visible to most people—even if you didn’t know that Wikipedia was a part of this—was actually, was translation tools. Wikipedia represents 300 different languages, is the largest natural language dataset of many of those languages in a digitized format for general knowledge. And so we were used very widely as the training dataset to develop a lot of these tools that are in wide use today and continue to improve. These days, Wikipedia’s still one of the largest training data sets. The Washington Post had a pretty good breakdown of one of Google’s models around this.
KATHERINE MAHER:
And what I think is interesting about that—and it comes back to this sort of question of who’s in the room—is this question of where bias exists. And one of the challenges that I’m immediately sort of attuned to is—we know Wikipedia is horribly biased; I’ve said this in many of my prior interviews, it’s not a shock. It’s, and that has to do with historical factors. It has to do with who participates. It has to do with what is actually citable information, what has been documented over time, in the sense of what has been written down in books, what has been deemed to be noteworthy knowledge. Tremendous gender, ethnic, geographic, colonialist bias in all of those questions. But the beauty of Wikipedia is that all of that is open and at least observable and also alterable. When we start looking at how that plays into these training datasets, first of all, there is often a lot of obscurity as to how much, or what percentage, or what weights, are afforded to this information, to this data.
KATHERINE MAHER:
Many of these systems are not continuously updated in ways that allow for edits or alterations in more of a public and transparent fashion that would allow us to adjust for that bias in real time. So they are that closed-loop sort of system of being able to identify and then correct is not available to the public in the same way that something like Wikipedia actually is. So that’s just sort of how do we, what are the lessons that we’ve learned from the compilation and construction of information? How can that go into the construction of more representative products coming out of large models? In terms of Wikipedia itself, I think there’s a very interesting question about, well, will AI write Wikipedia? But that question has always kind of been around, and we’ve always used automated tools to write and monitor Wikipedia.
KATHERINE MAHER:
The very first ones were built around identifying typos, then it was spam fighting, and then it was translation tools, and now it’s automated construction of article structures. And so I am far less concerned about what does this mean for Wikipedia, because I think the primary value set of people who write it is not to do it more efficiently. It’s to do it because they take joy in the act of information construction. To take joy in the act of negotiation of facts and representation of information. And they will use AI to help with that process, but ultimately it will still be a human-edited process. And I think that that is what will make Wikipedia continue to be viable and beloved for its humanity, as it always has been.
ARIA:
Yeah. I mean, I think the point you make is a good one about—you’ve constantly been talking about the, you know, the biases that Wikipedia has. And at the most recent Web Summit, you talk to the CEO of Unbabel about how we have to have generative AI training on non-English sources. You know, otherwise we’re just going to continue this. And you mentioned sort of the opacity of the large language models. What are the lessons that generative AI can take away from some of the strides you made at Wikipedia? Or is there a way that AI could be constantly contestable in the same way that Wikipedia is—like, what can we do? Because we know AI is good and that there’s so many positives, but we also want to always get better when it comes to, you know, bias, discrimination, that kind of thing.
KATHERINE MAHER:
I think that there, this is part of this whole question about what are the obligations around disclosure of datasets? What are the obligations around discernability? I think this is part of the false dichotomies question as well. One of my good friends is the person who like, who wrote Mozilla’s open license, right? And so this is a real like OG open source thinker. And we’ve talked a lot about, is it even possible to have open AI in a real sense? And I think, so this is Luis Villa. He’s a wonderful person. Look, he’s got a great newsletter. But we, he sort of pointed me to I think Martha Mitchell, whose testimony in front of Congress—which really focused on are there attributes of openness such as discernibility, such as scrutiny that we can bring into the, into questions and expectations about regulation and governance that can give us at least a fundamental baseline for how we might have a value of tools to be able to continue to update for bias. And I know that different places are taking different looks at this and there’s a lot of experimentation around it. I think for someone like myself who is, sits on more of the governance side of these questions, my encouragement would be that this needs to continue to be a priority and publish the results so that we can actually determine what’s working and what isn’t.
REID:
Well, speaking of publication, you’ve been a big advocate for citation, especially as a, you know, way to build and retain trust, you know, the way people have for Wikipedia. What do you think about citation here in generative AI? Like, you know, giving due credit and, and you know, what, given that the volume of things that are going to get published on the internet with AI-assisted, AI-generated is going to be [laugh], you know it’s going to be a quasi-exponential increase.
KATHERINE MAHER:
There are probably three different ways to think about this, right? One of which is a citation to copyrighted material. One of which is citation of critical information — so that might be in a medical or a scientific context. There are probably other contexts as well, that’s just the first thing that pops up in my mind. And then the third is sort of the creative construction that is based on open materials. I feel as though the lesson that we learned from Wikipedia is very, very few people engage with the citations [laugh]. But, but the fact that they are there allows people to do so and allows people to get engaged in this sort of confidence assessment as to where this information comes from. And so you might be the sort of person who checks the citations once at the beginning of your experience with Wikipedia. And if you are satisfied that that is roughly correct and roughly correctly attributed, you may then have a sort of a confidence transfer to your future engagements with the product.
KATHERINE MAHER:
You may be the sort of person who never checks, but in reality, we know about, in user-generated content in general, you only need a tiny percentage of people who are actually engaging with this process of sort of at attribution validation in order for that to work well. Further, citations within Wikipedia provide their own sort of graph that allows us to have better understanding of where information is being produced. It’s sort of, think of it as a equivalent to within academia, people look at, you know, what papers have the highest number of citations, and sort of how does that influence overall research and development? Within Wikipedia, we can do the same thing. We can isolate where most information is coming from, what are considered the most reliable sources, and then we can draw inferences around that, or we can direct research to areas where we have under-researched questions.
KATHERINE MAHER:
All of this is essential to how we construct information and has inherent value, in addition to sort of the validation value. Within AI, I’m very much of the opinion that there are going to be applications where having citations are critical. I’m actually on the board of an organization, a company that is doing work in this area around medical research, and that has basically built a model off all publicly available scientific and medical research. And we have an average of 36 citations per synthesis. So it is quite possible to do this work if you prioritize it and if you’re thinking about where your training data is actually coming from. And arguably, in that space of medical research, that’s exactly where you need to have that high fidelity of research because it tells you temporality, it tells you accuracy, it tells you scope of, of the research, applicability, it tells you the diversity of, of subject, of subjects in that research, et cetera, et cetera. Again, there are places where that’s not necessary at all, such as: What are a good set of parks to go hiking in Vancouver when I’m there for a weekend with my family? I mean, maybe the Vancouver public municipality wants us to have that, but from a, from a sort of critical citation purpose, it’s a little bit less relevant.
ARIA:
You know, these days it’s a common refrain that we’re so polarized politically, like no one can see, you know, either side of the equation. How has the internet impacted this sort of trust in a global context, and is there a way to rebuild it?
KATHERINE MAHER:
I think the internet has been catastrophic for trust, but perhaps not for all of the reasons we might think. It’s so complicated, isn’t it? And I—complex might be a better way of saying it. When I say catastrophic, I think that what has happened is not that the internet has destroyed trust. It is that the internet has surfaced fissures within systems and, and allowed them to grow, and grow publicly, at an exponential rate. And so when, for example, we think about trust in institutions—which is something that I’m very interested in—and it comes to sort of institutional governance, institutions fit for purpose, and I mean institutions in the abstract and in the literal sense—so the institution of universal suffrage is an institution, although it doesn’t have like a brand name—the issues that we’ve seen there is that many of these institutions were built around a sort of homogenous population that they were serving.
KATHERINE MAHER:
They were not terribly responsive, both in terms of accountability to that population, and then, when we started to see increasingly heterogeneous populations—due to immigration, diversification, civil rights movement, et cetera, et cetera—we started to see how those, those institutions were, were not actually sort of fit for purpose. What the internet has done is it has exposed those fissures in ways that are related to both. We now interface with all sorts of technologies, platforms, and services that are hyper-responsive to our needs and have created an anticipation of a much more frictionless, much more productive set of processes, services, outcomes, SLA, functionally human SLAs. We’ve also seen that we now have the ability to reckon in real time with, with examples of where those gaps actually are, which make the failures of those institutions even more evident to a greater number of people.
KATHERINE MAHER:
That, to me, is sort of the primary issue relative to trust in this day and age is not really around, “oh my goodness, my algorithm is serving me information that I can’t trust.” It is actually around a set of expectations about how institutions should function in our lives, and our, and where those institutions are falling down, as responsive institutions that are consistent and accountable in their purpose and effective in their service delivery. Absolutely then there are questions of, “Well, why is the media not trusted? And, you know, what does social media have to do with all of that?” And I think that those are useful questions, as are the questions around what is AI going to do to our, like, the construction of false information? But I think that they are secondary questions to this primary one around what has the internet shown to us about the institutions that have historically governed our nations, our lives, and where are they not fit for purpose? And I think that this is actually a really important—an essential question—if, for those of us, like myself who believe very foundationally that institutions should be A) responsive to all people, but that institutions are perhaps the most important part of stable, democratic, rights-respecting, representative governance.
REID:
Well, I’m curious what you think we should be doing vis-à-vis internet, social media, AI, et cetera. Because, you know, one of the things I think is funky about trust when it gets to institutions and society is: without it functioning, things break. Like, trust only works in systems of money because the belief is that next day, next week, the money system works. Trust only works with system of law because people believe the majority of people are going to follow the law, and that there’s a reason to do it. But part of the thing that obviously gets to a challenge is when you have kind of anti-institutionalists being fractionalists, it, you know, they are, they are themselves creating harm by the decay of trust in this kind of key ongoing function. A belief in the truth of discourse, and a belief in the fact that, you know, the 2020 election was as fair as, you know, any election in our lifetime. You know, and these kinds of things, it’s like important to work towards it. And so the people who attack the institutions, I’m always worried about saying, “We should be renovating, not destroying. And we should be building trust.” Because it’s really important. Do you have any, you know, reflections, thoughts, you know, suggestions, questions for this, kind of, how do we regain trust given that we’re technologists playing in this arena?
KATHERINE MAHER:
You know, it’s that whole, it’s the whole sort of, if it was easy, we would’ve figured it out by now [laugh]. I was just having a conversation with someone about this yesterday, about the role of something like a citizens’ assembly.
REID:
Pi, could you explain?
Pi:
A citizens’ assembly is a randomly selected group of ordinary citizens who come together to deliberate on a particular issue or policy. The goal of a citizens’ assembly is to give everyday people a voice in the decision-making process and to ensure that policies are informed by a wide range of perspectives.
KATHERINE MAHER:
This is a very useful model for addressing questions of values, as opposed to sort of questions of how we legislate on a particular policy. These citizens’ assemblies, what often ends up happening is, despite the diversity represented in the room, individuals, when faced with one another, are often able to find common ground on highly divisive and contested issues and produce outcomes with fairly significant nuance attached to how those issues are navigated. The reason I bring this up is I think that, of course, you’re absolutely right that there are the anti-institutionalists who operate on both the long and the short arc, in terms of defunding public institutions, which is something that’s been happening over a multi-decade period. It’s the whole pull-the-ladder-up -behind-us, sort of after the post-war economic boom. Great. Everybody’s got theirs. Now we’re going to defund and deconstruct all the institutions that made all that possible.
KATHERINE MAHER:
None of that is new to you. Then there are the short-term anti-institutionalists who use the tools available of broadcast media, of social media, of these highly accessible, highly niche communications tools to be able to continue to perpetuate that message. I think what I would want to be focused on if I had the resources and capacity, and sort of a whole secondary set of careers, is really around: How do you build out the infrastructure to allow for these spaces of discourse that are highly constructive and start with sort of a common set of facts, and then build and construct better understandings, or better truths around them? And then when you have that, what can you do to then bring that into public discourse in such a way that you’re perpetually expanding that circle outward? So that type of discursive understanding is part of the—of the body politics a whole.
KATHERINE MAHER:
And there’s obviously precedent to this. Like, I’ve come from New England and the, you know, the New England sort of assemblies in the old town hall are famous for, for the fact that this is what they would do. But the same is true. This has been happening in church basements, in grange halls, all over our country for years and years and years. And happens in various different forms all over the world. And so I think that, you know, we talk about this being a question of technology. Maybe there are ways for technology to mediate this. But I actually think technology has pulled us away from some of these tried and true, tested tools of engagement that are, are quite productive, quite constructive, and actually lead consistently to better outcomes time and time again.
REID:
Yeah, though I think one of the things, I agree with you; the technology has helped contribute to some of the alienation that people have—you know, people are in the same room. That’s positive. But one of the questions also is how we navigate with scale, right? You know, and so, you know, part of the reason I’m also very curious—and I agree with you entirely; if it was easy, we’ve done it already—but how do we use technology to bring back that, that almost like, “okay, I’ll, I’ll give you a benefit of a doubt if you’re in a room”? Because I do believe most normal people do that. I don’t think House of Representatives people do that. But I think most—
ARIA:
Except your mom, of course, to be clear.
REID:
Yes. But, you know, I think it’s really important, because what people don’t realize is that trust in the institution, trust in how we function—that doesn’t mean that don’t be critical—but that trust in it is critical for the functioning of our society, let alone the functioning of our democracy. You’re one of the thinkers, leaders, and experiencers on this, and I was wondering, I was like, “Ah, let’s, let’s see if there’s, there’s a silver bullet.” And the answer is, “Nope, lots of work.” [laugh]
KATHERINE MAHER:
No, I mean, you know, there—yeah, the answer is always lots of work, right? That is just the answer to doing good things in the world. But I think that there’s sort of couple things there that I just want to pick up on. One of which is—of course the answer is not don’t be critical. At Wikipedia, I always used to talk about the loyal opposition. And the loyal opposition are loyal because they believe in what you’re doing. They’re opposed because—or what the broader project is, they may not believe in what I’m doing; that’s why they’re the opposition—but they’re loyal to the broader project. And they are opposed because they believe that there is a better way to do it. And it is absolutely essential to being able to govern or to engage or to support or steward a project like that forward to engage with the loyal opposition.
KATHERINE MAHER:
If they didn’t have a legitimate concern, they wouldn’t have stuck around to express it is very often the case. Maybe the way it’s expressed is not the most constructive. I mean, there’s all sorts of things that, that I’m sure—and this is, and I’m deliberately drawing a parallel to our political discourse right now. Maybe the way it’s [laugh] expressed is not the most constructive. Maybe there is sort of the heckler’s veto component piece to it. But if you don’t engage it, it will always be the heckler’s veto on the outside as opposed to having conversations that weigh and debate trade-offs with candor, make determinations and decisions with transparency about the factors that go into them, and continue to engage people with respect throughout that process. I think the thing that Wikipedia taught us, to come back to this scale piece, is that nothing at Wikipedia happened at scale.
KATHERINE MAHER:
That the whole idea of Wikipedia is anti-scale. Yes, it is huge, absolutely, but it is all micro-communities of discourse. And that means that the people who work on things like 16th-century Islamic art are not the same people who are working on breaking news events. Not by, well, maybe there’s one or two, but, but not by a large, you know, a long shot. And that’s okay because that allowed for expertise and energy to flow into the places where people cared about it the most. And then you have people who work across the entire system to think about continuity of governance, continuity of policies and principles, applications of policies, both in a corrective and sort of constructive way. So corrective would be like a code of conduct, for example. What brings everybody together is a common set of, of sort of an intrinsic common set of values or common ambition, with some extrinsic markers around that as well.
KATHERINE MAHER:
Everybody wants to be right on the internet, right? So I think to the extent that we want to harness a way of engaging at these debates at scale, you both have to have an intrinsic understanding of what it is that we’re here to do. What’s the value proposition? What’s the purpose? What’s the motivation? Maybe the problem is, is that democratic governance is failing to articulate its value proposition to the vast majority of the American public right now. And failing to articulate the value proposition of participation in that process. And an extrinsic motivation, which is: you get the outcome you want. And you’re not going to get the outcome you want all the time. You’re going to get the outcome you want part of the time. But you’re going to get it frequently enough that you’re motivated to continue participating in the process. And I think that that is how you scale these kinds of conversations, is you don’t try to scale them from the center outward. You try to replicate them ad infinitum, to the extent necessary to be able to cover comprehensively all the issues and geographies and representative identities that you need to engage with.
ARIA:
Yeah, I mean, it’s interesting. So you said in a 2012 interview that governments have increasingly become aware of the way that the internet is a powerful tool for dissent in terms of political organization and individual freedom. And you know, you were just talking about that loyal opposition. And I think we all want the loyal opposition. It makes us better. If no one criticizes our companies, our governments, our organizations, you know, we’re never going to move forward. I think I get particularly frustrated in the, you know, in shit posters—the people who aren’t actually trying to, you know, make things better, whether it’s in an organizational, corporate, or government context. How do we ensure that the internet is sort of a positive place for dissent, and that, you know, it does keep people accountable, but that it doesn’t sort of devolve into the bad actors? And like, how can we make sure that that’s still true in the age of AI? Any thoughts?
KATHERINE MAHER:
Again, this comes to, well, first of all, the question of what a shit-poster or dissent actually is is again, highly contextual. And I don’t want to keep prevaricating and say, “well, it all doesn’t, you know, it’s all context, right?” But the reality is that when it comes to normative expectations of expression or privacy or what harassment constitutes or any of these sorts of fundamental questions of how we conduct conversation with respect or constructiveness, we have very different expectations across the board as to what the parameters of that are, right? When you think about that expansion of—like, I’m a human rights person; I go back to the universal declaration of human rights, right? And those rights sort of come back to things like, well, we have the right to expression, we have the right to assembly, we have the right to privacy, et cetera, et cetera.
KATHERINE MAHER:
But those rights grew out of highly localized norms and are differently expressed everywhere you go. So I always use the example of privacy in the Netherlands versus Germany. I lived in Germany for a while. I’ve spent a lot of time in the Netherlands. If you go to Germany, people pull their curtain shut, and they do so because they both have internal, sort of, domestic privacy, but also because it is perceived culturally that exposing your internal workings to the world is actually a bit of an affront to those passing by on the street, right? Or your neighbors across the way. I remember being, having a picnic on a balcony, and someone pulling the, you know, sun blinds down in order to not disturb the people across from us on a Sunday, because we wanted to give them their privacy in the house across the way. In the Netherlands, privacy is also tremendously important.
KATHERINE MAHER:
People keep their curtains open. Because the cultural expectation and norm is that if you have something to hide—or if you close your curtains, you’ve got something to hide. That that is somehow deficient relative to your participation in the broader society. And yet there are both incredibly strong norm—there’s incredibly strong norms of privacy in both of those countries. And you can see that in the way that they approach data protection, for example. So there are different applications or expressions of privacy in two countries that are side by side that share quite a lot of common history. And so I say that to say: establishing norms for what we want on the internet is super hard. Because all of our norms, our normative values grew out of our, our local town. They grew out of the conversations that we had around the, you know, the town green, or the town square, or the, you know, the town council, so to speak. The expectations of expression, the expectations of the limitations on that expression, et cetera, et cetera.
KATHERINE MAHER:
How you then scale that to a five-billion-person network—we have no idea how to do that. So [laugh], how do we create space for dissent within those models? Well, you know, I’m a very, very big believer in the ability to have clear codes of conduct and policies for spaces online. It is—while I support the expectations of unlimited rights within the context of a rights-based framework offline—the sorts of spaces that we have online are scaled. They are private. And we need to be able to uphold expectations around policy—expectations around discourse or norms that are appropriate to those communities and to the purpose we’re trying to seek. So Wikipedia had very clear policies, for example. Other platforms have very clear policies. Even, you know, subreddits have their own policies. That’s what allows for constructive discourse. It’s also what allows for objection, dissent, friction in the process without it ever crossing—well, it does sometimes cross the line. But when it crosses the line into something else, into something that is antagonistic, that becomes harassment, that becomes incitement, it gives us the ability to, to either block or stop that behavior or call that behavior back in. And so I think that again, this comes down to this question of you can’t do it at scale. You have to do it in a way that is replicable and consistent to the communities and to the purpose that you’re trying to achieve.
REID:
One related question that you just reminded me of is: What do you think the rough swag—of how we should think about data and training these AI models? Obviously Wikipedia has a, has a, has a philosophy of openness and availability to everybody. But obviously there’s a whole mix of data on the internet, and yet the scale of data is really key and the creation of the AI stuff can be a public good in various ways. So what’s your—you must have had some thoughts, and it’s one of the things that I’ve been meaning to put my thinking cap on and kind of scratch my own head about.
KATHERINE MAHER:
The question is: How do you do this in a way that adds value? And so I’ll just give the Wikipedia example because why not? But I think that it offers potential directionality. I differentiate between types of data. Obviously tprivate data, like health data or communications data. And I, as you know, I’m on the chair of the board of Signal, the messaging app—very much focused on data protection, very much focused on metadata protection in addition to the content of your conversations. That to me is the sort of data that it should absolutely be about consent to disclose. Health data: consent to disclose, to participate. I know that there’s some really interesting efforts out there to really think about how one can do sort of private model or privacy-centric model training. And I’m curious to see where those efforts go.
KATHERINE MAHER:
I know that there’s challenges and weaknesses to those different architectures, but nonetheless, it’s good to see, good to see progress moving towards that. Where it comes to things like content, and I know that this is the big—content isn’t even the right word. The constructed news, art, published thoughts, blogs, Wikipedia on the internet, long form text, short form text there; what we did was we looked at this, and we said we know that there is an inherent value for companies that want to use this data to train their models. We know that they’re creating tremendous value out of this. We want to be on the receiving end of some of that value so that we can continue to perpetuate the value that we create, because in order to contribute to that net advancement, we’re going to need to continue to be a viable going concern, right?
KATHERINE MAHER:
Models trained on models are not demonstrating really, sort of, the high-quality outcomes that models trained on knowledge are demonstrating. And so—or vetted, researched, peer-reviewed, et cetera—which is the sort of work, edited, which is the sort of work that Wikipedia creates. So what we did was we built out a set of APIs that we were then able to offer for a nominal fee for companies to take advantage of. Now all of our content is free. They could have just taken it. But when we were able to offer them a means by which they could access that with consistency, transparency, accountability on our end relative to the maintenance of those APIs—that was a model that worked that was a net benefit to both the companies that wanted access to it and to Wikipedia, in terms of creating additional income for us to continue to be able to support our sites.
KATHERINE MAHER:
We weren’t selling private data. We weren’t even selling freely licensed content. We were selling a model of support that made it easier for people to do the work that they needed to do on their computational backend. And so for them it had value, and for us it had value. And I’d love to see more of those sorts of win-win type approaches. I think there is this instinctive lesson that we learned from a decade ago, which is we throw up a paywall and that’s the solution. But I don’t actually think that that is the long-term solution set. I think that there is a reasonable business model to explore there. But I think that there are other reasonable business models to explore as well that allow for the continued health and evolution of core industries that we feel are tremendously important—like the creative industry, the media industry, and the like—while also allowing for us to continue to build models that are more representative, and frankly give us a lot more control about what data goes into them. And a lot more accountability as well as to how we think about, again, representation, diversity inclusion, and the like.
REID:
So we’ll now move to our rapid-fire section. Is there a movie, song, or book that fills you with optimism for the future?
KATHERINE MAHER:
So, if I didn’t work on these issues, I wish that I worked on climate, because that, to me, is the existential risk of our time. That’s the, that’s the real existential risk I’m worried about. A friend of mine, Dr. Ayana Johnson, has a book coming out called What If We Get It Right?, where she goes and she interviews a number of individuals about what it would look like if we actually are able to reduce anthropogenic global warming and how that might lead—what our society could be. And I think that that is—it’s available for pre-order. I’m really excited for it to be available to the public as well.
ARIA:
So what is a question, could be personal or professional, that you wish people would ask you more often?
KATHERINE MAHER:
I think it’s a question we should all be asking each other more often, which is: How can I help? How can I make your life better today? How can I make your life better in general? What do you need? And I think if we asked that question more often, we’d all get more comfortable answering it too.
REID:
Yeah. We are—life is a team sport, and so that’s a very important thing to do. So where do you see progress or momentum outside of your industry that inspires you?
KATHERINE MAHER:
I started my career at UNICEF and I worked on public health services. I had, I was there at the beginning of sort of the mobile telephony revolution, where all of a sudden everywhere in the world was getting access to mobile networks. And that created unprecedented opportunity for what we then called service delivery. So, education and healthcare. And what I think we are not often aware of is how transformative global health efforts have been in terms of elevating people’s quality of life overall. I specifically worked on preventing the transmission of AIDS or HIV from mothers to children during childbirth. And if you do that right, you can prevent 99% of transmission and have an AIDS-free generation, which is absolutely spectacular, not just for that individual’s life, but for the health of that economy, the opportunity that that then creates within that population set, what that means in terms of the redirection of government resources into actually long-term durable solutions for people’s wellbeing. So I still look to public health and think it is one of the most inspiring, noble, consequential places that people can work. And when I get down about the world, that’s where I turn my gaze.
ARIA:
Can you leave us with a final thought on what is possible to achieve in the next 15 years if everything breaks humanity’s way, and what’s the first step in that direction?
KATHERINE MAHER:
Well, one thing that I think is incredibly cool—and this is one of, like, a bunch of things, but since it, it ties back to earlier themes that we’ve been talking about—is, I was talking to someone the other day who was underscoring the fact that for the last decade we’ve been saying every child needs to learn like foundational STEM skills, and what they really need to learn is how to code. Because learning how to code is the ability to have control over the future. And I know that like there’s arguments that with increasing modularity of services and things like that, that’s no longer the case. But, but what is so exciting—and this sort of comes back to what we were talking about, about the opportunity of AI—is the fact that there are now tools that are teaching people how to be able to control the development environments around them.
KATHERINE MAHER:
And so you actually may have a generation in the next 15 years that is so much more computer literate and has the capacity to build the solution sets to their problems in the world in ways that were previously inaccessible, because you just didn’t have the teacher showing up in the classroom or on the Zoom. And now you have the compute power on every device so that every kid all over the world can learn these skills and build solutions to the things that they need. And like that is, that is the promise, right? That is the promise of the computer revolution. That is the promise of the growth of the internet. That to me is how we could get it right. That gives real power to everyone to have real agency over the things that are important to their communities. And that would be super cool.
ARIA:
I love it.
REID:
Possible is produced by Wonder Media Network, hosted by Aria Finger and me, Reid Hoffman. Our showrunner is Shaun Young. Possible is produced by Edie Allard, Sara Schleede, and Paloma Moreno Jiménez. Jenny Kaplan is our executive producer and editor.
ARIA:
Special thanks to Katie Sanders, Surya Yalamanchili, Saida Sapieva, Ian Alas, Greg Beato, and Ben Relles. And a big thanks to Katherine Farrell, Jenny O’Donoghue, and Little Monster Media Company.