WEBVTT 00:00:00.000 --> 00:00:00.616 [SPOT] 00:00:00.616 --> 00:00:04.440 GRETCHEN HUIZINGA: Hey, listeners. It’s  host Gretchen Huizinga. Microsoft Research   00:00:04.440 --> 00:00:08.960 podcasts are known for bringing you stories  about the latest in technology research and   00:00:08.960 --> 00:00:14.480 the scientists behind it. But if you want to dive  even deeper, I encourage you to attend Microsoft   00:00:14.480 --> 00:00:21.000 Research Forum. Each episode is a series of talks  and panels exploring recent advances in research,   00:00:21.000 --> 00:00:25.560 bold new ideas, and important discussions  with the global research community in the   00:00:25.560 --> 00:00:30.880 era of general AI. The next episode is  coming up on June 4, and you can register   00:00:30.880 --> 00:00:40.082 now at aka.ms/MyResearchForum.  Now, here’s today’s show. [END OF SPOT] 00:00:40.066 --> 00:00:45.240 [TEASER] [MUSIC PLAYS UNDER DIALOGUE] ABIGAIL SELLEN: I'm not saying that we shouldn't  take concerns seriously about AI or be hugely   00:00:45.240 --> 00:00:51.080 optimistic about the opportunities, but rather,  my view on this is that we can do research to get,   00:00:51.080 --> 00:00:56.880 kind of, line of sight into the future and what  is going to happen with AI. And more than this,   00:00:56.880 --> 00:01:01.720 we should be using research to not just get line  of sight but to steer the future, right. We can   00:01:01.720 --> 00:01:09.084 actually help to shape it. And especially being  at Microsoft, we have a chance to do that. 00:01:09.084 --> 00:01:11.000 [TEASER ENDS] GRETCHEN HUIZINGA: You’re listening to Ideas, 00:01:11.000 --> 00:01:16.440 a Microsoft Research Podcast that dives deep into  the world of technology research and the profound   00:01:16.440 --> 00:01:29.280 questions behind the code. I'm Dr. Gretchen  Huizinga. In this series, we'll explore the   00:01:29.280 --> 00:01:35.903 technologies that are shaping our future and  the big ideas that propel them forward. [MUSIC FADES] 00:01:35.920 --> 00:01:41.440 My guest on this episode is Abigail Sellen,  known by her friends and colleagues as Abi.   00:01:41.440 --> 00:01:45.960 A social scientist by training and an  expert in human-computer interaction,   00:01:45.960 --> 00:01:50.800 Abi has a long list of accomplishments and honors,  and she's a fellow of many technical academies and   00:01:50.800 --> 00:01:56.800 societies. But today I'm talking to her in her  role as distinguished scientist and lab director   00:01:56.800 --> 00:02:02.960 of Microsoft Research Cambridge, UK, where she  oversees a diverse portfolio of research, some of   00:02:02.960 --> 00:02:08.680 which supports a new initiative centered around  the big idea of AI, Cognition, and the Economy,   00:02:08.680 --> 00:02:14.784 also known as AICE. Abi Sellen. I'm so excited  to talk to you today. Welcome to Ideas! 00:02:14.784 --> 00:02:15.900 ABIGAIL SELLEN: Thanks! Me, too. 00:02:15.900 --> 00:02:21.560 HUIZINGA: So before we get into an  overview of the ideas behind AICE research,   00:02:21.560 --> 00:02:26.760 let's talk about the big ideas behind you.  Tell us your own research origin story,   00:02:26.760 --> 00:02:33.000 as it were, and if there was one, what  big idea or animating “what if?” captured   00:02:33.000 --> 00:02:37.040 your imagination and inspired you  to do what you're doing today? 00:02:37.040 --> 00:02:41.080 SELLEN: OK, well, you're asking me to go  back in the mists of time a little bit,   00:02:41.080 --> 00:02:49.960 but let me try. [LAUGHTER] So I would say, going  … this goes back to my time when I started doing   00:02:49.960 --> 00:02:56.960 my PhD at UC San Diego. So I had just graduated as  a psychologist from the University of Toronto, and   00:02:56.960 --> 00:03:03.520 I was going to go off and do a PhD in psychology  with a guy called Don Norman. So back then,   00:03:03.520 --> 00:03:09.440 I really had very little interest in computers.  And in fact, computers weren't really a thing that   00:03:09.440 --> 00:03:15.440 normal people used. [LAUGHTER] They were things  that you might, like, put punch cards into. Or, in   00:03:15.440 --> 00:03:21.800 fact, in my undergrad days, I actually programmed  in hexadecimal, and it was horrible. But at UCSD,   00:03:21.800 --> 00:03:27.480 they were using computers everywhere, and it was,  kind of, central to how everyone worked. And we   00:03:27.480 --> 00:03:33.360 even had email back then. So computers weren't  really for personal use, and it was clear that   00:03:33.360 --> 00:03:40.000 they were designed for engineers by engineers.  And so they were horrible to use, people grappling   00:03:40.000 --> 00:03:44.400 with them, people were making mistakes. You could  easily remove all your files just by doing rm*.   00:03:45.480 --> 00:03:50.840 So the big idea that was going around the lab at  the time—and this was by a bunch of psychologists,   00:03:50.840 --> 00:03:57.760 not just Don, but other ones—was that we could  design computers for people, for people to use,   00:03:57.760 --> 00:04:02.600 and take into account, you know, how people  act and interact with things and what they   00:04:02.600 --> 00:04:07.440 want. And that was a radical idea at the  time. And that was the start of this field   00:04:07.440 --> 00:04:12.600 called human-computer interaction, which is …  you know, now we talk about designing computers   00:04:12.600 --> 00:04:17.286 for people and “user-friendly” and that's a,  kind of, like, normal thing, but back then … 00:04:17.286 --> 00:04:19.960 HUIZINGA: Yeah … SELLEN: … it was a radical idea. And so, to me, 00:04:19.960 --> 00:04:25.920 that changed everything for me to think about how  we could design technology for people. And then,   00:04:25.920 --> 00:04:27.779 if I can, I'll talk about one  other thing that happened … 00:04:27.779 --> 00:04:28.736 HUIZINGA: Yeah, please. 00:04:28.736 --> 00:04:34.240 SELLEN: … during that time. So at that time,  there was another gang of psychologists,   00:04:34.240 --> 00:04:37.840 people like Dave Rumelhart,  Geoff Hinton, Jay McClelland,   00:04:37.840 --> 00:04:43.000 people like that, who were thinking about,  how do we model human intelligence—learning,   00:04:43.000 --> 00:04:49.320 memory, cognition—using computers? And so  these were psychologists thinking about,   00:04:49.320 --> 00:04:53.320 how do people represent ideas and knowledge,  and how can we do that with computers? 00:04:53.320 --> 00:04:54.068 HUIZINGA: Yeah … 00:04:54.068 --> 00:04:57.440 SELLEN: And this was radical at the time  because cognitive psychologists back then   00:04:57.440 --> 00:05:03.600 were thinking about … they did lots of, kind  of, flow chart models of human cognition. And   00:05:03.600 --> 00:05:07.377 people like Dave Rumelhart did  networks, neural networks, … 00:05:07.377 --> 00:05:08.228 HUIZINGA: Ooh … 00:05:08.228 --> 00:05:13.960 SELLEN: … and they were using what were then  called spreading activation models of memory   00:05:13.960 --> 00:05:19.760 and things, which came from psychology. And  that's interesting because not only were they   00:05:19.760 --> 00:05:24.920 modeling human cognition in this, kind of, what  they called parallel distributed processing,   00:05:24.920 --> 00:05:29.280 but they operationalized it. And that's  where Hinton and others came up with the   00:05:29.280 --> 00:05:37.280 back-propagation algorithm, and that was a huge  leap forward in AI. So psychologists were actually   00:05:37.280 --> 00:05:42.240 directly responsible for the wave of AI we see  today. A lot of computer scientists don't know   00:05:42.240 --> 00:05:48.640 that. A lot of machine learning people don't know  that. But so, for me, long story short, that time   00:05:48.640 --> 00:05:56.920 in my life and doing my PhD at UC San Diego led to  me understanding that social science, psychology   00:05:56.920 --> 00:06:03.160 in particular, and computing should be seen as  things which mutually support one another and that   00:06:03.160 --> 00:06:09.440 can lead to huge breakthroughs in how we design  computers and computer algorithms and how we do   00:06:09.440 --> 00:06:15.420 computing. So that, kind of, set the path for the  rest of my career. And that was 40 years ago! 00:06:15.420 --> 00:06:22.120 HUIZINGA: Did you have what we'll call  metacognition of that being an aha moment for you,   00:06:22.120 --> 00:06:26.000 and like, I'm going to embrace this, and  this is my path forward? Or was it just,   00:06:26.000 --> 00:06:30.400 sort of, more iterative: these things  interest you, you take the next step,   00:06:30.400 --> 00:06:33.600 these things interest you  more, you take that step? 00:06:33.600 --> 00:06:37.720 SELLEN: I think it was an aha moment  at certain points. Like, for example,   00:06:37.720 --> 00:06:43.000 the day that Francis Crick walked into our seminar  and started talking about biologically inspired   00:06:43.000 --> 00:06:46.020 models of computing, I thought, “Ooh,  there's something big going on here!” 00:06:46.020 --> 00:06:46.920 HUIZINGA: Wow, yeah. 00:06:46.920 --> 00:06:51.840 SELLEN: Because even then I knew that he was a big  deal. So I knew there was something happening that   00:06:51.840 --> 00:06:57.000 was really, really interesting. I didn't think  so much about it from the point of view of,   00:06:57.000 --> 00:07:02.240 you know, I would have a career of helping  to design human-centric computing, but more,   00:07:02.240 --> 00:07:07.160 wow, there's a breakthrough in psychology and  how we understand the human mind. And I didn't   00:07:07.160 --> 00:07:12.180 realize at that time that that was going  to lead to what's happening in AI today. 00:07:12.180 --> 00:07:15.200 HUIZINGA: Well, let's talk about  some of these people that were   00:07:15.200 --> 00:07:22.440 influential for you as a follow-up to the  animating “big idea.” If I'm honest, Abi,   00:07:22.440 --> 00:07:27.640 my jaw dropped a little when I read your bio  because it's like a who's who of human-centered   00:07:27.640 --> 00:07:33.960 computing and UX design. And now these people  are famous. Maybe they weren't so much at the   00:07:33.960 --> 00:07:40.100 time. But tell us about the influential people in  your life, and how their ideas inspired you? 00:07:40.100 --> 00:07:45.880 SELLEN: Yeah, sure, happy to. In fact, I'll start  with one person who is not a, sort of, HCI person,   00:07:45.880 --> 00:07:52.080 but my stepfather, John Senders, was this  remarkable human being. He died three years ago at   00:07:52.080 --> 00:07:59.560 the age of 98. He worked almost to his dying day.  Just an amazing man. He entered my life when I   00:07:59.560 --> 00:08:05.840 was about 13. He joined the family. And he went to  Harvard. He trained with people like Skinner. He   00:08:05.840 --> 00:08:10.760 was taught by these, kind of, famous psychologists  of the 20th century, and they were his friends and   00:08:10.760 --> 00:08:16.240 his colleagues, and he introduced me to a lot of  them. You know, people like Danny Kahneman and,   00:08:16.240 --> 00:08:21.840 you know, Amos Tversky and Alan Baddeley, and all  these people that, you know, I had learned about   00:08:21.840 --> 00:08:28.440 as an undergrad. But the main thing that John did  for me was to open my eyes to how you could think   00:08:28.440 --> 00:08:34.560 about modeling humans as machines. And he really  believed that. He was not only a psychologist,   00:08:34.560 --> 00:08:39.680 but he was an engineer. And he, sort of, kicked  off or he was one of the founders of the field   00:08:39.680 --> 00:08:45.560 of human factors engineering. And that's what  human factors engineers do. They look at people,   00:08:45.560 --> 00:08:50.320 and they think, how can we mathematically model  them? So, you know, we'd be sitting by a pool,   00:08:50.320 --> 00:08:56.480 and he'd say, “You can use information sampling  to model the frequency with which somebody has   00:08:56.480 --> 00:09:01.400 to watch a baby as they go towards the  pool. And it depends on their velocity   00:09:01.400 --> 00:09:05.320 and then their trajectory … !” [LAUGHTER]  Or we go into a bank, and he'd say, “Abi,   00:09:05.320 --> 00:09:10.480 how would you use queuing theory to, you know,  estimate the mean wait time?” Like, you know,   00:09:10.480 --> 00:09:14.880 so he got me thinking like that, and he recognized  in me that I had this curiosity about the   00:09:14.880 --> 00:09:22.080 world and about people, but also, that I loved  mathematics. So he was the first guy. Don Norman,   00:09:22.080 --> 00:09:27.840 I've already mentioned as my PhD supervisor,  and I've said something about already how he,   00:09:27.840 --> 00:09:33.240 sort of, had this radical idea about designing  computers for people. And I was fortunate   00:09:33.240 --> 00:09:38.520 to be there when the field of human-computer  interaction was being born, and that was mainly   00:09:38.520 --> 00:09:43.080 down to him. And he's just [an] incredible  guy. He's still going. He's still working,   00:09:43.080 --> 00:09:49.240 consulting, and he wrote this famous book called  The Psychology of Everyday Things, which now is,   00:09:49.240 --> 00:09:53.760 I think it's been renamed The Design of Everyday  Things, and he was really influential and been   00:09:53.760 --> 00:09:58.049 a huge supporter of mine. And then the third  person I'll mention is Bill Buxton. And … 00:09:58.049 --> 00:09:59.699 HUIZINGA: Yeah … SELLEN: Bill, Bill … 00:09:59.699 --> 00:10:01.672 HUIZINGA: Bill, Bill, Bill! [LAUGHTER] 00:10:01.672 --> 00:10:05.760 SELLEN: Yeah. I met Bill at, first, well,  actually first at University of Toronto;   00:10:05.760 --> 00:10:09.800 when I was a grad student, I went up and told  him his … the experiment he was describing was   00:10:09.800 --> 00:10:14.160 badly designed. And instead of, you know,  brushing me off, he said, “Oh really, OK,   00:10:14.160 --> 00:10:18.280 I want to talk to you about that.” And then I  met him at Apple later when I was an intern,   00:10:18.280 --> 00:10:23.520 and we just started working together. And he  is, he's just … amazing designer. Everything   00:10:23.520 --> 00:10:28.880 he does is based on, kind of, theory and  deep thought. And he's just so much fun.   00:10:28.880 --> 00:10:33.320 So I would say those three people  have been big influences on me. 00:10:33.320 --> 00:10:38.987 HUIZINGA: Yeah. What about Marilyn Tremaine?  Was she a factor in what you did? 00:10:38.987 --> 00:10:41.140 SELLEN: Yes, yeah, she was  great. And Ron Baecker. So… 00:10:41.140 --> 00:10:41.788 HUIZINGA: Yeah … 00:10:41.788 --> 00:10:48.440 SELLEN: … after I did my PhD, I did a postdoc at  Toronto in the Dynamic Graphics Project Lab. And   00:10:48.440 --> 00:10:54.240 they were building a media space, and they  asked me to join them. And Marilyn and Ron   00:10:54.240 --> 00:10:59.240 and Bill were building this video telepresence  media space, which was way ahead of its time. 00:10:59.240 --> 00:11:00.040 HUIZINGA: Yeah. 00:11:00.040 --> 00:11:03.420 SELLEN: So I worked with all three  of them, and they were great fun. 00:11:03.420 --> 00:11:07.280 HUIZINGA: Well, let’s talk about the  research initiative AI, Cognition,   00:11:07.280 --> 00:11:13.480 and the Economy. For context, this is a global,  interdisciplinary effort to explore the impact   00:11:13.480 --> 00:11:19.280 of generative AI on human cognition and  thinking, work dynamics and practices,   00:11:19.280 --> 00:11:24.320 and labor markets and the economy. Now, we've  already lined up some AICE researchers to come   00:11:24.320 --> 00:11:29.480 on the podcast and talk about specific projects,  including pilot studies, workshops, and extended   00:11:29.480 --> 00:11:34.680 collaborations, but I'd like you to act as a,  sort of, docent or tour guide for the initiative,   00:11:34.680 --> 00:11:39.480 writ large, and tell us why, particularly  now, you think it's important to bring   00:11:39.480 --> 00:11:43.060 this group of scientists together  and what you hope to accomplish. 00:11:43.060 --> 00:11:47.800 SELLEN: I think it's important now because  I think there are so many extreme views   00:11:47.800 --> 00:11:51.960 out there about how AI is going to  impact people. A lot of hyperbole,   00:11:51.960 --> 00:11:57.480 right. So there's a lot of fear about, you  know, jobs going away, people being replaced,   00:11:57.480 --> 00:12:01.120 robots taking over the world. And there's  a lot of enthusiasm about how, you know,   00:12:01.120 --> 00:12:05.680 we're all going to be more productive, have more  free time, how it's going to be the answer to all   00:12:05.680 --> 00:12:12.880 our problems. And so I think there are people at  either end of that conversation. And I always … I   00:12:12.880 --> 00:12:16.704 love the Helen Fielding quote … I don't know  if you know Helen Fielding. She wrote… 00:12:16.704 --> 00:12:17.330 HUIZINGA: Yeah, Bridget Jones’s Diary … 00:12:17.330 --> 00:12:21.120 SELLEN: … Bridget Jones’s Diary. Yeah. [LAUGHTER]  She says, “Nothing is either as bad or as good   00:12:21.120 --> 00:12:27.240 as it seems,” right. And I live by that because I  think things are usually somewhere in the middle.   00:12:27.240 --> 00:12:34.160 So I'm not saying that we shouldn't take concerns  seriously about AI or be hugely optimistic about   00:12:34.160 --> 00:12:39.200 the opportunities, but rather, my view on this  is that we can do research to get, kind of,   00:12:39.200 --> 00:12:45.000 line of sight into the future and what is going to  happen with AI. And more than this, we should be   00:12:45.000 --> 00:12:49.880 using research to not just get line of sight but  to steer the future, right. We can actually help   00:12:49.880 --> 00:12:55.280 to shape it. And especially being at Microsoft,  we have a chance to do that. So what I mean here   00:12:55.280 --> 00:13:01.160 is that let's begin by understanding first the  capabilities of AI and get a good understanding of   00:13:01.160 --> 00:13:05.756 where it's heading and the pace that it's heading  at because it's changing so fast, right. 00:13:05.756 --> 00:13:06.390 HUIZINGA: Mm-hmm … 00:13:06.390 --> 00:13:09.760 SELLEN: And then let's do some research  looking at the impact, both in the short term   00:13:09.760 --> 00:13:15.320 and the long term, about its impact on tasks, on  interaction, and, most importantly for me anyway,   00:13:15.320 --> 00:13:22.000 on people. Yeah, and then we can extrapolate  out how this is going to impact jobs, skills,   00:13:22.000 --> 00:13:27.360 organizations, society at large, you know. So  we get this, kind of, arc that we can trace,   00:13:27.360 --> 00:13:32.800 but we do it because we do research. We don't  just rely on the hyperbole and speculation,   00:13:32.800 --> 00:13:39.040 but we actually try and do it more systematically.  And then I think the last piece here is that if   00:13:39.040 --> 00:13:44.400 we're going to do this well and if we think  about what AI's impact can be, which we think   00:13:44.400 --> 00:13:48.520 is going to impact on a global scale, we  need many different skills and disciplines.   00:13:48.520 --> 00:13:54.440 We need not just machine learning people and  engineering and computer scientists at large,   00:13:54.440 --> 00:13:59.280 but we need designers, we need social scientists,  we need even philosophers, and we need domain   00:13:59.280 --> 00:14:04.640 experts, right. So we need to bring all of  these people together to do this properly. 00:14:04.640 --> 00:14:09.560 HUIZINGA: Interesting. Well, let's do break it  down a little bit then. And I want to ask you a   00:14:09.560 --> 00:14:16.040 couple questions about each of the disciplines  within the acronym A-I-C-E, or AICE. And I'll   00:14:16.040 --> 00:14:21.960 start with AI and another author that we can  refer to. Sci-fi author and futurist Arthur   00:14:21.960 --> 00:14:28.120 C. Clarke famously said that “any sufficiently  advanced technology is indistinguishable from   00:14:28.120 --> 00:14:34.080 magic,” and for many people, AI systems  seem to be magic. So in response to that,   00:14:34.080 --> 00:14:39.560 many in the industry have emphatically stated that  AI is just a tool. But you've said things like   00:14:39.560 --> 00:14:45.760 AI is more a “collaborative copilot than a mere  tool,” and recently, you said we might even think   00:14:45.760 --> 00:14:51.920 of it as a “very smart and intuitive butler.”  So how do those ideas from the airline industry   00:14:51.920 --> 00:14:58.920 and Downton Abbey help us better understand  and position AI and its role in our world? 00:14:58.920 --> 00:15:04.960 SELLEN: Well, I'm going to use Wodehouse here  in a minute as well, but um … so I think AI is   00:15:04.960 --> 00:15:10.760 different from many other tech developments in a  number of important ways. One is, it has agency,   00:15:10.760 --> 00:15:17.080 right. So it can take initiative and do things on  your behalf. It's highly complex, and, you know,   00:15:17.080 --> 00:15:22.440 it's getting more complex by the day. It  changes. It's dynamic. It's probabilistic   00:15:22.440 --> 00:15:25.560 rather than deterministic, so it will give  you different answers depending on when,   00:15:25.560 --> 00:15:29.480 you know, when you ask it and what you  ask it. And it's based on human-generated   00:15:29.480 --> 00:15:34.200 data. So it's a vastly different  kind of tool than HCI, as a field,   00:15:34.200 --> 00:15:38.840 has studied in the past. There are lots  of downsides to that, right. One is it   00:15:38.840 --> 00:15:42.320 means it's very hard to understand how  it works under the hood, right … 00:15:42.320 --> 00:15:42.814 HUIZINGA: Yeah … 00:15:42.814 --> 00:15:47.240 SELLEN: … and understanding the output. It's  fraught with uncertainty because the output   00:15:47.240 --> 00:15:52.240 changes every time you use it. But then let's  think about the upsides, especially, large   00:15:52.240 --> 00:15:59.080 language models give us a way of conversationally  interacting with AI like never before,   00:15:59.080 --> 00:16:04.880 right. So it really is a new interaction paradigm  which has finally come of age. So I do think it's   00:16:04.880 --> 00:16:10.000 going to get more personal over time and more  anticipatory of our needs. And if we design   00:16:10.000 --> 00:16:16.720 it right, it can be like the perfect butler. So  if you know P.G. Wodehouse, Jeeves and Wooster,   00:16:16.720 --> 00:16:22.000 you know, Jeeves knows that Bertie has had a rough  night and has a hangover, so he's there at the   00:16:22.000 --> 00:16:28.480 bedside with a tonic and a warm bath already ready  for him. But he also knows what Wooster enjoys and   00:16:28.480 --> 00:16:32.960 what decisions should be left to him, and he knows  when to get out of the way. He also knows when to   00:16:32.960 --> 00:16:38.920 be very discreet, right. So when I use that  butler metaphor, I think about how it's going   00:16:38.920 --> 00:16:43.880 to take time to get this right, but eventually,  we may live in a world where AI helps us with   00:16:43.880 --> 00:16:49.440 good attention to privacy of getting that kind of  partnership right between Jeeves and Wooster. 00:16:49.440 --> 00:16:52.180 HUIZINGA: Right. Do you think that's possible? 00:16:52.180 --> 00:16:55.600 SELLEN: I don't think we'll  ever get it exactly right,   00:16:55.600 --> 00:17:01.320 but if we have a conversational system  where we can mutually shape the interaction,   00:17:01.320 --> 00:17:06.680 then even if Jeeves doesn't get things right,  Wooster can train him to do a better job. 00:17:06.680 --> 00:17:11.960 HUIZINGA: Go back to the copilot analogy,  which is a huge thing at Microsoft—in fact,   00:17:11.960 --> 00:17:19.960 they've got products named Copilot—and  the idea of a copilot, which is, sort of,   00:17:19.960 --> 00:17:24.272 assuaging our fears that  it would be the pilot … 00:17:24.272 --> 00:17:24.960 SELLEN: Yeah … HUIZINGA: … AI. 00:17:24.960 --> 00:17:25.920 SELLEN: Yeah, yeah. 00:17:25.920 --> 00:17:29.000 HUIZINGA: So how do we envision that in a way   00:17:29.000 --> 00:17:34.020 that … you say it's more than a mere  tool, but it's more like a copilot? 00:17:34.020 --> 00:17:38.840 SELLEN: Yeah, I actually like the copilot  metaphor for what you're alluding to,   00:17:38.840 --> 00:17:44.320 which is that the pilot is the one who  has the final say, who has the, kind of,   00:17:44.320 --> 00:17:49.400 oversight of everything that's happening  and can step in. And also that the copilot   00:17:49.400 --> 00:17:54.720 is there in a supportive role, who kind of  trains by dint of the fact that they work   00:17:54.720 --> 00:17:59.520 next to the pilot, and that they have, you  know, specialist skills that can help. 00:17:59.520 --> 00:18:00.072 HUIZINGA: Right … 00:18:00.072 --> 00:18:03.680 SELLEN: So I really like that metaphor.  I think there are other metaphors that we   00:18:03.680 --> 00:18:08.400 will explore in future and which will make  sense for different contexts, but I think,   00:18:08.400 --> 00:18:12.520 as a metaphor for a lot of the things we're  developing today, it makes a lot of sense. 00:18:12.520 --> 00:18:17.200 HUIZINGA: You know, it also  feels like, in the conversation,   00:18:17.200 --> 00:18:24.520 words really matter in how people perceive  what the tool is. So having these other   00:18:24.520 --> 00:18:30.320 frameworks to describe it and to implement  it, I think, could be really helpful. 00:18:30.320 --> 00:18:32.015 SELLEN: Yes, I agree. 00:18:32.015 --> 00:18:41.150 [MUSIC BREAK] 00:18:41.150 --> 00:18:45.320 HUIZINGA: Well, let's talk about intelligence  for a second. One of the most interesting things   00:18:45.320 --> 00:18:52.200 about AI is it's caused us to pay attention to  other kinds of intelligence. As author Meghan   00:18:52.200 --> 00:18:56.960 O'Gieblyn puts it, “God, human, animal,  machine … ” So why do you think, Abi,   00:18:56.960 --> 00:19:01.880 it's important to understand the characteristics  of each kind of intelligence, and how does that   00:19:01.880 --> 00:19:07.340 impact how we conceptualize, make, and use  what we're calling artificial intelligence? 00:19:07.340 --> 00:19:09.480 SELLEN: Yeah, well, I actually prefer the   00:19:09.480 --> 00:19:12.200 term machine intelligence to  artificial intelligence … 00:19:12.200 --> 00:19:13.832 HUIZINGA: Me too! Thank you! [LAUGHTER] 00:19:13.832 --> 00:19:17.800 SELLEN: Because the latter implies  that there's one kind of intelligence,   00:19:17.800 --> 00:19:23.560 and also, it does allude to the fact that that is  human-like. You know, we're trying to imitate the   00:19:23.560 --> 00:19:28.880 human. But if you think about animals, I think  that's really interesting. I mean, many of us   00:19:28.880 --> 00:19:32.960 have good relationships with our pets, right.  And we know that they have a different kind of   00:19:32.960 --> 00:19:37.080 intelligence. And it's different from ours,  but that doesn't mean we can't understand it   00:19:37.080 --> 00:19:42.680 to some extent, right. And if you think about …  animals are superhuman in many ways, right. They   00:19:42.680 --> 00:19:48.400 can do things we can't. So whether it's an ox  pulling a plow or a dog who can sniff out drugs   00:19:48.400 --> 00:19:53.200 or ferrets who can, you know, thread electrical  cables through pipes, they can do things. And   00:19:53.200 --> 00:19:59.240 bee colonies are fascinating to me, right. And  they work as a, kind of, a crowd intelligence,   00:19:59.240 --> 00:20:04.880 or hive mind, right. [LAUGHTER] That's where  that comes from. And so in so many ways,   00:20:04.880 --> 00:20:10.840 animals are smarter than humans. But it doesn't  matter—like this “smarter than” thing also bugs   00:20:10.840 --> 00:20:15.960 me. It's about being differently intelligent,  right. And the reason I think that's important   00:20:15.960 --> 00:20:20.680 when we think about machine intelligence is that  machine intelligence is differently intelligent,   00:20:20.680 --> 00:20:27.520 as well. So the conversational interface allows us  to explore the nature of that machine intelligence   00:20:27.520 --> 00:20:33.560 because we can speak to it in a kind of human-like  way, but that doesn't mean that it is intelligent   00:20:33.560 --> 00:20:37.848 in the same way a human is intelligent. And in  fact, we don't really want it to be, right. 00:20:37.848 --> 00:20:38.515 HUIZINGA: Right … 00:20:38.515 --> 00:20:42.240 SELLEN: Because we want it, we want it to be  a partner with us, to do things that we can't,   00:20:42.240 --> 00:20:46.720 you know, just like using the plow and  the ox. That partnership works because   00:20:46.720 --> 00:20:51.720 the ox is stronger than we are. So I think  machine intelligence is a much better word,   00:20:51.720 --> 00:20:56.440 and understanding it's not human  is a good thing. I do worry that,   00:20:56.440 --> 00:21:01.406 because it sounds like a human, it can  seduce us into thinking it's a human … 00:21:01.428 --> 00:21:05.880 SELLEN: … and that can be problematic. You know,  there are instances where people have been on,   00:21:05.880 --> 00:21:12.400 for example, dating sites and a bot is sounding  like a human and people get fooled. So I think we   00:21:12.400 --> 00:21:17.640 don't want to go down the path of fooling people.  We want to be really careful about that. 00:21:17.640 --> 00:21:22.960 HUIZINGA: Yeah, this idea of conflating  different kinds of intelligences to our   00:21:22.960 --> 00:21:27.720 own … I think we can have a separate  vision of animal intelligence,   00:21:27.720 --> 00:21:33.929 but machines are, like you say, kind  of seductively built to be like us. 00:21:33.950 --> 00:21:40.720 And so back to your comment about shaping how this technology moves forward and   00:21:40.720 --> 00:21:46.760 the psychology of it, how might we envision  how we could shape, either through language   00:21:46.760 --> 00:21:54.280 or the way these machines operate, that we build  in a “I'm not going to fool you” mechanism? 00:21:54.280 --> 00:21:58.800 SELLEN: Well, I mean, there are things that we  do at the, kind of, technical level in terms   00:21:58.800 --> 00:22:04.000 of guardrails and metaprompts, and we have  guidelines around that. But there's also the   00:22:04.000 --> 00:22:10.360 language that an AI character will use in terms  of, you know, expressing thoughts and feelings and   00:22:10.360 --> 00:22:15.600 some suggestion of an inner life, which … these  machines don't have an inner life, right. 00:22:15.600 --> 00:22:16.380 HUIZINGA: Right! 00:22:16.380 --> 00:22:18.720 SELLEN: So … and one of the  reasons we talk to people is   00:22:18.720 --> 00:22:21.760 we want to discover something  about their inner life. 00:22:21.760 --> 00:22:22.272 HUIZINGA: Yessss … 00:22:22.272 --> 00:22:26.200 SELLEN: And so why would I talk to a machine to  try and discover that? So I think there are things   00:22:26.200 --> 00:22:32.120 that we can do in terms of how we design these  systems so that they're not trying to deceive   00:22:32.120 --> 00:22:38.080 us. Unless we want them to deceive us. So  if we want to be entertained or immersed,   00:22:38.600 --> 00:22:41.920 maybe that's a good thing, right? That  they deceive us. But we enter into that   00:22:41.920 --> 00:22:44.920 knowing that that's what's happening,  and I think that's the difference. 00:22:44.920 --> 00:22:52.600 HUIZINGA: Well, let's talk about the C in A-I-C-E,  which is cognition. And we've just talked about   00:22:52.600 --> 00:22:57.320 other kinds of intelligence. Let's broaden the  conversation and talk about the impact of AI on   00:22:57.320 --> 00:23:03.320 humans themselves. Is there any evidence to  indicate that machine intelligence actually   00:23:03.320 --> 00:23:09.180 has an impact on human intelligence, and if  so, why is that an important data point? 00:23:09.180 --> 00:23:14.680 SELLEN: Yeah, OK, great topic. This is one of my  favorite topics. [LAUGHTER] So, well, let me just   00:23:14.680 --> 00:23:20.200 backtrack a little bit for a minute. A lot of the  work that's coming out today looking at the impact   00:23:20.200 --> 00:23:25.680 of AI on people is in terms of their productivity,  in terms of how fast they can do something,   00:23:25.680 --> 00:23:30.880 how efficiently they can do a job, or the quality  of the output of the tasks. And I do think that's   00:23:30.880 --> 00:23:36.880 important to understand because, you know, as we  deploy these new tools in peoples’ hands, we want   00:23:36.880 --> 00:23:42.760 to know what's happening in terms of, you know,  peoples’ productivity, workflow, and so on. But   00:23:42.760 --> 00:23:48.520 there's far less of it on looking at the impact  of using AI on people themselves and on how people   00:23:48.520 --> 00:23:55.760 think, on their cognitive processes, and how are  these changing over time? Are they growing? Are   00:23:55.760 --> 00:24:01.800 they atrophying as they use them? And, relatedly,  what's happening to our skills? You know, over   00:24:01.800 --> 00:24:06.000 time, what's going to be valued, and what's going  to drop away? And I think that's important for all   00:24:06.000 --> 00:24:11.640 kinds of reasons. So if you think about generative  AI, right, these are these AI systems that will   00:24:11.640 --> 00:24:16.600 write something for us or make a slide deck or  a picture or a video. What they're doing is they   00:24:16.600 --> 00:24:22.240 are taking the cognitive work of generation of  an artifact or the effort of self-expression   00:24:22.240 --> 00:24:25.880 that most of us, in the old-fashioned  world, will do, right—we write something,   00:24:25.880 --> 00:24:31.520 we make something—they're doing that for us on our  behalf. And so our job then is to think about how   00:24:31.520 --> 00:24:36.760 do we specify our intention to the machine, how do  we talk to it to get it to do the things we want,   00:24:36.760 --> 00:24:43.080 and then how do we evaluate the output at the  end? So it's really radically shifting what we do,   00:24:43.080 --> 00:24:46.920 the work that we do, the cognitive and mental  work that we do, when we engage with these   00:24:46.920 --> 00:24:55.400 tools. Now why is that a problem? Or should it be  a problem? One concern is that many of us think   00:24:55.400 --> 00:25:00.040 and structure our thoughts through the process  of making things, right. Through the process of   00:25:00.040 --> 00:25:05.960 writing or making something. So a big question  for me is, if we're removed from that process,   00:25:05.960 --> 00:25:12.760 how deeply will we learn or understand  what we're writing about? A second one is,   00:25:12.760 --> 00:25:17.840 you know, if we're not deeply engaged in the  process of generating these things, does that   00:25:17.840 --> 00:25:22.129 actually undermine our ability to evaluate the  output when we do get presented with it? 00:25:22.152 --> 00:25:27.320 Like, if it writes something for  us and it's full of problems and errors,   00:25:27.320 --> 00:25:31.360 if we stop writing for ourselves, are we  going to be worse at, kind of, judging the   00:25:31.360 --> 00:25:38.160 output? Another one is, as we hand things over  to more and more of these automated processes,   00:25:38.160 --> 00:25:43.720 will we start to blindly accept  or over-rely on our AI assistants,   00:25:43.720 --> 00:25:46.950 right. And the aviation industry  has known that for years … 00:25:46.950 --> 00:25:47.548 HUIZINGA: Yeah … 00:25:47.548 --> 00:25:52.600 SELLEN: … which is why they stick pilots in  simulators. Because they rely on autopilot so   00:25:52.600 --> 00:25:57.640 much that they forget those key skills. And  then another one is, kind of, longer term,   00:25:57.640 --> 00:26:02.960 which is like these new generations of people  who are going to grow up with this technology,   00:26:02.960 --> 00:26:09.320 what are the fundamental skills that they're going  to need to not just to use the AI but to be kind   00:26:09.320 --> 00:26:17.040 of citizens of the world and also be able to judge  the output of these AI systems? So the calculator,   00:26:17.040 --> 00:26:21.120 right, is a great example. When it was first  introduced, there was a huge outcry around,   00:26:21.120 --> 00:26:25.800 you know, kids won't be able to do math anymore!  Or we don't need to teach it anymore. Well,   00:26:25.800 --> 00:26:30.040 we do still teach it because when you  use a calculator, you need to be able   00:26:30.040 --> 00:26:34.531 to see whether or not the output the machine is  giving you is in the right ballpark, right. 00:26:34.531 --> 00:26:35.031 HUIZINGA: Right … 00:26:35.031 --> 00:26:39.040 SELLEN: You need to know the basics. And so what are the basics that kids are going to need to   00:26:39.040 --> 00:26:44.640 know? We just don't have the answer to those  questions. And then the last thing I'll say   00:26:44.640 --> 00:26:49.000 on this, because I could go on for a long time,  is we also know that there are changes in the   00:26:49.000 --> 00:26:55.560 brain when we use these new technologies.  There are shifts in our cognitive skills,   00:26:55.560 --> 00:27:01.960 you know, things get better and things do  deteriorate. So I think Susan Greenfield is   00:27:01.960 --> 00:27:08.400 famous for her work looking at what happens to  the neural pathways in the age of the internet,   00:27:08.400 --> 00:27:11.920 for example. So she found that all the  studies were pointing to the fact that   00:27:11.920 --> 00:27:18.800 reading online and on the internet meant that  our visual-spatial skills were being boosted,   00:27:18.800 --> 00:27:24.640 but our capacity to do deep processing, mindful  knowledge acquisition, critical thinking,   00:27:24.640 --> 00:27:31.560 reflection, were all decreasing over time. And I  think any parent who has a teenager will know that   00:27:31.560 --> 00:27:36.560 focus of attention, flitting from one thing  to another, multitasking, is, sort of, the   00:27:36.560 --> 00:27:41.320 order of the day. Well, not just for teenagers. I  think all of us are suffering from this now. It's   00:27:41.320 --> 00:27:46.021 much harder. I find it much harder to sit down  and read something in a long, focused way … 00:27:46.021 --> 00:27:46.548 HUIZINGA: Yeah … 00:27:46.548 --> 00:27:51.960 SELLEN: … than I used to. So all of this  long-winded answer is to say, we don't understand   00:27:51.960 --> 00:27:58.120 what the impact of these new AI systems is going  to be. We need to do research to understand it.   00:27:58.760 --> 00:28:02.880 And we need to do that research  both looking at short-term impacts   00:28:02.880 --> 00:28:06.520 and long-term impacts. Not to say  that this is all going to be bad,   00:28:06.520 --> 00:28:10.460 but we need to understand where it's  going so we can design around it. 00:28:10.460 --> 00:28:14.280 HUIZINGA: You know, even as you  asked each of those questions,   00:28:14.280 --> 00:28:19.760 Abi, I found myself answering it preemptively,  “Yes. That's going to happen. That's going to   00:28:19.760 --> 00:28:24.960 happen.” [LAUGHS] And so even as you say  all of this and you say we need research,   00:28:24.960 --> 00:28:33.160 do you already have some thinking about, you know,  if research tells us the answer that we thought   00:28:33.160 --> 00:28:39.280 might be true already, do we have a plan in place  or a thought process in place to address it? 00:28:39.280 --> 00:28:42.880 SELLEN: Well, yes, and I think we've got  some really exciting research going on in   00:28:42.880 --> 00:28:47.760 the company right now and in the AICE  program, and I'm hoping your future   00:28:47.760 --> 00:28:51.760 guests will be able to talk more in-depth  about these things. But we are looking at   00:28:51.760 --> 00:28:58.040 things like the impact of AI on writing, on  comprehension, on mathematical abilities.   00:28:59.160 --> 00:29:04.320 But more than that. Not just understanding  the impact on these skills and abilities,   00:29:04.320 --> 00:29:09.954 but how can we design systems better  to help people think better, right? 00:29:09.954 --> 00:29:10.521 HUIZINGA: Yeah … 00:29:10.521 --> 00:29:15.200 SELLEN: To help them think more deeply,  more creatively. I don't think AI needs to   00:29:15.200 --> 00:29:21.640 necessarily de-skill us in the critical skills  that we want and need. It can actually help us   00:29:21.640 --> 00:29:27.400 if we design them properly. And so that's the  other part of what we're doing. It's not just   00:29:27.400 --> 00:29:31.680 understanding the impact, but now saying,  OK, now that we understand what's going on,   00:29:31.680 --> 00:29:37.120 how do we design these systems better  to help people deepen their skills,   00:29:37.120 --> 00:29:42.640 change the way that they think in ways that they  want to change—in being more creative, thinking   00:29:42.640 --> 00:29:47.000 more deeply, you know, reading in different ways,  understanding the world in different ways. 00:29:47.000 --> 00:29:53.320 HUIZINGA: Right. Well, that is a brilliant segue  into my next question. Because we're on the last   00:29:53.320 --> 00:30:02.120 letter, E, in AICE: the economy. And that I think  instills a lot of fear in people. To cite another   00:30:02.120 --> 00:30:07.880 author, since we're on a citing authors roll,  Clay Shirky, in his book Here Comes Everybody,   00:30:07.880 --> 00:30:13.040 writes about technical revolutions in general  and the impact they have on existing economic   00:30:13.040 --> 00:30:18.880 paradigms. And he says, "Real revolutions don't  involve an orderly transition from point A to   00:30:18.880 --> 00:30:25.720 point B. Rather, they go from A, through a long  period of chaos, and only then reach B. And in   00:30:25.720 --> 00:30:31.560 that chaotic period the old systems get broken  long before the new ones become stable.” Let's   00:30:31.560 --> 00:30:39.720 take Shirky’s idea and apply it to generative AI.  If B equals the future of work, what's getting   00:30:39.720 --> 00:30:45.600 broken in the period of transition from how things  were to how things are going to be, what do we   00:30:45.600 --> 00:30:52.384 have to look forward to, and how do we progress  toward B in a way that minimizes chaos? 00:30:52.384 --> 00:30:53.630 SELLEN: Hmm … oh, those are  big questions! [LAUGHS] 00:30:53.630 --> 00:30:54.952 HUIZINGA: Too many questions! [LAUGHS] 00:30:54.952 --> 00:31:00.400 SELLEN: Yeah, well, I mean, Shirky was  right. Things take a long time to bed in,   00:31:00.400 --> 00:31:06.080 right. And much of what happens over time, I don't  think we can actually predict. You know, so who   00:31:06.080 --> 00:31:11.000 would have predicted echo chambers or the rise of  deepfakes or, you know, the way social media could   00:31:11.000 --> 00:31:16.160 start revolutions in those early days of social  media, right. So good and bad things happen,   00:31:16.160 --> 00:31:21.840 and a lot of it's because it rolls out over time,  it scales up, and then people get involved. And   00:31:21.840 --> 00:31:26.800 that's the really unpredictable bit, is when  people get involved en masse. I think we're   00:31:26.800 --> 00:31:31.760 going to see the same thing with AI systems.  They are going to take a long time to bed in,   00:31:31.760 --> 00:31:38.440 and its impact is going to be global, and it's  going to take a long time to unfold. So I think   00:31:38.440 --> 00:31:43.560 what we can do is, to some extent, we can see the  glimmerings of what's going to happen, right. So I   00:31:43.560 --> 00:31:48.720 think it's the William Gibson quote is, you know,  “The future's already here; it's just not evenly   00:31:48.720 --> 00:31:54.440 distributed,” or something like that, right. We  can see some of the problems that are playing out,   00:31:54.440 --> 00:31:59.360 both in the hands of bad actors and things that  will happen unintentionally. We can see those, and   00:31:59.360 --> 00:32:05.400 we can design for them, and we can do things about  it because we are alert and we are looking to see   00:32:05.400 --> 00:32:09.165 what happens. And also, the good things, right.  And all the good things that are playing out, 00:32:09.188 --> 00:32:14.440 we can make the most of those. Other  things we can do is, you know, at Microsoft,   00:32:14.440 --> 00:32:22.040 we have a set of responsible AI principles that  we make sure all our products go through to make   00:32:22.040 --> 00:32:27.680 sure that we look into the future as much as we  can, consider what the consequences might be,   00:32:27.680 --> 00:32:33.040 and then deploy things in very careful  steps, evaluating as we go. And then,   00:32:33.040 --> 00:32:36.320 coming back to what I said earlier, doing  deep research to try and get a better line   00:32:36.320 --> 00:32:41.640 of sight. So in terms of what's going to  happen with the future of work, I think,   00:32:41.640 --> 00:32:46.320 again, we need to steer it. Some of the things I  talked about earlier in terms of making sure we   00:32:46.320 --> 00:32:52.360 build skills rather than undermine them, making  sure we don't over automate, making sure that we   00:32:52.360 --> 00:32:58.440 put agency in the hands of people. And always  making sure that we design our AI experiences   00:32:58.440 --> 00:33:05.000 with human hope, aspirations, and needs in mind.  If we do that, I think we're on a good track,   00:33:05.000 --> 00:33:09.400 but we should always be vigilant, you know,  to what’s evolving, what's happening here. 00:33:09.400 --> 00:33:10.032 HUIZINGA: Yeah … 00:33:10.032 --> 00:33:12.520 SELLEN: I can't really predict  whether we're headed for chaos   00:33:12.520 --> 00:33:15.760 or not. I don't think we are,  as long as we're mindful. 00:33:15.760 --> 00:33:25.200 HUIZINGA: Yeah. And it sounds like there's a  lot more involved outside of computer science,   00:33:25.200 --> 00:33:34.840 in terms of support systems and education  and communication, to acclimatize people to   00:33:34.840 --> 00:33:40.000 a new kind of economy, which like you say, you  can't … I'm shocked that you can't predict it,   00:33:40.000 --> 00:33:42.760 Abi. I was expecting that you  could, but … [LAUGHTER] 00:33:42.760 --> 00:33:44.080 SELLEN: Sorry. 00:33:44.080 --> 00:33:51.400 HUIZINGA: Sorry! But yeah, I mean,  do you see the ancillary industries,   00:33:51.400 --> 00:33:57.800 we'll call them, in on this? And how can,  you know, sort of, a lab in Cambridge,   00:33:57.800 --> 00:34:05.400 and labs around the world that are doing AI, how  can they spread out to incorporate these other   00:34:05.400 --> 00:34:12.880 things to help the people who know nothing about  what's going on in your lab move forward here? 00:34:12.880 --> 00:34:17.720 SELLEN: Well, I think, you know, there are  lots of people that we need to talk to and   00:34:17.720 --> 00:34:21.480 to take account of. The word stakeholder  … I hate that word stakeholder! I'm not   00:34:21.480 --> 00:34:27.240 sure why. [LAUGHTER] But anyway, stakeholders in  this whole AI odyssey that we're on … you know,   00:34:27.240 --> 00:34:31.280 public perceptions are one thing. I'm  a member of a lot of societies where we   00:34:31.280 --> 00:34:36.160 do a lot of outreach and talks about AI and  what's going on, and I think that's really,   00:34:36.160 --> 00:34:41.769 really important. And get people excited also  about the possibilities of what could happen. 00:34:41.792 --> 00:34:44.600 Because I think a lot of the media,  a lot of the stories that get out there are   00:34:44.600 --> 00:34:52.120 very dystopian and scary, and it's right that we  are concerned and we are alert to possibilities,   00:34:52.120 --> 00:34:57.800 but I don't think it does anybody any good  to make people scared or anxious. And so   00:34:57.800 --> 00:35:02.600 I think there's a lot we can do with the  public. And there's a lot we can do with,   00:35:02.600 --> 00:35:08.360 when I think about the future of work, different  domains, you know, and talking to them about   00:35:08.360 --> 00:35:13.700 their needs and how they see AI fitting  into their particular work processes. 00:35:13.700 --> 00:35:19.960 HUIZINGA: So, Abi, we're kind of [LAUGHS]  dancing around these dystopian narratives,   00:35:19.960 --> 00:35:25.920 and whether they're right or wrong, they have  gained traction. So it's about now that I ask   00:35:25.920 --> 00:35:32.080 all of my guests what could go wrong if you got  everything right? So maybe you could present,   00:35:32.080 --> 00:35:38.440 in this area, some more hopeful, we'll call them  “-topias,” or preferred futures, if you will,   00:35:38.440 --> 00:35:45.220 around AI and how you and/or your lab and other  people in the industry are preparing for them. 00:35:45.220 --> 00:35:51.920 SELLEN: Well, again, I come back to the idea that  the future is all around us to some extent, and   00:35:51.920 --> 00:35:58.640 we're seeing really amazing breakthroughs, right,  with AI. For example, scientific breakthroughs   00:35:58.640 --> 00:36:04.280 in terms of, you know, drug discovery, new  materials to help tackle climate change, all   00:36:04.280 --> 00:36:09.640 kinds of things that are going to help us tackle  some of the world's biggest problems. Better   00:36:09.640 --> 00:36:15.680 understandings of the natural world, right, and  how interventions can help us. New tools in the   00:36:15.680 --> 00:36:20.840 hands of low-literacy populations and support for,  you know, different ways of working in different   00:36:20.840 --> 00:36:26.400 cultures. I think that's another big area in  which AI can help us. Personalization—personalized   00:36:26.400 --> 00:36:31.720 medicine, personalized tutoring systems, right.  So we talked about education earlier. I think   00:36:31.720 --> 00:36:38.240 that AI could do a lot if we design it right  to really help in education and help support   00:36:38.240 --> 00:36:43.840 people's learning processes. So I think there's  a lot here, and there's a lot of excitement—with   00:36:43.840 --> 00:36:48.720 good reason. Because we're already seeing these  things happening. And we should bear those things   00:36:48.720 --> 00:36:53.320 in mind when we start to get anxious about AI.  And I personally am really, really excited about   00:36:53.320 --> 00:36:57.680 it. I'm excited about, you know, what the  company I work for is doing in this area and   00:36:57.680 --> 00:37:03.360 other companies around the world. I think that  it's really going to help us in the long term,   00:37:03.360 --> 00:37:08.260 build new skills, see the world in new ways,  you know, tackle some of these big problems. 00:37:08.260 --> 00:37:14.000 HUIZINGA: I recently saw an ad—I'm not making  this up—it was the quote-unquote “productivity   00:37:14.000 --> 00:37:19.120 app,” and it was simply a small wooden box  filled with pieces of paper. And there was a   00:37:19.120 --> 00:37:25.520 young man who had a how-to video on how to use it  on YouTube. [LAUGHS] He was clearly born into the   00:37:25.520 --> 00:37:32.880 digital age and found writing lists on paper  to be a revolutionary idea. But I myself have   00:37:32.880 --> 00:37:38.560 toggled back and forth between what we'll call  the affordances of the digital world and the   00:37:38.560 --> 00:37:44.960 familiarity and comfort of the physical world. And  you actually studied this and wrote about it in a   00:37:44.960 --> 00:37:51.280 book called The Myth of the Paperless Office. That  was 20 years ago. Why did you do the work then,   00:37:51.280 --> 00:37:57.080 what's changed in the ensuing years, and why  in the age of AI do I love paper so much? 00:37:57.080 --> 00:38:02.640 SELLEN: Yeah, so, that was quite a while ago now.  It was a book that I cowrote with my husband. He's   00:38:02.640 --> 00:38:09.040 a sociologist, so we, sort of, came together  on that book, me as a psychologist and he as   00:38:09.040 --> 00:38:15.920 a sociologist. What we were responding to at  the time was a lot of hype about the paperless   00:38:15.920 --> 00:38:22.200 office and the paperless future. At the time, I  was working at EuroPARC, you know, which is the   00:38:22.200 --> 00:38:27.440 European sister lab of Xerox PARC. And so,  obviously, they had big investment in this.   00:38:28.240 --> 00:38:32.600 And there were many people in that lab who  really believed in the paperless office,   00:38:32.600 --> 00:38:37.160 and lots of great inventions came out of the  fact that people were pursuing that vision. So   00:38:37.160 --> 00:38:41.760 that was a good side of that, but we also saw  where things could go horribly wrong when you   00:38:41.760 --> 00:38:45.280 just took a paper-based system away and you  just replaced it with a digital system. 00:38:45.280 --> 00:38:45.992 HUIZINGA: Yeah … 00:38:45.992 --> 00:38:49.760 SELLEN: I remember some of the disasters  in air traffic control, for example,   00:38:49.760 --> 00:38:53.320 when they took the paper flight strips  away and just made them all digital. And   00:38:53.320 --> 00:38:56.285 those are places where you don't want to  mess around with something that works. 00:38:56.285 --> 00:38:56.985 HUIZINGA: Right. 00:38:56.985 --> 00:39:01.360 SELLEN: You have to be really careful about  how you introduce digital systems. Likewise,   00:39:01.360 --> 00:39:06.800 many people remember things that went wrong  when hospitals tried to go paperless with   00:39:06.800 --> 00:39:14.240 health records being paperless. Now, those things  are digital now, but we were talking about chaos   00:39:14.240 --> 00:39:20.280 earlier. There was a lot of chaos on the path.  So what we've tried to say in that book to some   00:39:20.280 --> 00:39:26.880 extent is, let's understand the work that paper  is doing in these different work contexts and the   00:39:26.880 --> 00:39:32.720 affordances of paper. You know, what is it doing  for people? Anything from, you know, I hand a    00:39:32.720 --> 00:39:37.017 document over to someone else; a physical document gives me the excuse to talk to that person … 00:39:37.017 --> 00:39:37.708 HUIZINGA: Right… 00:39:37.708 --> 00:39:42.040 SELLEN: … through to, you know, when I place  a document on somebody's desk, other people in   00:39:42.040 --> 00:39:48.640 the workplace can see that I've passed it on to  someone else. Those kind of nuanced observations   00:39:48.640 --> 00:39:53.400 are useful because you then need to think, how's  the digital system going to replace that? Not in   00:39:53.400 --> 00:39:58.200 the same way, but it's got to do the same  job, right. So you need to talk to people,   00:39:58.200 --> 00:40:03.360 you need to understand the context of their work,  and then you need to carefully plan out how you're   00:40:03.360 --> 00:40:08.960 going to make the transition. So if we just try  to inject AI into workflows or totally replace   00:40:08.960 --> 00:40:15.600 parts of workflows with AI without a really deep  understanding of how that work is currently done,   00:40:15.600 --> 00:40:22.880 what the workers get from it, what is the  value that the workers bring to that process,   00:40:22.880 --> 00:40:27.640 we could go through that chaos. And so it's  really important to get social scientists   00:40:27.640 --> 00:40:32.560 involved in this and good designers, and  that's where the, kind of, multidisciplinary   00:40:32.560 --> 00:40:37.741 thing really comes into its own. That's  where it's really, really valuable. 00:40:37.741 --> 00:40:44.000 HUIZINGA: Yeah … You know, it feels super  important, that book, about a different thing,   00:40:44.000 --> 00:40:49.440 how it applies now and how you can  take lessons from that arc to what   00:40:49.440 --> 00:40:54.840 you're talking about with AI. I feel like  people should go back and read that book. 00:40:54.840 --> 00:41:00.750 SELLEN: I wouldn't object! [LAUGHTER] 00:41:00.750 --> 00:41:05.012 [MUSIC BREAK] 00:41:05.013 --> 00:41:07.520 HUIZINGA:   00:41:07.520 --> 00:41:14.720 Let's talk about some research ideas that are  on the horizon. Lots of research is basically   00:41:14.720 --> 00:41:20.760 just incremental building on what's been done  before, but there are always those moonshot   00:41:20.760 --> 00:41:27.080 ideas that seem outrageous at first. Now,  you're a scientist and an inventor yourself,   00:41:27.080 --> 00:41:30.800 and you're also a lab director, so you've seen  a lot of ideas over the years. [LAUGHS] You've   00:41:30.800 --> 00:41:36.880 probably had a lot of ideas. Have any of  them been outrageous in your mind? And if so,   00:41:36.880 --> 00:41:40.000 what was the most outrageous,  and how did it work out? 00:41:40.000 --> 00:41:45.520 SELLEN: OK, well, I'm a little reluctant to say  this one, but I always believed that the dream   00:41:45.520 --> 00:41:53.080 of AI was outrageous. [LAUGHTER] So, you know,  going back to those early days when, you know,   00:41:53.080 --> 00:41:59.760 I was a psychologist in the ’80s and seeing those  early expert systems that were being built back   00:41:59.760 --> 00:42:08.080 then and trying to codify and articulate expert  knowledge into machines to make them artificially   00:42:08.080 --> 00:42:12.920 intelligent, it just seemed like they were on a  road to nowhere. I didn't really believe in the   00:42:12.920 --> 00:42:21.160 whole vision of AI for many, many years. I think  that when deep learning, that whole revolution’s   00:42:21.160 --> 00:42:27.240 kicked off, I never saw where it was heading. So  I am, to this day, amazed by what these systems   00:42:27.240 --> 00:42:34.440 can do and never believed that these things would  be possible. And so I was a skeptic, and I am no   00:42:34.440 --> 00:42:40.600 longer a skeptic, [LAUGHTER] with a proviso of  everything else I've said before, but I thought it   00:42:40.600 --> 00:42:47.020 was an outrageous idea that these systems would  be capable of what they're now capable of. 00:42:47.020 --> 00:42:49.000 HUIZINGA: You know, that's funny because,   00:42:49.000 --> 00:42:53.640 going back to what you said earlier  about your stepdad walking you around   00:42:53.640 --> 00:43:00.000 and asking you how you'd codify a human into  a machine … was that just outrageous to you,   00:43:00.000 --> 00:43:06.240 or is that just part of the exploratory mode  that your stepdad, kind of, brought you into? 00:43:06.240 --> 00:43:10.400 SELLEN: Well, so, back then I was quite  young, and I was willing to believe him,   00:43:10.400 --> 00:43:17.440 and I, sort of, signed up to that. But later,  especially when I met my husband, a sociologist,   00:43:17.440 --> 00:43:21.560 I realized that I didn't agree with any of  that at all. [LAUGHTER] So we had great,   00:43:21.560 --> 00:43:26.833 I'll say, “energetic” discussions with  my stepdad after that, which was fun. 00:43:26.833 --> 00:43:27.767 HUIZINGA: I bet. 00:43:27.767 --> 00:43:31.080 SELLEN: But yeah, but so, it was how I  used to think and then I went through   00:43:31.080 --> 00:43:36.200 this long period of really rejecting all  of that. And part of that was, you know,   00:43:36.200 --> 00:43:42.040 seeing these AI systems really struggle and  fail. And now here we are today. So yeah. 00:43:42.040 --> 00:43:48.280 HUIZINGA: Yeah, I just had Rafah Hosn on the  podcast and when we were talking about this   00:43:48.280 --> 00:43:52.920 “outrageous ideas” question, she said, “Well,  I don't really see much that's outrageous.”   00:43:52.920 --> 00:43:58.280 And I said, “Wait a minute! You're living  in outrageous! You are in AI Frontiers at   00:43:58.280 --> 00:44:02.661 Microsoft Research.” Maybe it's just because  it's so outrageous that it's become normal? 00:44:02.661 --> 00:44:03.710 SELLEN: Yeah … 00:44:03.710 --> 00:44:10.160 HUIZINGA: And yeah, well … Well, finally,  Abi, your mentor and adviser, Don Norman   00:44:10.160 --> 00:44:15.480 … you referred to a book that he wrote, and  I know it as The Design of Everyday Things,   00:44:15.480 --> 00:44:20.920 and in it he wrote this: “Design is really  an act of communication, which means having   00:44:20.920 --> 00:44:27.520 a deep understanding of the person with whom  the designer is communicating.” So as we close,   00:44:27.520 --> 00:44:33.640 I'd love it if you'd speak to this statement in  the context of AI, Cognition, and the Economy.   00:44:33.640 --> 00:44:38.600 How might we see the design of AI systems as  an act of communication with people, and how   00:44:38.600 --> 00:44:43.840 do we get to a place where an understanding  of deeply human qualities plays a larger role   00:44:43.840 --> 00:44:49.520 in informing these ideas, and ultimately the  products, that emerge from a lab like yours? 00:44:49.520 --> 00:44:54.560 SELLEN: So this is absolutely critical  to getting AI development and design   00:44:54.560 --> 00:45:01.600 right. It's deeply understanding people and  what they need, what their aspirations are,   00:45:01.600 --> 00:45:07.040 what human values are we designing for. You  know, I would say that as a social scientist,   00:45:07.040 --> 00:45:11.600 but I also believe that most of the  technologists and computer scientists   00:45:11.600 --> 00:45:16.520 and machine learning people that I interact  with on a daily basis also believe that.   00:45:16.520 --> 00:45:22.280 And that's one thing that I love about the  lab that I'm a part of, is that it's very   00:45:22.280 --> 00:45:28.320 interdisciplinary. We're always putting the, kind  of, human-centric spin on things. And, you know,   00:45:28.320 --> 00:45:34.480 Don was right. And that's what he's been all about  through his career. We really need to understand,   00:45:34.480 --> 00:45:39.720 who are we designing this technology for?  Ultimately, it's for people; it's for society;   00:45:39.720 --> 00:45:44.680 it's for the, you know, it's for the common  good. And so that's what we're all about. Also,   00:45:44.680 --> 00:45:50.360 I'm really excited to say we are becoming, as  an organization, much more globally distributed.   00:45:50.360 --> 00:45:57.040 Just recently taken on a lab in Nairobi. And  the cultural differences and the differences   00:45:57.040 --> 00:46:04.640 in different countries casts a whole new light  on how these technologies might be used. And so   00:46:04.640 --> 00:46:09.240 I think that it's not just about understanding  different people's needs but different cultures   00:46:09.240 --> 00:46:16.141 and different parts of the world and how this  is all going to play out on a global scale. 00:46:16.141 --> 00:46:21.840 HUIZINGA: Yeah … So just to, kind of, put a  cap on it, when I said the term “deeply human   00:46:21.840 --> 00:46:28.920 qualities,” what I'm thinking about is the way we  collaborate and work as a team with other people,   00:46:28.920 --> 00:46:33.640 having empathy and compassion,  being innovative and creative,   00:46:33.640 --> 00:46:38.600 and seeking well-being and prosperity.  Those are qualities that I have a hard   00:46:38.600 --> 00:46:46.560 time superimposing onto or into a machine.  Do you think that AI can help us? 00:46:46.560 --> 00:46:53.480 SELLEN: Yeah, I think all of these things that  you just named are things which, as you say,   00:46:53.480 --> 00:47:00.000 are deeply human, and they are the aspects  of our relationship with technology that we   00:47:00.000 --> 00:47:07.000 want to not only protect and preserve but  support and amplify. And I think there are   00:47:07.000 --> 00:47:13.320 many examples I've seen in development  and coming out which have that in mind,   00:47:13.320 --> 00:47:19.440 which seek to augment those different  aspects of human nature. And that's   00:47:19.440 --> 00:47:24.240 exciting. And we always need to keep that in  mind as we design these new technologies. 00:47:24.240 --> 00:47:30.720 HUIZINGA: Yeah. Well, Abi Sellen, I'd love to  stay and chat with you for another couple hours,   00:47:30.720 --> 00:47:36.480 but how fun to have you on the show.  Thanks for joining us today on Ideas. 00:47:36.480 --> 00:47:56.840 SELLEN: It's been great. I  really enjoyed it. Thank you. 00:47:56.840 --> 00:47:57.340 [MUSIC]