Spotify<\/a>.<\/p>\nTranscript<\/h4>\n
Allison Ryder:<\/strong> What inspired a chemical engineer to work in program management for trust and safety? Find out on today\u2019s episode.<\/p>\nAngela Nakalembe:<\/strong> I\u2019m Angela Nakalembe from YouTube, and you\u2019re listening to Me, Myself, and AI<\/cite>. <\/p>\nSam Ransbotham:<\/strong> Welcome to Me, Myself, and AI<\/cite>, a podcast from \u9ebb\u7701\u7406\u5de5\u5b78\u9662\u65af\u9686\u7ba1\u7406\u8a55\u8ad6<\/cite> exploring the future of artificial intelligence. I\u2019m Sam Ransbotham, professor of analytics at Boston College. I\u2019ve been researching data, analytics, and AI at MIT SMR<\/cite> since 2014, with research articles, annual industry reports, case studies, and now 12 seasons of podcast episodes. In each episode, corporate leaders, cutting-edge researchers, and AI policy makers join us to break down what separates AI hype from AI success.<\/p>\nHey, listeners. Thanks for joining us again. Today, we\u2019ve got Angela Nakalembe, engineering program manager at YouTube, with us in the virtual studio. Angela, thanks for joining us.<\/p>\n
Angela Nakalembe:<\/strong> Thanks for having me, Sam. I\u2019m excited to be here.<\/p>\nSam Ransbotham:<\/strong> Typically, we start with some background, but it\u2019s really hard to imagine any listener that doesn\u2019t know about YouTube. If you\u2019re listening to a podcast, you probably know about YouTube. But still, can you give us a brief overview of YouTube and, in particular, what your role there is?<\/p>\nAngela Nakalembe:<\/strong> YouTube is an online video sharing platform, started about 20 years ago as an online community for folks to share videos on any topic that they were interested in. It got acquired by Google (Alphabet) shortly [there]after. And since then, it\u2019s grown into this behemoth of an online presence, with almost 2 billion, 3 billion users monthly and created such an incredible online community for creators to connect and also make a living as well.<\/p>\nWhat I do specifically at YouTube is I work behind the scenes within YouTube\u2019s trust and safety department, where, as an engineering program manager, I\u2019m responsible for overseeing a lot of the tools and feature launches \u2014 a lot of them [are] now using AI \u2014 that help keep the platform safe from things like hate speech, misinformation, and graphic content and the like. So it\u2019s pretty rewarding work. I feel like as we\u2019re in this ever-evolving world of AI, [as] we\u2019re seeing this technology get ever more integrated into our lives, it\u2019s really important that we find ways to not only use this technology to power our products and make it easier for us to send emails but also to find ways to protect the people who are forming online communities.<\/p>\n
Sam Ransbotham:<\/strong> You teed it up. What is happening with artificial intelligence [at] YouTube right now? <\/p>\nAngela Nakalembe:<\/strong> One of the ways we\u2019re using AI is to basically turbocharge our content moderation efforts and make it a lot faster but also a lot more humane. Imagine your job is to sit for eight hours a day, five days a week, reviewing videos filled with sexually explicit content or graphic violence or misinformation. That\u2019s the reality that exists for a lot of the human moderators YouTube and other online social media platforms use to moderate content and keep that violative content off the platform. And doing that day in, day out can have such a huge emotional and psychological toll on our reviewers.<\/p>\nBut now with AI and machine learning, we\u2019re able to really change that and evolve the role of the reviewer and have them work hand in hand, not just solo but hand in hand with these AI tools. Specifically, we\u2019re developing tools that can act as a first line of defense [to] flag or catch harmful content before it ever reaches a person\u2019s eyes, whether they\u2019re on the platform or a human reviewer.<\/p>\n
So while human reviewers are still needed as a step to kind of verify what these AI models are doing, they no longer have to carry the weight and that burden alone. And I think to me, that\u2019s one of the most powerful uses of AI. We are using AI [by] not particularly replacing people but helping protect them in a sense. Because now, we are reporting much higher reviewer satisfaction scores in terms of well-being.<\/p>\n
We\u2019re also receiving a lot of positive feedback from our users or creators who notice a much faster handle time for a lot of issues or cases that they\u2019ve reported. So things are getting taken down a lot faster, or things are not actually making it onto the platform, so people feel it\u2019s a lot safer. It\u2019s been incredible to see how we\u2019re able to use this technology to overall make it a much better experience for everyone across the board, both within YouTube and out in the world.<\/p>\n
Sam Ransbotham:<\/strong> We had a family friend who\u2019s a policeman, and he worked in the parts of the police force that dealt with things you didn\u2019t want to see. He talked about the turnover that job has, even though \u2026 they believe in the job. It\u2019s a rough job, and it\u2019s the kind of thing that they end up taking home with them inadvertently, even if they don\u2019t want to. They can\u2019t just forget [what] they\u2019ve seen all day.<\/p>\n[Now that] you mentioned it here kind of makes it sink in, this idea of you are in some sense replacing human workers, and that\u2019s a place where we really don\u2019t want humans to be doing something. Typically, with artificial intelligence, we think about dirty, dull, and dangerous jobs.<\/p>\nAngela Nakalembe:<\/strong> There\u2019s still an incredible need for these human moderators, right? There\u2019s a lot of nuance that AI just can\u2019t understand, you know? \u2026 I think AI is fantastic because it gives that first line of defense and tackles a lot of the more gory, very obvious instances of policy guideline violations, but then that leaves space for humans to handle the work that needs a little bit more nuance. So there\u2019s still need for human support there. <\/p>\nSam Ransbotham:<\/strong> At the same time, we were talking about the time spent curating content and you\u2019re not having a decreasing number of submissions. At the same time, we have artificial intelligence helping people produce content much more quickly. And so then you\u2019re getting much more content that you need to curate. But at the same time, I know you\u2019re having some struggles there with exactly what you do with AI content. Alphabet is working on developing these models, and at the same time, you\u2019ve got an interesting position of curating them. What\u2019s the YouTube take on that?<\/p>\nAngela Nakalembe:<\/strong> That\u2019s a really good point. I do want to underscore that one of the benefits of AI has been frictionless content creation. It\u2019s helping people express themselves a lot easier across language barriers as well, and scaling ideas that they previously weren\u2019t able to, right? But the downside of that becomes when anyone can generate high-quality text or video or images. Instantly, we\u2019re flooding the internet with a lot of content that looks legitimate but really isn\u2019t. <\/p>\nThat\u2019s a big thing that we\u2019ve been struggling with, not just YouTube but in general, with the rise in AI. A good example could be last year. [It] was a really big election season globally, and we saw an uptick in huge AI-generated videos that people would create about political opponents saying the most [rage-baiting] information you can think of to get a rise out of people and spread all that misinformation.<\/p>\n
We\u2019ve put [in] a lot of time at YouTube on thinking about how we can prevent the spread of misinformation and make sure people are more aware of the modality of creation of the content that they\u2019re consuming. A really good example I\u2019d want to call out here is at Google and YouTube, what we\u2019re working on is this big initiative called [the] C2PA alliance. It\u2019s the Coalition for Content Provenance and Authenticity, or C2PA [for] short.<\/p>\n
It\u2019s basically an open technical standard for publishers, creators, [and] consumers to establish the origins of whatever content they\u2019re observing. It\u2019s a group of tech companies from across the world: Adobe and Microsoft, a bunch of other companies are involved as well, and are committed to using this, and at least with YouTube, they\u2019re requiring our creators to validate whether their content was created with a camera or with AI.<\/p>\n
That\u2019s one of the ways we\u2019re mitigating it. We\u2019re also training our models to flag any misleading or low-effort or manipulated content a lot faster. The goal is really trying to restore context or what we call cognitive security, which is people\u2019s ability to be able to distinguish what\u2019s real or not in a world where it\u2019s becoming increasingly difficult to tell the difference. Back to the example I shared earlier, last year it was somewhat easy to tell the difference between what was AI-generated versus not in terms of those political videos and such.<\/p>\n
But fast-forward eight months later \u2014 Sam, there are some videos that I watch that [at] first glance, second glance I could not tell that [they were] AI-generated. It was very, very realistic, right? As the technology evolves, I think we as a company \u2014 Google, YouTube \u2014 are trying our best to keep up with that and make sure that we can find ways to help people distinguish what\u2019s real versus what\u2019s not.<\/p>\n
Sam Ransbotham:<\/strong> I\u2019m glad you mentioned provenance though, because I think, in general, not just [with] video, provenance is a big deal for any sorts of data that we\u2019re working with. Where data comes from, how it\u2019s used, what its source was, I think that\u2019s gotten much more difficult. It\u2019s wonderful that we have all this data and this video and content available, but at the same time, it has made it much harder to know where that came from. If I collected data myself, I knew where it came from, but if I used data from someone else, I\u2019m one step removed. And I think provenance in general is for everyone to start paying more attention to.<\/p>\nIn some sense, [with] such market power that YouTube has, a lot of the onus then comes to you, and what you do ends up becoming to some degree a de facto standard in the marketplace and, you know, not to put more pressure on you, but. \u2026<\/p>\n
Angela Nakalembe:<\/strong> None taken.<\/p>\nSam Ransbotham:<\/strong> The things you\u2019re doing are pretty important there. I want to switch a little bit. \u2026 Your title is engineering program manager, and you\u2019re actually an engineer. I always have to bring this out. I was a chemical engineer back in the day. You were a chemical engineer. <\/p>\nAngela Nakalembe:<\/strong> Yes.<\/p>\nSam Ransbotham:<\/strong> Tell us a little bit about how you ended up at YouTube. <\/p>\n