00:00 Welcome,
00:00 welcome back everyone.
00:02 Please
00:03 take your seats
00:04 and uh we're about to start the session,
00:06 the 2nd session today,
00:07 which is all about the future,
00:09 specifically
00:10 how artificial intelligence is revolutionizing technology,
00:13 reshaping competition policy,
00:15 and
00:15 redefining education.
00:17 We'll hear from experts about the opportunities
00:20 and risks of AI for developing countries
00:22 and how policymakers can use this technology for inclusive growth.
00:26 Moderating the session is Mr.
00:28 Gurav Nayar,
00:29 director of the World Development Report 2026,
00:32 and he'll introduce our three panelists and one discussion.
00:35 So over
00:36 to you,
00:36 please.
00:44 Hi,
00:44 good afternoon,
00:45 everyone,
00:45 to the post-lunch siesta session.
00:48 Uh,
00:49 but I,
00:50 I'm hoping that just the mention of the
00:52 word artificial intelligence will wake everyone up.
00:54 So,
00:55 thank you for being here.
00:56 Uh,
00:57 the esteemed guest panelists had a
01:00 preference to not be introduced as rock stars and entered
01:02 the stage one by one with the spotlight on them.
01:05 So,
01:05 they've all taken their seats.
01:07 Um,
01:08 I just want to quickly give a very,
01:10 very quick introduction to the session.
01:12 In the morning we heard about populism and
01:16 populism is a reaction to
01:19 distributional concerns both
01:21 between countries and within countries,
01:23 and I think the discussion around AI
01:26 dovetails almost perfectly into this discussion on how
01:30 Uh,
01:30 the diffusion of the technology might affect
01:33 the distribution of incomes both between
01:36 and within countries.
01:37 Um,
01:38 here at the bank,
01:39 I think there's a lot of optimism around the development potential of AI.
01:45 Uh,
01:46 just in the sense of being the technology being able to solve
01:50 market failures and addressing long-standing development challenges.
01:54 Think about access to credit markets
01:56 where AI can devise new ways of assessing credit histories.
02:00 Think about filling in skill gaps,
02:01 say,
02:02 in education and health services where quality teachers
02:05 and health service providers are often lacking,
02:08 or just making
02:09 the large numbers of small businesses
02:11 more profitable through AI advisory services.
02:15 However,
02:15 such optimism is not unbridled.
02:18 Uh,
02:18 I think everyone recognizes and is aware
02:21 of the risks that come through growing digital divides,
02:25 uh,
02:25 competition concerns,
02:27 and also risks of automations and,
02:29 and worker displacement.
02:31 Uh,
02:31 so the aim of today's session is really to
02:34 cut through some of the hype around AI.
02:37 Provide some conceptual clarities around
02:40 specific sets of issues
02:42 and then also present some new evidence,
02:44 uh,
02:44 which will hopefully give us a better sense
02:47 of the potential uh development impacts of AI,
02:50 uh,
02:50 going forward.
02:51 To do that,
02:52 we have a stellar,
02:53 um,
02:54 set of panelists,
02:56 um,
02:57 to begin with,
02:57 we will have Arvind Narayinan,
02:59 who is
03:00 professor of computer science at Princeton University,
03:03 talk about AI as a normal technology.
03:06 Uh,
03:06 after that,
03:07 we will have Katherine Tucker,
03:09 the Sloan Distinguished Professor of Management at MIT,
03:13 talk about how does competition policy need to
03:16 change in a world of artificial intelligence.
03:20 Last but not least,
03:21 we would have Dan
03:23 Bjorkeren from Columbia University.
03:26 Who will talk about could AI leapfrog the web
03:29 and provide evidence from teachers in Sierra Leone.
03:33 And also immediately after the presentation of the three papers,
03:36 we will move into a panel discussion
03:38 where we will also be joined by
03:40 Han Cheng Xia,
03:41 who's a policy fellow
03:42 at the Center for Global Development.
03:45 So with that,
03:46 I'd like to invite Arvind
03:48 to give us the first presentation.
03:49 Thank you.
03:54 Thank you Gurav and hello everybody.
03:56 Good afternoon.
03:57 My goal for the next 20 minutes
04:00 is to say things that will sound obvious to you hopefully.
04:04 Uh,
04:05 that's an unusual goal for a talk,
04:06 so let me explain why that's my goal.
04:08 Perhaps if we could,
04:09 uh,
04:09 have the slides up in the meantime.
04:12 Now there are many technologists,
04:14 AI researchers,
04:15 developers,
04:16 and industry CEOs who have predicted imminent
04:20 and rapid transformation because of AI,
04:22 that we will have
04:23 super intelligence in the next couple of years,
04:26 before the end of the decade,
04:27 certainly,
04:28 and that this might obviate the need for human
04:30 labor and perhaps even the concept of money,
04:32 and that we should be
04:33 worried about AI taking over the world.
04:36 The more modest of these predictions are talking about 10% growth annually.
04:41 But as economists,
04:42 a lot of this might seem
04:44 strange.
04:45 It seems to be from the realm of sci-fi,
04:47 so it might be natural to wonder,
04:49 do these folks know something that we don't know.
04:52 And I'm here as a computer scientist to say I think
04:55 the answer is no.
04:56 I have,
04:56 I have looked at these predictions and the research behind them,
05:00 and my view
05:01 is that
05:02 together with my co-author Sayash Kapoor,
05:05 that AI is not going to be that different from
05:07 many past waves of technological disruption that we've seen.
05:10 Of course,
05:11 no two technologies are exactly like each other,
05:13 but there is a lot that we can learn from history.
05:16 And so that's what our Paper AI as normal technology is all about,
05:19 we're
05:19 expanding it now into a book.
05:21 So
05:22 the paper has 4 parts.
05:23 The first part is about timeline
05:25 on what time scale will AI's transformative effects unfold,
05:29 and the second part is about,
05:30 OK,
05:30 whatever time scale that is,
05:31 what will that world look like?
05:33 Now
05:34 we accept
05:35 the possibility of cognitive automation that most of the cognitive
05:38 tasks that we do today might one day be automatable,
05:41 and yet we do not envision something radical such as
05:44 human labor being superfluous.
05:46 So why is that?
05:47 What is the resolution to that seeming contradiction?
05:50 The third part I'm going to mostly skip for
05:52 today what do we do about potential superintelligence,
05:54 etc.
05:55 and then I'll briefly spend a couple of minutes on policy implications,
05:59 and I look forward to the discussion.
06:01 OK,
06:01 so our starting point
06:03 is well established patterns of the diffusion
06:06 of general purpose technologies that we've seen over and over.
06:09 The first stage is invention.
06:11 So if we think about electricity,
06:12 for instance,
06:13 the invention of electromagnetic principles,
06:15 AC and DC,
06:16 those are all part of invention.
06:18 But people don't directly use electricity.
06:20 We use electrical appliances,
06:22 so that requires innovation.
06:23 It requires the grid.
06:24 It requires a whole lot of other complementary things.
06:27 And then finally that technology gradually diffuses
06:30 throughout society and throughout the economy.
06:32 So we took this
06:33 and we
06:34 expanded it and adapted it for AI.
06:36 So we introduced one extra stage and we talked about
06:39 what each These stages look like in the context of AI.
06:42 So the first invention stage is about the increasing capabilities of models.
06:46 So today,
06:47 the models,
06:47 the kind of model that's at the frontier,
06:50 are large language models.
06:51 It wasn't always the case.
06:53 It might not necessarily always be the case,
06:55 but for now you can think of
06:56 the invention stage as improvements in LLM capabilities.
07:00 We don't necessarily directly use LLMs for most tasks.
07:03 We use it through the integration of those LLMs into various applications.
07:08 So throughout this slide,
07:09 I'm using software engineering itself as an example,
07:12 because it is one
07:14 occupation that's really at the forefront of
07:17 integration of AI into uh into workflows.
07:20 So,
07:22 Even even though software engineering is pretty advanced
07:27 along this
07:28 causal chain,
07:29 if you will,
07:30 we're still pretty early in figuring out
07:32 how to architect AI code editors.
07:34 We're having really fundamental debates in
07:36 the computer science community about whether
07:38 command line interfaces are better or graphical interfaces are
07:41 better when it comes to software engineering using large.
07:43 Language models,
07:44 right?
07:44 So it seems like very early days.
07:47 And then you have the early adoption phase where people are
07:49 learning how to use these tools.
07:51 That phase is very much underway.
07:52 You have things like vibe coding where people are able to
07:55 create small apps using AI,
07:57 but when they start to edit,
07:58 you know,
07:58 billion line enterprise code bases using AI,
08:01 it's not going well so far.
08:02 We're not yet in the adaptation phase.
08:04 We're not yet in the phase where people are.
08:06 Figuring out
08:07 how to rearchitect
08:09 software both technically as well as software companies to
08:11 best take advantage of the capabilities of AI tools.
08:15 So there was a study that showed that right now most software engineers are not yet
08:19 achieving productivity benefits even though they think they are.
08:22 Their self-reported
08:23 time savings are actually
08:25 very positive,
08:25 but actual reported time savings are negative.
08:28 So.
08:29 Our view here is that adaptation is going to require a lot of changes.
08:33 One example is what I'm calling extreme personalization.
08:36 This is all well into the future,
08:37 so it's necessarily speculative,
08:39 but let me explain what I mean.
08:40 So if it's going to be the case
08:42 that AI's ability to write code
08:44 will dramatically decrease the cost of producing software,
08:47 then the manner in which we're going to
08:48 use software is going to change dramatically.
08:51 It will no longer make sense for one software company to produce
08:54 one piece of software and then millions of other individuals and other companies
08:57 to all be shoehorned into that same workflow,
09:00 the same set of abstractions that's defined into that software.
09:03 The way we produce software
09:05 might be in a very in-house way.
09:07 Each company is producing its own software for its own teams and for its own users.
09:11 So in other words,
09:12 if the cost of software drops,
09:14 the
09:15 elasticity of demand for software might be very high.
09:18 We might see extremely personalized software where
09:20 everybody is using software tailored to themselves.
09:23 So if you think about that magnitude of transformation.
09:25 Hopefully it's intuitively clear why that feels like a few decades away as opposed
09:29 to something that's going to happen next
09:30 year because of increasing AI capabilities.
09:33 So that's really the central idea of the paper,
09:34 why we think these timelines are going to be pretty slow.
09:38 OK,
09:38 so let me focus though on the first era,
09:41 the gap between the model capabilities and the products.
09:44 Now in the first year after Chat GPT was released,
09:47 many companies tried to create these simple wrappers
09:50 around these large language models and build these personal agents,
09:53 but they failed terribly in the market.
09:55 There's a whole graveyard of products like this.
09:57 So here's one where
09:59 this product,
10:00 you know,
10:00 does a pretty good job of automatically ordering food,
10:03 so it's an impressive capability,
10:04 but it ordered it to the wrong address.
10:07 Right,
10:07 so we've had unreliable AI like this before.
10:10 If traditional Siri or Alexa played the wrong song 10% of the time,
10:14 that's just annoying.
10:15 But if a product orders
10:17 something,
10:17 you know,
10:17 where you spent money to the wrong address 10% of the time,
10:20 that's dead on arrival,
10:22 right?
10:22 So this is something we call the capability reliability gap,
10:25 and we think there's years of research to go before
10:27 we smooth out these kinks and build reliable products using large language models.
10:32 OK,
10:33 so
10:34 the capability reliability gap is one reason why the
10:37 way we're evaluating AI today doesn't make sense.
10:39 We're evaluating models and assuming that that means useful products,
10:42 but that's not the case.
10:44 Another reason is construct validity.
10:47 So for instance,
10:48 there are studies that
10:50 AI can pass the bar exam,
10:52 but it's not as if lawyers' job is to answer bar exam questions all day,
10:55 right?
10:55 So
10:56 The evaluations are looking at the wrong construct,
10:59 so we don't have a lot of information about how useful these are
11:02 to actual workers unless you do RCTs of those workers using those tools,
11:06 and we have lots of anecdotal evidence now that the hype doesn't match the reality.
11:10 The RCTs in many domains are still just getting started,
11:13 so this is a company DoNotPay that got
11:15 a lot of attention by saying they'd built a robot lawyer,
11:18 and any attorney who used it to argue
11:20 in front of the Supreme Court would get
11:22 $1 million.
11:24 First of all,
11:24 this was just a publicity stunt.
11:26 Electronic devices are not even allowed in the Supreme Court,
11:28 and
11:29 of course this company knew this,
11:31 but the reason so many people believed this was even possible was because they
11:34 had believed the headlines and they thought that a robot lawyer was imminent.
11:37 But when you look at lawyers actually using AI to do the hard parts of being a lawyer,
11:41 we're,
11:42 you know,
11:42 it's still very,
11:43 very far to go,
11:45 and we teamed up with our colleague
11:47 Peter Henderson,
11:48 who's a JD PhD,
11:49 and we wrote a paper breaking down the different types
11:52 of tasks that are involved in being a lawyer.
11:55 And it turns out,
11:55 and this is the central reason why so many technologists are overconfident,
11:59 the tasks that are easier to measure,
12:00 which is the top bucket,
12:02 are also the ones that are much easier for AI.
12:04 So we just measure the things that are easy to measure
12:06 and conclude that AI is doing really well at it.
12:11 OK,
12:11 so,
12:12 so far I've talked about the first error,
12:14 but the much harder thing is diffusion,
12:16 the second error.
12:17 So let me talk about that a little bit.
12:19 So one important point I want to emphasize is
12:22 that it doesn't matter how fast capabilities are increasing.
12:26 Adoption of these technologies requires behavioral changes,
12:28 and behavioral changes are going to happen
12:30 at a certain
12:31 time scale based on how fast humans can adapt,
12:33 and that's not really changing very much.
12:35 There was this paper that made the headlines.
12:37 It's a good paper.
12:37 I have no concerns about the paper,
12:39 but
12:40 the thing that the media Picked up is that 40%
12:42 of US adults are using generative AI as of,
12:45 you know,
12:45 about a year ago,
12:45 so two years after chat GPT.
12:47 That seemed very striking
12:49 and it's framed as rapid adoption of generative AI.
12:52 I don't think this tells the whole story
12:54 because what the paper also measured.
12:56 Excuse me,
12:57 is the intensity of adoption and productivity effects,
12:59 and those are still very low.
13:01 So there is still a very long way to go
13:03 because
13:04 that 40% number
13:05 includes even someone who used chat GPT once a week to write a limerick or something.
13:10 It's not necessarily looking at how much productivity
13:12 improvements people are getting in the workplace.
13:16 OK,
13:17 and think about this.
13:18 I mean,
13:19 so often,
13:20 and I'm part of this,
13:21 I often forget that Google Search exists and things that,
13:23 you know,
13:24 I'm debating with other people can be quickly resolved through a Google search,
13:27 and this person is telling a story of
13:29 how much
13:29 of a competitive edge you get just from remembering
13:32 to Google the things that can be Googled,
13:33 and that's a quarter century after Google Search was released,
13:36 right?
13:36 So that's kind of the timescale
13:38 on which these behavioral adaptations happen.
13:41 OK,
13:42 so that's on the speed of diffusion.
13:43 However,
13:44 one of the things we point out in the paper
13:47 is that
13:47 in the last few years,
13:48 we've been in an unusual period of AI innovation,
13:52 and that has already ended.
13:53 And what I mean is this,
13:55 the reason why we had such rapid AI capability advances.
13:59 Was building bigger and bigger models using
14:01 bigger and bigger slices of the internet
14:03 that has mostly ended,
14:05 you know,
14:06 not completely,
14:07 but for the most part,
14:08 and what it means going forward is that these models will have to be improved
14:12 using experiential data from their use in the field by actual workers,
14:16 and that's going to be a feedback loop that's going to happen much
14:19 slower and we're already seeing those much slower cycles of progress kick in.
14:24 And
14:25 we can reasonably predict what that will look like
14:27 by looking at sectors in which those
14:29 cycles of feedback loops have already happened,
14:33 notably
14:33 self-driving cars.
14:35 So what happened in self-driving cars is that 2-3 decades
14:37 ago we had convincing prototype demonstrations of self-driving cars.
14:41 However,
14:42 they weren't yet reliable enough to autonomously put on the roads.
14:45 So what car companies had to do.
14:47 Or you know AI companies had to do
14:49 was to get it reliable enough to drive 100 miles,
14:52 drive it 100 miles,
14:53 get enough data,
14:54 improve their reliability so that you're comfortable driving at 1000 miles,
14:57 and so on and so on.
14:58 So each order of magnitude of scaling takes a couple of years.
15:01 So we've seen those slow feedback loops,
15:03 feedback loops play out,
15:04 and we predict
15:05 that the same thing is going to happen in AI for law,
15:07 AI for medicine,
15:08 and every other sector.
15:12 And finally,
15:13 what we say about the speed of change is that
15:14 a lot of this might require changes to organizations,
15:17 structural transformations,
15:18 changes to
15:20 even laws and norms,
15:21 and so forth.
15:22 We use the analogy of how slow the electrification
15:24 of factories was because they had to figure out
15:27 how to reorient the whole assembly line and so on.
15:30 And you know,
15:32 looking at that example of software again,
15:34 so if it's going to be the case that it makes no sense to have these
15:37 large software companies and every individual company is
15:39 going to have its own software engineers,
15:41 that is a kind of
15:43 slow
15:44 transformation of the whole sector
15:45 that is again going to play out
15:47 on a time scale of decades.
15:49 OK,
15:50 so there's lots of other examples of this,
15:52 but really the The summary of the first part is that we identify speed limits
15:56 in every stage of this invention to
15:58 innovation to early adoption to diffusion pipeline.
16:01 I haven't gotten into the details of most of the speed limits,
16:04 but they're in the paper,
16:05 and that is what leads us to conclude
16:07 that we're not going to see some kind of sci-fi transformation in 2027,
16:11 but rather we're going to see gradual diffusion throughout the next few decades.
16:16 So
16:16 let me spend 5 minutes on part 2 and then maybe 2 minutes on part 4
16:21 because I know I'm running out of time.
16:23 So
16:24 again,
16:24 like I said earlier,
16:25 so we accept the possibility that it might be the
16:27 case that one day most cognitive tasks that we do,
16:30 most things we do sitting in front of the computer
16:32 are going to be automatable.
16:34 But let's think about what that will mean.
16:36 So there have been concerns about mass job
16:38 displacement because of automation for many years.
16:41 This was a headline from
16:42 15 years ago.
16:43 However,
16:45 from what I understand of the literature and economics,
16:47 those are naive ways of thinking about things.
16:49 There is the task model that goes back to Ator,
16:52 Levy,
16:53 and Murnane more than two decades ago and has been.
16:55 Refined over the decades
16:57 we can kind of see jobs as bundles of tasks,
16:59 and each job might have,
17:00 you know,
17:01 several
17:02 dozen tasks as part of the bundle.
17:04 Not every task gets automated at the same moment in time,
17:07 and as one task gets automated,
17:09 new tasks get created.
17:10 The job composition changes.
17:12 Automation is counterbalanced by a reinstatement effect.
17:15 And it could even be the case that if it becomes
17:18 cheaper for lawyers to provide their services because of automation,
17:21 there's more demand for legal work and therefore
17:23 the total employment of lawyers actually increases.
17:25 OK,
17:26 so that much hopefully is straightforward,
17:28 but now
17:29 I want to make a couple of points that are even more skeptical
17:33 of mass automation compared to what you might take away from the literature.
17:38 in labor economics
17:39 and for that I want to give you an example.
17:42 So radiology
17:43 was famously predicted
17:44 by computer science pioneer Geoff Hinton and Nobel Prize winner
17:48 to be going obsolete because AI can automate the reading of scans,
17:52 and he said almost a decade ago
17:54 that if you work as a radiologist,
17:56 you're like the coyote that's already over the edge
17:58 of the cliff but hasn't yet looked down.
18:01 So in my world,
18:01 this is a famous prediction because of how wrong it was.
18:04 The demand for radiologists has only gone up
18:06 even though
18:08 radiologists are not
18:10 shy of the tech.
18:11 They have enthusiastically adopted it,
18:13 and it has had an effective augmentation rather than automation.
18:17 And interestingly,
18:19 this does not appear to be because there is some particular task that we can identify
18:23 at which
18:24 human radiologists still outperform AI.
18:27 And this is the central claim that I want to make.
18:30 I think the task model of jobs is incomplete.
18:32 Yes,
18:33 to a first approximation,
18:34 you can break down jobs into tasks,
18:36 but think about your own job.
18:38 Can you identify a series of tasks such that if each of those tasks was automated,
18:43 then your job as a whole would be automated?
18:46 I would,
18:47 I would say no,
18:47 and I would imagine that most of you would say no,
18:49 and you might even find the question to be a little bit silly,
18:52 and yet
18:53 that's exactly how I think a lot of this
18:55 literature approaches the question of AI and job displacement.
19:00 So in my understanding of radiology and many other professions,
19:03 a lot of the hardest to automate things are at the boundaries between tasks.
19:07 So by the time you've broken down
19:09 jobs into a set of tasks that you can put
19:11 into the OED database that everybody seems to use,
19:14 you've already lost the most nuanced and interesting aspects of
19:17 jobs that are going to be hardest to automate.
19:19 And then you have to consider
19:21 human preferences for dealing with other humans
19:23 instead of AI and various other things.
19:25 So considering all this,
19:26 we're very,
19:27 very skeptical about some of these claims of,
19:30 you know,
19:31 broad-based labor displacement due to AI.
19:35 So and that leads to the point,
19:37 I think there are a lot of these predictions made by,
19:39 you know,
19:40 computer scientists like us,
19:42 made by economists who are not necessarily experts in
19:45 the specific domains that they're making predictions about.
19:47 I think there is some value in those.
19:49 I mean,
19:49 it is a lot of what I do
19:50 and I'm very much turning this criticism back onto myself as well.
19:53 But I think it would be much more useful
19:56 if we teamed up
19:57 with experts in different domains who have a
19:59 much more nuanced understanding of what tasks entail,
20:02 what are the boundaries between tasks,
20:03 and so forth
20:04 to make predictions and recommendations
20:06 about those specific occupations and sectors
20:08 as opposed to these overarching,
20:10 sweeping pronouncements.
20:12 OK,
20:13 so having said that,
20:13 I'm going to return to the mode of making sweeping pronouncements.
20:17 Um,
20:17 so one thing I want to point out is that
20:19 there is actually a way to reconcile some of the radical
20:22 optimists and the radical skeptics,
20:24 skeptics,
20:25 and we can do that by looking back at the effect that the internet has had.
20:29 Now.
20:31 Let's accept for a second that generative AI will
20:34 not necessarily automate but at least mediate most knowledge work
20:37 just like the internet did.
20:38 If you went back in time 30 years and told people that most of the
20:41 cognitive work that they do will one day
20:43 be mediated by this new communication technology,
20:46 they would have said that's ridiculous.
20:48 A lot of people did say that's ridiculous,
20:50 and yet that happened
20:51 even though it sounded crazy at the beginning.
20:53 What were its effects?
20:54 The effects on productivity,
20:56 I would say,
20:57 have been relatively small and only while it's
21:00 true that some professions have been decimated,
21:02 for the most part we see an augmentation effect,
21:05 and this all unfolded gradually over a period of a few decades.
21:09 And therefore our thesis is
21:11 that you can have workplaces that look in some ways radically different
21:15 and yet
21:16 without necessarily having,
21:17 you know,
21:18 transformative effects on total factor productivity
21:20 growth or something like that.
21:22 So I'll skip this slide
21:25 to be a little bit flippant about it,
21:27 our view on the role of labor in a world with advanced AI is that
21:32 even though everything might change in a qualitative sense
21:35 in terms of our experience of what work looks like
21:38 in an economic sense,
21:39 things might mostly stay the same.
21:41 So one thing I want to add to
21:43 this is whether superintelligence will change the equation.
21:46 A lot of people who have very different predictions from us will say,
21:49 no,
21:49 no,
21:49 this is what past technologies have done,
21:51 and so far AI is on the same path,
21:54 but something is going to change in a couple
21:55 of years where recursive self-improvement of AI becomes possible,
21:59 and then it will be an exception to all of these well established patterns.
22:02 We have a whole section in the paper on this,
22:04 but I'll just spend one minute on it.
22:06 We say no.
22:06 And the reason we say no is that we have a very different view of what
22:10 human intelligence entails and to what extent AI
22:13 will be able to exceed human intelligence.
22:16 You can identify three types of limits to our capabilities.
22:19 We kind of suck at chess,
22:21 and AI is much better than even the
22:23 human world champion because of our computational limits.
22:27 But
22:28 for most tasks,
22:29 the limits are not computational.
22:30 There are things like writing where there are intrinsic limits to how persuasive,
22:34 let's say,
22:34 a piece of writing can get,
22:35 or
22:36 there are knowledge limits.
22:37 Some people say AI is going to be a superhuman medical researcher,
22:41 and we're like,
22:42 what?
22:43 Let's take a second to imagine what that means
22:45 for AI to be a superhuman medical researcher.
22:48 It can only happen if we allow AI to perform
22:50 autonomous experiments on thousands of people without any supervision.
22:54 Right,
22:55 we will never allow it to do that.
22:56 So there are intrinsic knowledge limits to how good AI can get.
22:59 So we reject this notion that superintelligence is going to introduce
23:03 a new causal pathway that's gonna
23:05 blow past all of these limits that are identified from,
23:08 uh,
23:08 from previous research.
23:10 Uh,
23:10 OK,
23:10 so I will stop here in the interest of time.
23:12 We have lots of thoughts on what policymakers should do,
23:15 but maybe we can get into that during the discussion period.
23:17 Thank you so much.
23:32 Can I just say how brilliant that was?
23:34 And we've got a computer scientist summarizing all labor
23:39 economics and AI for the last 10 years.
23:41 It was really well done.
23:42 Um,
23:43 is there a chance of my slides?
23:45 Brilliant.
23:45 OK.
23:46 So,
23:47 um,
23:47 now for something completely different because I'm an IO economist,
23:52 which means I've been tasked not to talk so much about labor,
23:56 but to talk a little bit more about how,
23:58 AI will change
24:00 how firms relate to consumers and
24:03 how that will change in the world of AI.
24:05 Now,
24:06 um,
24:09 I have,
24:10 was asked
24:11 to talk about a paper
24:14 I wrote on this topic,
24:16 um,
24:16 for this panel.
24:18 And
24:19 I must admit,
24:20 when I was asked this,
24:22 I was like,
24:22 oh,
24:24 that's my paper which has got least to do
24:28 with any low-income countries.
24:30 It's completely obsessed with America,
24:33 completely inappropriate for this audience.
24:34 So,
24:35 What am I going to do with the fact that this paper,
24:38 um,
24:39 is,
24:39 you know,
24:39 very US obsessed?
24:41 What I'm gonna do is I'm going to set out the sort of broader lessons,
24:46 because in some ways,
24:47 the reason I wrote this paper
24:49 was because much like Arvind,
24:52 I often sit on panels
24:54 with lots of
24:57 Policy people
24:59 who have
25:01 very dramatic perspectives about AI.
25:04 And you sit there,
25:05 and the general view about competition policy and how firms relate to consumers and
25:10 how everything's going to change is that everything is going to change very,
25:13 very swiftly.
25:14 And then there's always one person a little bit more like Arvin,
25:16 who says,
25:17 No,
25:17 no,
25:17 I'm going to be different.
25:19 Um,
25:20 and then I'm gonna be more economists and be slightly in the middle and say,
25:24 look,
25:24 in the end,
25:25 uh,
25:26 economics doesn't change.
25:27 That's the good news if you're an econo economist.
25:30 Um,
25:30 on the other hand,
25:31 process change is gonna be important.
25:33 Now,
25:35 the great thing about writing this paper for me
25:38 was that I wrote it in response to being on a lot of policy panels
25:44 and really taking it as a chance to try and educate lawyers largely
25:50 about how economists think about artificial intelligence.
25:55 And I have run the,
25:57 um,
25:57 NBR meetings on digital economics and AI
26:02 for nearly a decade now.
26:04 And so I want to just take this opportunity,
26:06 not just,
26:07 uh,
26:07 given the article,
26:09 to just
26:10 tell this group a little bit about the history
26:12 of the economics of AI.
26:16 So I think it's fair to say we started off
26:19 back in 2018
26:22 with our first sort of NBR official meeting
26:26 about the economics of AI.
26:29 And what was the revelation back in 2018?
26:34 I think the revelation for economists back in 2018
26:38 is that AI was not about robots.
26:42 I think we never really written many papers about
26:44 it because we thought it was about robots,
26:46 and therefore about Terminator and Judgment Day.
26:49 And instead,
26:50 what we worked out was instead that really if you think about AI,
26:55 um,
26:55 it is a cost saving technology.
27:00 And the moment you start to describe AI as a cost-saving technology,
27:05 it is something which economists start to think,
27:07 uh,
27:08 they can analyze,
27:08 unlike Terminator.
27:10 So,
27:11 um,
27:12 The sort of forward to back in 2018 is the way to think about it
27:18 is that AI lowers the cost of prediction.
27:23 And as you may have guessed from Arvin's,
27:26 uh,
27:26 talk,
27:27 who,
27:27 who,
27:28 um,
27:28 he talked about Hinton,
27:29 what he said about radiologists back in 2016,
27:32 we were obsessed
27:35 about
27:36 radiologists.
27:38 And
27:39 what,
27:39 where we came into the radiologist debate is we said,
27:42 and we thought we're being very smart,
27:44 is that,
27:45 look,
27:46 uh,
27:46 you know,
27:46 the prediction is that radiology is going away.
27:50 But in the end,
27:51 as economists we believe in complementarity and the complementarity of tasks.
27:57 And therefore really what's going to happen is when you reduce
28:01 the costs of doing certain types of tasks such as prediction.
28:05 Other forms of humanity are going to be
28:08 more important than the one we emphasized there
28:11 was judgment,
28:12 or the idea that there's going to be
28:15 large returns to judgment in our AI,
28:18 uh,
28:18 economy.
28:19 Now,
28:20 you know,
28:20 at the time,
28:22 we were all joking around.
28:23 I remember quite a few laughs were had at our 2018 conference
28:28 by someone suggesting that,
28:31 you know,
28:31 what we should all be doing is encouraging our children to become
28:35 journalists and poets because those occupations were not going to be disrupted.
28:41 What I learned from that is that economists who study AI are usually wrong.
28:45 Because,
28:46 you know,
28:46 back in now in 2025,
28:48 that sort of viewpoint,
28:50 that AI is not coming for creativity seems completely wrong,
28:53 uh,
28:54 especially if you're a poet.
28:56 So,
28:57 um,
28:58 what I've learned to sort of,
29:00 when I talk about the economics of AI more broadly,
29:02 especially to PhD students,
29:04 what I've always tried to do now is to encourage them to be a little bit more humbled,
29:08 express humility,
29:09 don't have our hubris that we had back in 2018.
29:13 And I was returned to this Oxford English dictionary definition.
29:19 Which is that in the idea,
29:21 in the end,
29:21 what is AI?
29:23 It's
29:24 The
29:27 set of systems and technologies
29:29 which is capable
29:31 of performing tasks which require human intelligence.
29:35 And what I want you to notice is that is a moving target.
29:40 That definition can never be true or stable in the long run.
29:45 Because
29:45 the nature of AI is such that,
29:49 of course,
29:49 we're always going to be finding new ways
29:52 of replacing human intelligence.
29:55 So,
29:56 I've learned to be sort of very humble about how I talk about AI.
30:00 And that was one of the things I was trying to get across
30:03 in,
30:04 uh,
30:04 in my article,
30:06 which was
30:07 really a debate,
30:07 as I say,
30:08 with antitrust lawyers.
30:10 Now,
30:10 the one thing that,
30:11 uh,
30:12 all AI economists do agree about
30:15 is very much,
30:16 and this is so similar,
30:18 and that's why I was congratulating Arvind on,
30:19 on his talk,
30:20 in that it was a,
30:22 uh,
30:23 the one thing we all agree about
30:25 is that AI is probably a general-purpose technology.
30:28 And that makes technology economists happy.
30:31 Why?
30:31 Because we sort of have so many examples littered around.
30:35 And now as you'll have gathered,
30:37 you know,
30:37 one of our favorite examples is electrification and making that analogy.
30:42 And all I want to really add to the sort of
30:45 very good evidence that we just heard in the previous talk
30:49 is to remind everyone in the room that when
30:52 electricity was first introduced,
30:56 Everyone thought it was going to be about illumination.
31:00 We all thought that the jobs that were going to be
31:03 gone are the people who go around the streets lighting gas,
31:06 gas lamps.
31:08 We had no idea of where the real impacts would come from,
31:12 which was the electrification of factories.
31:17 And again,
31:18 in 2025,
31:19 I think we're probably at that stage.
31:22 In that we're still really thinking about use cases,
31:26 which are the analogy
31:28 of illumination rather than really reflecting
31:32 fundamental process change,
31:34 which is what underpins general purpose technologies.
31:39 OK,
31:40 so I've got to set all this out in the article about the history
31:44 of the economics of AI and how economists have tended to think about it.
31:48 And then I realized that what I'm really doing is we're
31:53 I've set myself the task of writing an article about a technology which is fluid,
31:57 unpredictable,
31:58 has no established business model,
32:01 and bring it to competition policy.
32:03 And that's,
32:04 if you know anything about competition policy,
32:06 there's a big mismatch here.
32:08 And why is that?
32:09 Well,
32:09 in the end,
32:10 competition policy is the belief that governments
32:13 do good
32:14 by encouraging competition.
32:17 And that competition is a way of benefiting the citizenry and the average consumer,
32:23 and that's the aim of competition policy.
32:27 Now,
32:27 the reason I say that the unpredictability of AI
32:32 poses trouble for competition policy.
32:36 Is that in the end,
32:37 the way that competition policy tends to work.
32:42 Is that it tends to be very retrospective.
32:46 Um,
32:46 if you think
32:48 about US competition policy,
32:50 uh,
32:50 often it is implemented through,
32:52 uh,
32:53 the courts,
32:54 and you are therefore always going to be looking backwards.
32:59 And again,
32:59 we've got this technology which is changing its nature every 2 years.
33:03 Uh,
33:04 the other thing I'd say about competition policy is
33:07 across the world,
33:07 there's generally a set of norms about how you conduct it.
33:11 And
33:13 how we think and analyze substitution by people,
33:16 uh,
33:17 across products and firms.
33:19 And there's also a lot of norms of which are sort of established
33:23 in the competition policy tends to be very obsessed with looking at price.
33:28 If you look at all the tools that our economists tend to bring to competition policy,
33:34 they require you to look at a price.
33:36 And now I've got to say something about,
33:39 uh,
33:40 technology where we literally don't know what the business model is.
33:43 Uh,
33:44 we have lots of tech companies spending billions and
33:47 billions of dollars on it with no return,
33:49 giving a lot away for free.
33:50 And so again,
33:51 that is going to make it difficult to predict.
33:54 Now,
33:55 when I said I was going to give this talk,
33:57 I inwardly panicked
33:59 and reached out to Daniel.
34:00 Why,
34:01 Daniel?
34:01 Well,
34:01 you'll hear in the next talk,
34:03 Daniel's a brilliant development economist.
34:05 Good news.
34:06 And he's gonna talk about AI in developing countries.
34:09 Again,
34:10 good news.
34:11 And I was like,
34:12 you know,
34:12 sort of how do I say,
34:14 you know,
34:14 how do I bridge this more to a group of development economists?
34:18 And he was like,
34:18 no,
34:18 no,
34:19 Catherine,
34:19 you're really underestimating the extent to which antitrust is important.
34:22 You know,
34:23 for example,
34:23 look at these evolutions in Africa.
34:26 You know,
34:26 I'm going to sort of give credit to that and let Daniel talk about that.
34:30 But what I would just say is that I think this
34:32 debate is so important about how we regulate competition around AI,
34:36 um,
34:37 even if at the moment it has a US focus.
34:40 And the reason is this,
34:41 is,
34:42 you know,
34:42 if we were to look backwards,
34:44 as competition economists often do.
34:47 At
34:47 what's happening
34:49 in antitrust and technology more generally.
34:53 Decisions being made about competition policy in the US are still
34:58 going to have implications for the rest of the world.
35:01 So I'll just give you an example of both Google and Meta.
35:06 Both of them,
35:07 I know we're probably moving in completely parallel worlds in DC right now,
35:12 but there've been two really large antitrust cases for both these companies
35:17 in DC in the past year.
35:20 And
35:21 I would argue that those cases are going to have implications
35:26 worldwide.
35:28 And why is that?
35:29 Well,
35:29 if you think about current tech platforms,
35:32 there is this
35:35 I think the less talked about aspect of them
35:38 is that they tend to be worldwide,
35:41 and there's an awful lot of cross-subsidization involved.
35:45 Basically,
35:46 they are built around business models
35:48 that thrive on the fact that
35:50 rich people's eyeballs
35:52 are very valuable,
35:54 and if you put ads in front of those rich people's eyeballs,
35:56 you can make a lot of money.
35:58 And
36:00 The products though,
36:02 are worldwide,
36:02 even if all the revenues are being generated
36:06 in a few rich cities.
36:07 And so therefore,
36:08 when antitrust policy or competition policy comes in
36:11 and tries to alter those business models,
36:13 you know,
36:14 there might be some more worldwide implications.
36:18 OK,
36:19 so I've got 5 minutes,
36:20 and I should say a little bit more about my views about competition policy and how it
36:24 should change in the world of AI given that's what the title of the talk was,
36:28 or the paper was.
36:30 And again,
36:31 my message,
36:32 as I said,
36:33 to summarize,
36:34 is that economics doesn't change.
36:36 We're still going to have exactly the same digital economics concerns
36:40 that we've had for the last 30 years when thinking about
36:43 antitrust and technology.
36:45 And I tend to group these into 3 separate concerns,
36:50 which we might want to think about in the advent of AI.
36:54 The first,
36:55 uh,
36:55 obsession of competition economists
36:59 is a phenomenon called network effects.
37:01 And
37:03 just to remind everyone in the room,
37:04 a network effect occurs when products become
37:07 more valuable as more people use them.
37:10 Now,
37:11 network effects are very,
37:13 very popular term in competition policy.
37:16 People just love saying they're a network effects.
37:18 I don't know why,
37:19 but it's a little bit of obsession.
37:21 Um,
37:22 and so actually,
37:23 one of the reasons I wrote this paper was completely because I got irritated
37:27 by lawyers labeling anything that moved as a network effect.
37:32 Um,
37:33 what I'd argue is if you look right now though,
37:36 it's not clear that network effects are gonna operate in the same way
37:41 as we've seen with previous technologies.
37:43 If you think again back to Alvin's talk,
37:45 do you remember he was saying,
37:47 gosh,
37:47 you know,
37:47 one of the main
37:49 Implications of AI in computer science or in computing
37:53 could be that we end up with lots more little personalized systems.
37:57 In other words,
37:58 the opposite of a network effect.
38:00 And I think that's a far better example than I was gonna use,
38:03 which was if you think about the primary,
38:05 um,
38:06 use case
38:07 of AI right now in America,
38:10 which is helping high school students cheat on English essays.
38:14 There you might sort of say,
38:16 again,
38:16 if you're,
38:16 if you're cheating on an English essay,
38:18 you really like quite a bit of divergence in output,
38:20 right?
38:21 There's the opposite
38:22 of a network effect there.
38:24 So again,
38:24 so this primary move of competition policy network effects,
38:28 uh,
38:28 doesn't seem to be clearly present now.
38:30 We might see it as we get to the final stage,
38:34 uh,
38:34 sort of a layer two being built upon,
38:37 uh,
38:38 our current LLM models,
38:39 but it's not clear how it will evolve at the moment.
38:43 Now,
38:43 the next question will be one more of switching costs.
38:46 Uh,
38:46 Most digital antitrust cases are really about switching costs.
38:50 How easy is it to switch,
38:52 uh,
38:54 between search engines,
38:55 how easy is it to switch between,
38:58 uh,
38:58 entertainment sites.
39:00 This is the real question.
39:02 Um,
39:02 and I think there,
39:03 again,
39:04 we don't really have enough evidence.
39:06 And again,
39:07 because we don't really know what the business model's gonna be,
39:09 we don't really know where the switching costs are going to come from.
39:12 You know,
39:12 I think probably the right for experiment right now is to think,
39:17 well,
39:17 in the end,
39:19 The most common application of our model right now is you take it into a company,
39:25 uh,
39:25 you do some fine tuning.
39:27 Then there's a question about how easy it is to
39:29 switch the model once you've done that fine tuning.
39:32 But again,
39:32 I suspect that may be the wrong way of thinking about switching costs.
39:36 But do always be thinking about switching costs
39:38 when you're thinking about how AI
39:41 should be regulated in terms of competition.
39:44 Uh,
39:44 the last thing I'll just say about that
39:45 is what's wonderful about being a digital economist,
39:48 is that the moment you have a switching cost is
39:51 the moment that everyone is innovating around getting around it.
39:55 Um,
39:55 I always used to talk about iTunes when I
39:58 taught and say there's a brilliant switching cost.
40:00 No one's ever going to want to,
40:02 uh,
40:02 leave their iPod when they've got all their tunes on it.
40:05 And then,
40:06 of course,
40:06 came Spotify to just make me wrong.
40:09 And so,
40:10 again,
40:10 just remember the switching costs are always going to be transient.
40:14 Now in terms of data as an essential facility,
40:16 there's often this idea there may be barriers to entry to do data.
40:19 Um,
40:20 but
40:20 when you actually look
40:22 at the training data set it's already out here,
40:24 this is a screenshot of a website named Hugging Face,
40:27 which provides millions gig terabytes of data for free.
40:31 Uh,
40:31 you have to second guess it.
40:34 All right,
40:34 so just sort of what I say,
40:36 what,
40:36 what did I end up saying in this paper?
40:39 Or what's the big conclusion?
40:41 In the end,
40:41 it's really hard.
40:44 To be too definitive about
40:47 competition policy when we don't have a clear monetization strategy.
40:51 Um,
40:52 my concern right now is,
40:53 is our eagerness to regulate antitrust.
40:56 We may end up
40:58 regulating the,
41:00 um,
41:00 the analogy of the dial-up ISPs because we don't really
41:04 understand what's going to be built upon the essential innovation.
41:08 Anyway,
41:08 what we do know though,
41:10 is that in some sense,
41:11 the way we think about competition policy will have to change as a process.
41:16 Because as we just heard from Arvind,
41:17 a lot of how we might think about law being conducted
41:20 may change in the long run.
41:22 All right,
41:22 with that,
41:23 I will say thank you very much
41:25 and give you over to a development economist.
41:39 Thanks so much for,
41:41 for having me here.
41:42 So today I wanted to talk a bit about a couple of concrete examples of
41:47 projects
41:48 um that
41:49 are in on AI in low income settings to give us
41:52 a sense of what this technology is capable of and,
41:55 and put a little bit more concreteness around some of these,
41:58 these projections going into the future.
42:00 The main project I'll I'll present today is joined with three of my students,
42:04 Juno,
42:04 Divia,
42:05 and Dominic,
42:05 all at Columbia,
42:06 um,
42:07 and then Paul Atherton and,
42:08 and Oliver Gerard are at Fab Inc.
42:10 that created the chatbot that we'll be talking about.
42:13 But before we talk about that,
42:14 I wanted to,
42:15 to zoom back and,
42:16 and think about an earlier example,
42:19 uh,
42:19 and I,
42:20 I want to make the point that
42:21 we may think of artificial intelligence as some altogether new animal,
42:25 and I think Arvin has made this point that no,
42:27 it's actually kind of an extension of a normal
42:29 of technologies that we were familiar with.
42:32 And there are cases that where machine
42:34 learning has already impacted low income countries,
42:36 and I think it's useful to start from
42:38 that premise and really understand what happened there.
42:41 So,
42:41 so one project that I've worked on is
42:44 the need for financial services.
42:46 There are about 1.77 billion people around the
42:49 world who lack access to formal finance,
42:51 which means that not only do they not have bank accounts,
42:54 they also don't have credit histories.
42:56 And so as mobile phones came along,
42:59 people began to be able to use mobile money to transfer funds and to store funds.
43:04 And then it became an issue where you,
43:06 you could potentially imagine giving out loans via this platform,
43:10 but it's very hard to know whether someone is worth the
43:13 credit risk unless you have a credit score of some sort.
43:17 So I suggested that we might be able to come up with
43:19 credit scores just based on how a person uses their mobile phone
43:23 and predict their credit worthiness
43:25 and in,
43:26 and for example,
43:27 when you use a mobile phone.
43:29 It records patterns of how you place calls,
43:32 how you manage your balance over time,
43:34 etc.
43:36 and you can use machine learning,
43:37 even though if,
43:38 if you and I as humans look through this set of indicators,
43:41 there might be 5000,
43:42 6000 indicators,
43:43 we would have no idea how to
43:46 figure out and who
43:48 is a good credit risk just based on those indicators.
43:51 Machine learning can kind of connect the dots and
43:53 find patterns that are not obvious to us.
43:55 In the study that we did in a South American country,
43:58 we had a sample where some of the people also had credit
44:02 bureau histories and and all of them had mobile phone histories.
44:06 And we find that when you rely on the credit bureau score,
44:10 you're able to predict whether someone will repay a credit,
44:13 not perfectly,
44:15 but you have some ability to do that.
44:17 But as you go to people who have less
44:19 and less information in their credit bureau file,
44:22 the ability of a credit bureau score to predict repayment goes down.
44:27 And if you go to people who have no information in their credit bureau file,
44:31 you can't predict whether they'll repeal a loan at all.
44:34 In contrast,
44:35 we found with this digital method using the behavior in just how you use your phone,
44:40 we could come up with a credit score.
44:43 It wasn't perfect,
44:45 but it was something,
44:46 and it performed about as well as the standard credit bureau scores.
44:50 But the crucial thing is that we can also create those
44:52 scores for people who are completely missing from the financial system.
44:57 And so machine learning really allows you to extend loans
45:00 to people who were omitted from the financial system.
45:04 This has become a foundation for
45:07 a new industry providing
45:08 quick loans via mobile phone called digital credit.
45:11 It's available around the world.
45:13 And so you could imagine that we might have thought that the story would end there
45:17 and that this product is out now.
45:19 Now maybe credit is solved,
45:20 and I think kind of the answer,
45:21 and this comes to some of the points that Arvind had brought up,
45:23 is that there's a lot of difficulties after that point.
45:27 And so,
45:29 uh,
45:30 our team and,
45:30 and several other teams then try to
45:33 understand,
45:33 you know,
45:34 was this actually improving people's lives or could it
45:36 be that these were getting people into debt traps.
45:39 These loans are often short,
45:41 uh,
45:41 short in duration
45:42 and have high interest rates.
45:45 We found that
45:46 the welfare impacts seemed to be small.
45:48 On that,
45:48 they seem to be a little bit in the positive direction.
45:52 But there's also a lot of evidence that
45:54 firms who are extending loans were also taking advantage of consumers,
45:58 so this new kind of industry required
46:01 also some adaptations
46:02 to our infrastructure for monitoring harms and regulating credit.
46:07 So I wanted to put this up as an example just to suggest that,
46:10 you know,
46:10 the game doesn't end once AI enters
46:14 and tries to solve a problem that it raises,
46:16 raises a whole bunch of other questions that we need to address as societies.
46:21 Now,
46:22 that is one example of what we would
46:23 call predictive AI or predictive machine learning,
46:27 where you typically make decisions from digital data.
46:31 Credit scoring is one example.
46:33 A couple other examples that have seen some promise
46:35 in low income settings are one targeting digital aid,
46:39 where you can use the same type of data,
46:41 um,
46:41 that Josh Blumeno has been doing a lot of work on,
46:44 um,
46:44 and then also,
46:45 um,
46:46 weather and flood forecasting.
46:48 Now,
46:49 And you can hear you can use satellite images.
46:52 11 particular useful use of machine learning in this case is that
46:56 often we are working in settings that have
46:58 little access to data.
47:00 But we may increasingly have these digital data sources,
47:03 and machine learning allows us to extract signal from them
47:06 to make better decisions.
47:09 But the reason we're here is not just to talk about predictive AI.
47:12 What people are excited about is this new kind of era of AI where
47:16 these systems are able to do more,
47:18 and
47:19 you've probably heard this called generative AI,
47:22 and this is a technology underlying chatbots
47:25 and that can create images and video and all of these kind of
47:28 more recent things that you've seen in the last couple of years.
47:31 So I'm going to now think about
47:33 one example of what generative AI can do in this
47:37 project that we're working on with teachers in Sierra Leone.
47:40 But I want to first think about
47:42 connectivity and information in sub-Saharan Africa.
47:46 So,
47:47 right now,
47:47 85% of people in sub-Saharan Africa actually are within coverage
47:51 of signal that they could use to access the internet.
47:55 So,
47:55 Africa is to some extent,
47:57 quite connected,
47:58 or has the capability of being quite connected.
48:01 Despite that,
48:02 only 37% of people
48:05 are using the internet.
48:07 And if you look at how people use the internet,
48:11 even among this group of users,
48:12 what you'll find is a fairly striking pattern.
48:15 People use social networking,
48:17 people use instant messaging like WhatsApp,
48:20 but people are much less likely to use what maybe you and I would think of as,
48:24 as the bulk of the internet,
48:26 which is using the web and web search.
48:29 Now we're going to be thinking about this problem
48:32 or this gap
48:33 more specifically in the population of teachers in Sierra Leone.
48:38 Teachers are knowledge workers who you might think would be using the web a lot,
48:41 but a lot,
48:42 many of the teachers in our sample live in
48:44 remote areas where internet connectivity is quite poor.
48:48 And so in our sample of teachers,
48:50 95%
48:52 of them use WhatsApp every day.
48:54 We find that only 3% use web search each day to look for classroom materials.
49:01 And so I'm going to,
49:03 our partner organization introduced a chatbot over
49:06 WhatsApp that allows teachers to access AI
49:09 and what they find is that
49:11 teachers end up using this on balance a bit more
49:14 often than they use search.
49:17 So our paper is really trying to understand
49:19 what is going on here,
49:20 what is AI doing that could be useful here.
49:24 Now,
49:25 one,
49:26 there's a couple of reasons why you might think
49:28 the internet is not doing a good job here.
49:30 One is content.
49:32 If you're in the US and you search for something,
49:36 what you find
49:37 is content from the US.
49:40 If you're anywhere else in the world,
49:43 when you search
49:44 for something on the internet,
49:45 a lot of what you find is not local,
49:47 a lot of it is foreign.
49:49 So content is lacking in most countries around the world.
49:52 And the second issue is bandwidth.
49:55 If you ask
49:56 Africans in a survey,
49:59 Are there,
49:59 are there limits to your use of the internet,
50:01 most will respond yes,
50:03 and most will report that the cost of data is the major reason
50:06 why they don't use the internet as much as they otherwise would.
50:09 11% also report that the network is a struggle.
50:13 Now we're gonna think about this in Sierra Leone,
50:15 which has a GDP per capita of $509
50:18 so a fairly low income country,
50:20 and we're gonna be working with teachers
50:23 who are working in schools that are,
50:24 are fairly overloaded,
50:26 um,
50:26 and only about 6% of uh primary schools have electricity
50:30 and 1% have the internet.
50:33 So when we're thinking about the impacts of AI here,
50:35 this isn't a setting where you're going to drop some magic tablet,
50:38 and it's going to magically
50:39 teach all of the kids that may come in coming years,
50:42 uh,
50:43 but,
50:43 but that's not immediately,
50:44 uh,
50:45 it seems,
50:45 it seems to be on the table.
50:47 So,
50:48 in this context,
50:49 the,
50:49 the partner organization created this chatbot called the Teacher AI
50:52 and it basically links WhatsApp
50:55 to
50:56 GPT
50:57 with a prompt that says,
50:58 please be,
50:59 be helpful and,
51:00 and make useful suggestions for teachers.
51:03 We find that teachers ask questions like write
51:04 a lesson plan to teach two-digit addition.
51:07 How can I support students who fall behind,
51:10 describe immunization and give examples for 1/5 grade class in Sierra Leone.
51:15 Uh,
51:15 let me just give you one example here of how the chatbot responds.
51:18 So for the query activities to teach shapes,
51:22 the,
51:22 the chatbot came back and it,
51:23 it knows the teacher is teaching grade 6,
51:26 and it gave some examples of,
51:27 of creating an art collage,
51:29 creating bingo,
51:30 just activities that engage the students,
51:32 um,
51:32 which,
51:33 which seem pretty kind of straightforward things that
51:35 you might have read in a textbook,
51:36 but it's kind of tailored to this particular query.
51:40 Now,
51:41 the first question is,
51:42 is how do teachers use this chatbot?
51:44 So we analyzed the queries they submitted.
51:47 So this is a sample of 469 teachers
51:50 who asked
51:50 about 40,000 queries over the span of a year and a half,
51:54 and we used AI itself to classify these queries into categories
51:58 to understand what is it that teachers are using this for.
52:02 The biggest category is concepts and facts,
52:05 so this is often things that you might find on Wikipedia.
52:10 Though teachers also use this system for lesson planning assessment,
52:14 by creating new,
52:15 new exams
52:16 and supporting their writing and also for,
52:18 for professional guidance.
52:21 Uh,
52:22 we also asked them,
52:23 when you use AI rather than search,
52:25 what are the reasons,
52:27 and the top reason was that they said that AI gave them a,
52:30 gave them a useful answer right away and that it was concise.
52:34 When we asked the reverse,
52:36 not as many responded,
52:38 uh,
52:38 but the modal response was that the web has unique content
52:42 that AI does not have.
52:44 Now
52:45 teachers seem to rate AI as quite trustworthy,
52:49 but we all know
52:51 that AI can make mistakes or hallucinations,
52:53 and so we're going to come back to this
52:55 later to understand why are teachers trusting this system.
52:59 And then to flag,
53:00 a few of these teachers also mentioned in the survey that they
53:03 stopped using the web once they got access to this AI tool.
53:08 So to understand why they seem to be using this tool,
53:11 we took the queries that they submitted
53:14 and we classified these into threads
53:16 and then took the top,
53:18 the first query from each thread,
53:19 which is the one that would have had the most context.
53:23 Um,
53:24 we'd know how the teacher AI would have responded,
53:27 but we want to know,
53:28 had that teacher taken that same question to,
53:31 to the local search engine,
53:33 what would they have seen?
53:34 The most common local search engine in this case is google.com.sl
53:39 and so we submit that same query to Google and scrape the top five search results.
53:45 And then we're going to compare what is the information that the teacher would
53:48 have seen had they submitted this to a search engine on the web like
53:52 or versus what they got from the AI response.
53:56 Now,
53:57 the first comparison to to think about is that we came up with a ranking
54:02 of how similar the AI response is to existing search engine content.
54:08 And we came up with a score,
54:10 which we call the contained score,
54:11 which ranges from 1,
54:13 which means that the content that the AI generated is
54:16 very new and it's different from the search responses,
54:19 or 10,
54:20 which would suggest that the content in the AI response is completely contained
54:25 in
54:26 one of the search results.
54:28 We find that the majority of the queries that teachers submit
54:31 come up with responses that overlap with existing content.
54:34 It's already available on the web,
54:36 they just use the chatbot.
54:37 There is this small tale where the teachers are using this tool to
54:41 generate new content like a custom lesson plan or a new story,
54:45 but it's a minority in this population,
54:47 so
54:48 they're using AI mostly to retrieve existing content.
54:52 The second thing is that only 2%
54:55 of the results that come back from Google are actually from Sierra Leone.
55:02 And
55:03 most of the results come from the United States
55:06 and very few come from
55:08 similar countries to Sierra Leone at all.
55:10 So much comes from the from the United Kingdom or Australia,
55:13 for example.
55:15 And this means that not only is it foreign content,
55:17 but also
55:18 it is tailored for foreign users.
55:21 And if we're,
55:22 if we're accessing the internet on our
55:24 desktops or laptops with fast internet connections,
55:27 that content is tailored to us.
55:30 And so when you look at the data required to transfer a search result,
55:34 the average search result
55:36 that a teacher would have seen in our sample is about 2.5 megabytes.
55:41 The that includes both the text,
55:43 syntax,
55:44 fonts,
55:44 data,
55:45 a whole bunch of content
55:47 that's easy for us to download.
55:49 Um,
55:50 the AI response was under 1 kilobyte.
55:53 And so we find that the AI system is over 3000 times more data efficient
56:00 than using the web.
56:02 And
56:04 while
56:04 that data is easy and cheap for us sitting in offices,
56:08 these teachers are paying per gigabyte of data,
56:11 so this entails a cost.
56:13 And so,
56:13 if you look back in time,
56:16 at the beginning,
56:17 kind of prior to this study,
56:18 if you were to have submit 1000
56:21 or download 1000 search results from this sample,
56:25 uh,
56:25 you would have paid about $2.50.
56:28 In bandwidth fees
56:30 and we've,
56:30 we know that there's a discussion that AI is kind of expensive to use.
56:34 It's true that AI requires not only the bandwidth here,
56:37 but it also requires a server to compute the response.
56:41 And so,
56:41 if you went back in time prior to the release of chat GPT,
56:45 so earlier in 2022,
56:48 um,
56:49 an AI chatbot in this setting would have cost about $30.
56:52 It would have been about 13 times more expensive for these 1000 queries.
56:58 Over time,
56:59 the compute cost for AI has declined declined.
57:02 So,
57:02 during most of the study,
57:03 they cost about the same,
57:05 but then with recent drops,
57:06 the AI chatbot has become an order of magnitude cheaper.
57:09 So,
57:10 we find that it is actually 87% cheaper
57:13 to use the AI system
57:14 in the setting to get a response
57:17 than to access a search result.
57:19 And so,
57:19 this seems to suggest that there's a lot of potential
57:22 to,
57:22 to do better by,
57:24 by
57:24 communities that have low bandwidth connections.
57:27 We also
57:29 that that would be
57:30 Fine if it were cheaper,
57:31 but we also want to make sure that it is not providing low quality information.
57:36 So we also ask teachers to rate the system.
57:38 We find that
57:39 in this,
57:40 these ratings,
57:40 teachers rate this the AI responses as more relevant and helpful.
57:45 And then the next that wasn't surprising to us,
57:48 but we were surprised that teachers also seemed
57:50 to rate the AI responses as more correct
57:53 and less likely to include inaccuracies.
57:56 Now we all know that these systems make mistakes.
57:58 In this sample that we tested,
58:00 we found hallucinations in about 3%
58:03 of the responses.
58:05 So there definitely are mistakes
58:07 in how this AI system interacts.
58:10 But it seems like the teachers are using it for fairly
58:13 basic questions for which it does a pretty good job.
58:16 And also
58:17 what illuminate this exercise illuminates is that when I sit
58:20 at my desk and I use a Google search,
58:22 I don't expect one particular search result to be true.
58:26 I expect to have to vet multiple and triangulate between them to come to truth.
58:31 And when we asked the survey respondents here to do an
58:35 exercise just looking at whether a search engine result was true,
58:38 they seemed to rate it less,
58:39 less strongly.
58:40 So
58:41 this,
58:42 this paper we really think of as showing that AI here can do something.
58:47 Somewhat simple,
58:48 it can reformat content to work better on small screens and snow collections,
58:53 and it kind of reveals that the information services that we're
58:56 providing to much of the lower income parts of the world
58:59 are not meeting their needs,
59:02 and there may be ways to design better systems.
59:05 Uh,
59:06 this is a pretty early system.
59:08 It's pretty simple,
59:09 but we think it,
59:10 it suggests that there may be a lot of different directions we could go.
59:13 A few caveats,
59:15 one is that
59:16 the less educated teachers seem to use the system less,
59:19 so we do think you need some education to be able to be productive here.
59:23 This is English speaking,
59:24 these models perform a lot better in English,
59:26 um,
59:26 and it seemed like these teachers needed some training.
59:29 I also wanted to plug that the partner has an
59:31 AI for education website that has more resources that they're doing useful things,
59:35 but
59:35 um,
59:37 Anyway,
59:37 I'm very excited to think about this question about whether AI,
59:41 what,
59:42 what opportunities there are to actually embed AI in low income settings.
59:45 I think there's a lot of great things we could do,
59:46 but looking forward to your questions.
59:54 Thank you Irvin,
59:55 Katherine,
59:55 and Dan for your excellent presentations.
59:58 Uh,
59:59 Hancheng,
59:59 maybe I could request you to
1:00:02 Kick off the discussion and maybe
1:00:04 tell us your key takeaways from what we've just heard.
1:00:07 Thank you.
1:00:07 Sure.
1:00:08 Thank you so much to the three presenters,
1:00:10 uh,
1:00:10 for,
1:00:11 for really,
1:00:12 uh,
1:00:12 such in-depth
1:00:13 exploration and in some instances pushback to a lot of
1:00:16 the narrative that we've been hearing coming out of Silicon Valley
1:00:19 that says,
1:00:19 you know,
1:00:20 this,
1:00:20 this technology is gonna
1:00:22 transform how
1:00:23 the economy and society behaves,
1:00:26 you know,
1:00:26 tonight,
1:00:27 tomorrow,
1:00:28 um,
1:00:28 so,
1:00:29 so I think.
1:00:30 Um,
1:00:31 you know,
1:00:32 firstly,
1:00:32 pushing back on this spirit of inevitability,
1:00:34 I think is,
1:00:35 is incredibly important.
1:00:36 I think it,
1:00:37 I think this,
1:00:38 this,
1:00:38 you know,
1:00:38 there's a lot of marketing hype that hides
1:00:41 kind of very sloppy thinking.
1:00:42 I think that
1:00:43 the authors today challenge us to think about what are the specific pathways,
1:00:47 uh,
1:00:47 that determine whether these grand outcomes could or could not happen.
1:00:51 Um,
1:00:52 will we see high adoption because the technology is inexpensive
1:00:56 and valuable to teachers,
1:00:57 as Dan has pointed out,
1:00:59 or will we see lower and slower adoption
1:01:01 because as Arvin has pointed out,
1:01:03 um,
1:01:03 you know,
1:01:04 these capable technologies get intermediated through
1:01:08 business logic,
1:01:09 workflows in,
1:01:10 in a firm that's trying to adopt this technology.
1:01:13 Um,
1:01:14 and so
1:01:15 it's very helpful to hear these very specific examples from both Arvin and,
1:01:18 and Daniel.
1:01:20 Um,
1:01:21 I think
1:01:22 the,
1:01:23 you know,
1:01:24 for,
1:01:24 for many economists here,
1:01:25 that the concept of pathways and mechanisms for impact is not new,
1:01:30 right?
1:01:31 For,
1:01:31 for the last 20 years,
1:01:33 many of you in this room have probably run experiments and studies to try and
1:01:36 tease out what exactly it is that drives the development outcomes we care about.
1:01:40 So
1:01:42 we've learned that,
1:01:43 um,
1:01:44 you know,
1:01:44 necessarily,
1:01:45 you know,
1:01:45 sending kids textbooks is not necessarily what
1:01:48 will drive learning outcomes,
1:01:49 but that teaching at the right level,
1:01:51 targeting instruction to a child's learning level is particularly effective.
1:01:55 I think the question for us now as,
1:01:57 as,
1:01:58 as we Think about whether we're going to hit our
1:02:00 development outcomes and whether AI can help us with that
1:02:03 is can AI be applied specifically
1:02:07 to those pathways that we already know,
1:02:09 to those mechanisms that we have
1:02:11 already demonstrated to be particularly effective
1:02:14 in low and middle income settings.
1:02:17 Um,
1:02:17 and so,
1:02:18 so,
1:02:19 you know,
1:02:19 at the end of the day,
1:02:21 If I think,
1:02:22 and,
1:02:22 and,
1:02:22 and I think I,
1:02:23 Irvin has talked about this,
1:02:24 you know,
1:02:24 I,
1:02:24 I think it's a bit of a mirage to look at
1:02:27 capabilities
1:02:29 uh to predict outcomes,
1:02:31 uh,
1:02:31 but rather we should look at specific pathways and mechanisms.
1:02:34 It doesn't matter if
1:02:36 our AI is,
1:02:38 um,
1:02:38 is outperforming,
1:02:40 uh,
1:02:40 every math benchmark out there,
1:02:42 it doesn't necessarily translate into the user becoming more numerate.
1:02:46 So,
1:02:47 so I think that that's a fantastic contribution.
1:02:49 I think,
1:02:49 I think from this group.
1:02:51 Um I do want to
1:02:53 quickly
1:02:55 point out
1:02:56 um that,
1:02:58 that,
1:02:58 um,
1:02:59 you know,
1:03:02 Dan has presented some,
1:03:03 some,
1:03:03 some empirical evidence,
1:03:06 uh,
1:03:06 but really,
1:03:07 um,
1:03:08 uh,
1:03:08 Catherine and Arvin have made very bold forward-looking predictions,
1:03:12 uh,
1:03:12 into,
1:03:13 into the future.
1:03:14 Um,
1:03:15 and,
1:03:16 and I,
1:03:17 I read all of their papers in detail.
1:03:19 And Catherine has a fantastic,
1:03:22 uh,
1:03:22 uh,
1:03:23 uh,
1:03:23 last line
1:03:25 in her,
1:03:26 in her paper,
1:03:28 where she says,
1:03:28 you know,
1:03:29 and,
1:03:29 and sorry if I'm butchering the takeaway of your paper a little bit,
1:03:32 but,
1:03:32 but you say,
1:03:32 you know,
1:03:32 I think competition policy doesn't need to change
1:03:35 that much that that currently maybe our,
1:03:37 our regulatory frameworks might already be quite suitable for the age of AI
1:03:40 and you say,
1:03:41 but if you're having a real knee jerk reaction to this prediction,
1:03:45 Uh,
1:03:45 right now,
1:03:47 um,
1:03:47 uh,
1:03:47 I challenge you to think about all the ways,
1:03:50 uh,
1:03:50 I could be wrong,
1:03:51 and,
1:03:52 and the interesting question is,
1:03:54 is not so much,
1:03:55 am I right or am I wrong,
1:03:56 but,
1:03:57 but,
1:03:57 but,
1:03:58 but
1:03:59 why am I wrong?
1:04:00 And so,
1:04:01 uh,
1:04:02 perhaps it's a,
1:04:02 it's a spicy question that I can,
1:04:04 I can tee up for,
1:04:05 for,
1:04:05 for this panel.
1:04:06 I,
1:04:06 I love,
1:04:06 you know,
1:04:06 maybe in the comments,
1:04:08 uh,
1:04:09 uh,
1:04:09 uh,
1:04:09 through the Q&A
1:04:11 to,
1:04:11 to try and play the counterargument a little bit and,
1:04:13 and come,
1:04:14 you know,
1:04:14 I,
1:04:14 I'd love to hear,
1:04:15 uh,
1:04:16 uh,
1:04:17 what are some credible
1:04:19 Uh,
1:04:19 arguments for,
1:04:20 for,
1:04:21 for,
1:04:21 for why your positions,
1:04:22 uh,
1:04:23 today could be wrong?
1:04:24 Could,
1:04:24 could superintelligence actually transform us overnight?
1:04:27 What,
1:04:28 what would a credible counterpart
1:04:30 say to that?
1:04:32 Um,
1:04:32 but,
1:04:32 but let,
1:04:33 let,
1:04:33 let me turn it over,
1:04:34 uh,
1:04:34 to Q&A.
1:04:36 Uh,
1:04:36 thanks,
1:04:37 Hansheng.
1:04:37 So maybe uh I'll add a question from my side as well,
1:04:40 and then maybe we can get some responses.
1:04:42 So,
1:04:43 I just want to pull a word out of Dan's presentation of the title leapfrogging,
1:04:48 right?
1:04:48 So,
1:04:49 I think,
1:04:50 uh,
1:04:51 we've heard,
1:04:51 I think time and again that it's difficult to leapfrog
1:04:55 general purpose technologies,
1:04:57 uh,
1:04:57 although we have seen instances of leapfrogging
1:05:00 sector-specific technologies,
1:05:02 I think the landline and mobile phones was the
1:05:04 Most commonly cited examples.
1:05:06 So,
1:05:06 I think that what you presented tells us something a little bit different in that,
1:05:10 you know,
1:05:11 people are actually not using web searches,
1:05:13 but are going straight to AI searches.
1:05:15 So that's,
1:05:16 so,
1:05:16 so that's,
1:05:16 I think an interesting
1:05:18 result that we see.
1:05:20 But I think in general,
1:05:20 just linking it to
1:05:22 all the other presentations,
1:05:24 so it's the promise of leapfrogging.
1:05:27 Or is it
1:05:29 That you can get more from spending
1:05:31 or making similar investments,
1:05:33 right?
1:05:33 So if you are making investments around
1:05:36 digital infrastructure or around smartphone access,
1:05:40 it's just you can get a lot more
1:05:42 from that than what you were getting before.
1:05:44 So,
1:05:44 is that an aspect of leapfrogging um that's important,
1:05:48 um,
1:05:49 or is the claim that you're making a little bit bolder,
1:05:51 because it sort of a little bit sits
1:05:54 I think opposite to what Irvin said,
1:05:56 like,
1:05:56 this is just a normal technology,
1:05:58 so
1:05:58 are we gonna see something that's
1:06:01 You know,
1:06:02 speeds up the diffusion,
1:06:03 even if it's a normal technology.
1:06:05 So maybe if you could just take some of
1:06:07 the questions from Han and,
1:06:08 and this one and
1:06:10 we can keep the discussion going.
1:06:13 Great.
1:06:13 Uh,
1:06:14 yeah,
1:06:14 so
1:06:16 I
1:06:17 I,
1:06:17 I'll say something
1:06:19 maybe a little bit spicy,
1:06:20 but
1:06:21 every technology seems normal in retrospect,
1:06:24 uh,
1:06:25 and I think there's been a lot of technologies that,
1:06:27 that have had major impacts that
1:06:31 were quite destabilizing at the time.
1:06:33 And so I think
1:06:35 I think
1:06:37 that will be right.
1:06:38 Uh,
1:06:39 I don't think that diminishes what it
1:06:42 What impact it could,
1:06:43 could have,
1:06:44 uh,
1:06:44 on,
1:06:45 on the question of leapfrogging,
1:06:46 I mean,
1:06:46 I think I
1:06:48 Maybe it's helpful to split this into two buckets,
1:06:50 so
1:06:52 These are
1:06:53 our technologies that are seeing a lot of investment from Silicon Valley
1:06:56 and
1:06:57 from
1:06:59 Wealthy countries
1:07:00 and
1:07:01 that will improve its usefulness for many
1:07:04 of the tasks that happen in wealthy countries
1:07:07 and many of those will spill over to the equivalent
1:07:09 tasks in low income settings and so you think about office
1:07:14 jobs and and so on,
1:07:15 knowledge work
1:07:17 is going to be kind of exposed around the world.
1:07:20 And I think
1:07:22 where a lot of my work is thinking about is are there particular wins
1:07:26 in low income settings that might otherwise be overlooked.
1:07:30 And are those going to happen automatically?
1:07:33 So that
1:07:35 like mobile phones for example are we're just
1:07:37 such a great technology and the private market
1:07:39 really
1:07:41 allowed that to spread
1:07:42 around the world and reach
1:07:44 almost everyone with signal and and communication technology
1:07:48 or is this going to be a type of technology
1:07:50 that is going to require some active policy to,
1:07:52 to really push
1:07:54 and so
1:07:54 I think
1:07:55 maybe when I focus on leapfrogging,
1:07:57 I,
1:07:57 I'm not saying that overall this is going to always cause leapfrogging,
1:08:01 but.
1:08:01 That
1:08:02 I'm kind of drawn to these places where,
1:08:05 where there's something interesting happening in low income communities and,
1:08:08 and debating about whether they need some policy support.
1:08:13 Thanks,
1:08:14 um,
1:08:14 Arvind,
1:08:15 uh.
1:08:17 I'm not sure if you had a,
1:08:19 yeah,
1:08:19 sure,
1:08:20 so this is all,
1:08:22 you know,
1:08:22 as a computer scientist,
1:08:23 so I was listening to
1:08:25 Dan's presentation and I had two
1:08:27 simultaneous reactions one admiration for the research
1:08:31 and the other is horror at what the web has become.
1:08:34 So if you had gone back 30 years and told people that
1:08:37 this is what the web's going to look like in 2025,
1:08:40 the pioneers of the web,
1:08:41 right,
1:08:42 that
1:08:42 an average query for
1:08:45 an educational query,
1:08:46 right,
1:08:46 not even entertainment or video or anything.
1:08:49 It's going to be 2 megabytes
1:08:52 they would not know how to deal with that information.
1:08:54 That was beyond the worst case scenario of what they would have considered,
1:08:57 right?
1:08:58 Like all of the advantages you were talking about with AI today,
1:09:00 those were
1:09:02 a lot of the reasons that people were excited about the web originally
1:09:05 because you can,
1:09:06 you know,
1:09:06 it's this tremendous educational tool.
1:09:08 You can compress
1:09:09 a lot of information into just a few bytes of text,
1:09:11 and that it's going to have these tremendous effects,
1:09:13 especially in a in a development context.
1:09:16 And I think the wounds are entirely self-inflicted technologically,
1:09:20 of course,
1:09:20 things have not gotten worse.
1:09:21 You can still have,
1:09:22 you know,
1:09:23 very,
1:09:23 very efficient web pages.
1:09:25 We have chosen not to because,
1:09:28 uh,
1:09:28 search engines need to monetize and the people
1:09:31 creating those web pages want to monetize,
1:09:34 which goes back to Katherine's talk,
1:09:35 which is that we're in the pre-business model stage,
1:09:38 I think,
1:09:38 of AI.
1:09:39 So I think it's worth thinking about 10 years from
1:09:41 now what are those business models going to look like.
1:09:44 And could they make
1:09:46 AI look a lot like what the web looks like today?
1:09:49 Um,
1:09:50 I don't necessarily want to,
1:09:51 you know,
1:09:52 um,
1:09:52 constantly be the pessimist,
1:09:54 but it is
1:09:55 something I think about a lot when I look at,
1:09:57 um,
1:09:58 the claims around and shitification,
1:10:00 for instance,
1:10:00 if folks have heard that term,
1:10:02 the way that,
1:10:02 uh,
1:10:03 social media platforms and other
1:10:05 two-sided marketplaces have the,
1:10:07 uh,
1:10:07 rent-seeking effects of degrading product quality in the interest
1:10:10 of making things more addictive and making more revenue.
1:10:15 Again,
1:10:15 while completely agreeing with Katherine that it's very early days,
1:10:17 I think it's worth thinking
1:10:20 ahead about what are some of those potential
1:10:22 negative effects that are coming to these new platforms
1:10:25 and what perhaps can be done from
1:10:27 a policy perspective without being too prescriptive,
1:10:29 without
1:10:30 jumping the gun on competition policy
1:10:32 to
1:10:33 try to,
1:10:34 you know,
1:10:35 decrease the probability of some of those future negative consequences.
1:10:40 All right.
1:10:42 So I,
1:10:42 I love the question which was,
1:10:44 tell us why you're wrong.
1:10:46 Uh,
1:10:46 such a wonderful question.
1:10:48 So I,
1:10:48 I've been thinking about ways I'm wrong.
1:10:51 Um,
1:10:53 The first thing I would like to highlight about why economists are often wrong.
1:10:59 Is this analogy.
1:11:01 Our problem is that often when we think about labor displacement,
1:11:06 we have in mind something analogous to thinking
1:11:11 that what
1:11:12 Uber is doing is or ride sharing is doing is replacing taxi dispatches.
1:11:21 And that in some sense is completely
1:11:25 underestimating the real change.
1:11:29 And if you think what ride-sharing has done is it's completely transformed
1:11:33 the process surrounding
1:11:36 how you get into something akin to a taxi.
1:11:39 And that's what makes us wrong and potentially
1:11:43 maybe that what could be what makes Arvind and I wrong that we just
1:11:47 don't anticipate enough how much process change there will be.
1:11:52 Um,
1:11:53 the second reason I think we could be wrong,
1:11:56 or I don't,
1:11:56 I don't know,
1:11:56 I'm meant to be only talking about myself and now I'm making everyone else wrong.
1:12:00 But
1:12:01 I think,
1:12:02 you know,
1:12:02 in a talk like this,
1:12:04 it is very easy to fall into the trap of calling
1:12:09 everything artificial intelligence.
1:12:11 Um,
1:12:11 when often what we're really talking about is something more akin to
1:12:15 machine learning,
1:12:16 predictive analytics.
1:12:18 And you know,
1:12:19 this reminds me of a,
1:12:22 and again,
1:12:22 that could lead us to
1:12:24 make mistakes.
1:12:25 I'm always reminded of one of my more famous papers in the realms of AI.
1:12:29 I was looking at how machine learning did,
1:12:32 um,
1:12:33 uh,
1:12:34 this,
1:12:34 these sort of machine learning algorithms run
1:12:36 by large credit bureaus and similar data,
1:12:38 data brokers.
1:12:40 And what I showed there was that machine learning can generally,
1:12:44 if given web search or web browsing data,
1:12:48 classify gender right about 50% of the time.
1:12:52 And if you think about it,
1:12:53 that's quite funny.
1:12:55 So given that machine learning
1:12:57 can often get things wrong,
1:13:00 predictive analytics as of now can often get things wrong.
1:13:03 We're maybe being very bold by sort of overselling it in terms of this AI
1:13:09 label,
1:13:10 which I think we will,
1:13:11 we've all been using.
1:13:12 So I think those are the ways I'm wrong,
1:13:14 right?
1:13:15 Misusing language.
1:13:17 Under predicting process change.
1:13:19 Thanks,
1:13:19 thanks,
1:13:19 Catherine.
1:13:20 Maybe I,
1:13:21 I have one
1:13:23 more question before we open up uh the,
1:13:25 the,
1:13:26 you know,
1:13:26 the floor to questions.
1:13:27 Um,
1:13:28 just,
1:13:28 I wanted to pick up one,
1:13:30 I think one of the pictures that Arvind you had early on in your presentation just
1:13:34 on on sort of invention,
1:13:36 innovation,
1:13:37 and,
1:13:37 uh,
1:13:38 you know,
1:13:38 wider adoption or diffusion.
1:13:40 So,
1:13:42 I think just because this conference is about development,
1:13:44 I just wanted to see if how we
1:13:46 can place or contextualize developing countries along this,
1:13:50 this process,
1:13:52 um,
1:13:52 and perhaps,
1:13:53 you know,
1:13:53 there is room for them to be in all parts of the process,
1:13:57 but more so,
1:13:58 um,
1:13:59 in,
1:13:59 in,
1:14:00 you know,
1:14:00 one particular part of the process.
1:14:02 And,
1:14:03 you know,
1:14:03 maybe,
1:14:04 uh Catherine,
1:14:04 like,
1:14:05 also extending sort of your
1:14:07 talk about competition and value,
1:14:09 I think the,
1:14:10 the sense that people have is that all the value is in the invention
1:14:15 stage that Irvin put up or perhaps a little bit in the innovation,
1:14:18 but not so much at,
1:14:19 at the bottom end.
1:14:20 So,
1:14:20 how should we think about that?
1:14:22 Yeah,
1:14:23 I'll maybe say a couple words and then would love to hear from,
1:14:25 uh,
1:14:26 folks who have more expertise in developmental economics.
1:14:29 I mean,
1:14:29 I'll,
1:14:30 I'll start by saying that developing countries focusing industrial policy on
1:14:34 data centers,
1:14:34 for instance,
1:14:35 seems bonkers to me.
1:14:37 Uh,
1:14:38 like I'm,
1:14:39 I'm very convinced by Jeff Ding's book Technology and the
1:14:41 Rise of Great Powers looking historically at general purpose technologies.
1:14:45 It
1:14:46 ended up not mattering that much
1:14:49 in which country the innovation happened.
1:14:51 What was much more
1:14:53 useful to explain the variance and outcomes between countries was how good
1:14:57 those countries were at diffusing those technologies
1:15:00 throughout various sectors of the economy,
1:15:01 and policy can play a big role there,
1:15:05 you know,
1:15:05 just looking at regulation as one form of policy can both inhibit
1:15:09 diffusion and spur diffusion.
1:15:11 So my favorite example of this really trivial but I think really powerful,
1:15:16 is
1:15:17 in the early days of the web e-commerce was just such a strange notion to people
1:15:22 and what actually helped that along in the US,
1:15:25 but I think you know it applies anywhere,
1:15:27 is the e-Sign Act that said that companies are forced to accept.
1:15:32 An online signature in any contract in which
1:15:35 they would have accepted a physical signature,
1:15:38 and this one thing really changed the game because it allowed people
1:15:41 to digitize their processes without worrying that
1:15:44 suddenly they're going to run into a roadblock
1:15:47 where a party they're trying to contract with
1:15:49 will not accept online ways of doing things,
1:15:52 right?
1:15:52 So this was a huge it's a simple example of how,
1:15:56 you know,
1:15:56 policy can play a big role in diffusion,
1:15:58 right?
1:15:58 And
1:15:59 And that's where I think the impacts are really going
1:16:02 to be felt in terms of where countries can,
1:16:04 you know,
1:16:05 leapfrog,
1:16:05 etc.
1:16:06 as opposed to
1:16:07 putting up a data center.
1:16:09 It's not clear what they're achieving by having the
1:16:11 model training done in their country when it's owned by
1:16:15 an American company anyway and then can be used by anyone in the world,
1:16:18 so something does not compute as far as I'm concerned.
1:16:23 Catherine,
1:16:24 Dan.
1:16:25 Oh,
1:16:25 well,
1:16:25 I'll just say,
1:16:26 you know,
1:16:26 you may,
1:16:27 again,
1:16:27 with,
1:16:28 you may be thinking,
1:16:29 so sitting here looking at me
1:16:31 going,
1:16:31 of course,
1:16:32 it's a business school professor
1:16:34 going on and on and on about monetization models,
1:16:37 so predictable,
1:16:38 completely irrelevant for my world.
1:16:40 Now,
1:16:41 I would like you to have this Ford experiment about why monetization models
1:16:46 are important.
1:16:48 And if we were to go back in time.
1:16:51 Uh,
1:16:52 to 1995,
1:16:54 and I was to tell you that monetization models would develop,
1:16:59 which meant that mapping would be free for the rest of the world.
1:17:03 You would have fought me completely insane.
1:17:06 And especially if I told you that that was going to be
1:17:09 advertising.
1:17:10 Again,
1:17:11 sounds nuts.
1:17:12 And so I would argue that in some sense,
1:17:15 where we land with monetization models is going to be important
1:17:20 really just in terms of access because
1:17:22 it's going to very much shape access internationally.
1:17:28 Yeah,
1:17:29 so I,
1:17:29 I'd agree that a lot of the work here is figuring out what these,
1:17:34 what these systems are capable of and how that integrates
1:17:38 into local economies.
1:17:39 What does that mean for business processes like your specific business process?
1:17:43 There,
1:17:44 there is a challenge here in that.
1:17:47 Some of these models will already slot in well into some tasks.
1:17:52 Some of them will currently fail
1:17:55 for reasons that are locally specific.
1:17:57 It may be bad in the local language.
1:17:59 It may not know the local conditions.
1:18:01 And so I think there's,
1:18:03 there's some
1:18:05 danger here.
1:18:06 I think we need a combination of kind of exploration by entrepreneurs and etc.
1:18:10 and really pushing it to understand.
1:18:12 How this technology can be useful in local conditions,
1:18:15 but then
1:18:15 we will at some point need more coordinated investment to make sure that if,
1:18:20 if there's latent capabilities where these models could be useful
1:18:24 for many people around the world,
1:18:26 but they would require some some public intervention,
1:18:28 how do we design systems for that and I think
1:18:31 are the are the business models going to support that
1:18:34 in wealthy countries where this will happen automatically or do we need
1:18:38 other institutions to step in is a huge question.
1:18:42 Yeah,
1:18:42 if I,
1:18:43 if I may just build on,
1:18:44 on,
1:18:44 on that,
1:18:45 and,
1:18:45 and to Arvin's question about,
1:18:46 you know,
1:18:47 why are lower income countries just,
1:18:50 you know,
1:18:50 so obsessed about building data centers and having their own,
1:18:53 their own
1:18:54 model development capabilities,
1:18:55 I think,
1:18:56 you know,
1:18:56 the,
1:18:57 the,
1:18:58 the pushback to,
1:18:58 to that,
1:18:59 to that point I've been hearing is that.
1:19:01 Uh,
1:19:02 competition among the major AI labs in high-income countries are not going to
1:19:06 deliver on the services that ultimately we need in our local context.
1:19:09 I think some of the,
1:19:10 some of,
1:19:11 some of the obvious
1:19:12 cases here would be minority language development,
1:19:15 um,
1:19:16 you know,
1:19:17 there's absolutely very,
1:19:18 you know,
1:19:19 no business incentive,
1:19:20 very difficult business models that,
1:19:22 that will incentivize,
1:19:24 uh,
1:19:24 uh,
1:19:24 uh,
1:19:25 the major tech labs to come in to do that.
1:19:27 Um,
1:19:27 competition policy,
1:19:28 you know,
1:19:29 to Catherine's point,
1:19:30 uh,
1:19:31 uh,
1:19:31 focusing on lower prices for consumers,
1:19:33 allowing switching costs.
1:19:34 Again,
1:19:34 that doesn't quite,
1:19:35 uh,
1:19:36 uh,
1:19:36 you know,
1:19:37 account for that.
1:19:38 And so,
1:19:39 so I think just,
1:19:40 just,
1:19:40 you know,
1:19:41 what I'm hearing from policymakers in developing countries is,
1:19:43 is that,
1:19:44 is
1:19:44 that,
1:19:45 that's the push for,
1:19:46 for the need for local capacity.
1:19:49 Thank you.
1:19:50 Uh,
1:19:50 the floor is open,
1:19:51 please,
1:19:51 uh,
1:19:52 come up to the mic,
1:19:53 uh,
1:19:54 short introduction and a short question would be ideal.
1:19:58 Hello,
1:19:59 I'm short.
1:20:01 Daria Taoni,
1:20:02 uh,
1:20:02 we're back,
1:20:03 um,
1:20:04 so thanks.
1:20:05 I mean,
1:20:05 I want to,
1:20:06 um,
1:20:07 I don't know,
1:20:08 uh,
1:20:08 I guess reflect on two takeaways that I took from,
1:20:11 uh,
1:20:12 this chat.
1:20:13 The one is that,
1:20:14 uh,
1:20:14 AI automates routine tasks,
1:20:16 and there is nothing new.
1:20:18 And the second one
1:20:20 is that we are in the pre-business model.
1:20:23 So it might be true that we are in the pre-business model,
1:20:27 but I think it's important to think that this pre-business model
1:20:31 is already reshaping labor and it is having systemic implications,
1:20:35 and it has to do with the fact that AI is based on the gig economy,
1:20:39 right?
1:20:40 And so the supply chain of AI is
1:20:42 structured around squeezing the maximum value from vulnerable
1:20:47 populations.
1:20:48 The annotators,
1:20:49 which are the people checking the code that are paid 90 cents.
1:20:53 Per hour
1:20:54 they are taken from countries that normally
1:20:56 are imploding Venezuela,
1:20:59 Kenya,
1:20:59 and then in turn all the countries that were uh imploding uh after the COVID
1:21:06 crisis
1:21:07 and so all these people on top there are algorithms that check peaks of productivity
1:21:13 and they take away tasks from people that are or or communities that
1:21:17 are becoming solely reliant on that gig
1:21:20 economy because everything else has imploded.
1:21:22 And so all these people are part of a growing global informal
1:21:27 and regulated workforce where there is some basic laws around labor that
1:21:33 we established some 100 years ago that are not reflected
1:21:37 in the economy because there is no stable employment,
1:21:40 no benefits,
1:21:41 no protection,
1:21:42 and not even an employer you can go to and and complain.
1:21:46 So and hence
1:21:50 my plea is and comes to some of the points that were made
1:21:54 is that we need to recognize in this pre-business model
1:21:58 and quantify and price in the human and social costs.
1:22:02 Otherwise the
1:22:04 conversation about AI will be dangerously incomplete.
1:22:08 So labor protections,
1:22:09 transparency,
1:22:10 fair compensation,
1:22:11 community control
1:22:12 is necessary
1:22:13 unless we are,
1:22:15 uh,
1:22:15 we want to replicate some colonial year extraction
1:22:19 and call it progress.
1:22:21 Thanks.
1:22:22 Thanks.
1:22:22 We'll just in the interest of time,
1:22:24 just collect all the questions and then we can have one round of replies.
1:22:27 Thank you.
1:22:27 Please,
1:22:27 please,
1:22:28 no,
1:22:28 no commentary,
1:22:29 just a short question.
1:22:30 OK,
1:22:31 so,
1:22:31 uh,
1:22:31 hi,
1:22:32 this is Elia from Syracuse University.
1:22:33 Thank you all for your,
1:22:35 uh,
1:22:35 thought-profiling,
1:22:36 uh,
1:22:36 talks.
1:22:37 I specifically appreciated Arvin's,
1:22:40 uh,
1:22:40 contextualization of AI within the,
1:22:42 uh,
1:22:43 history of technology innovations,
1:22:44 and,
1:22:45 um,
1:22:46 uh,
1:22:46 power and progress from Ajamola and Johnson received a lot of praise from,
1:22:51 uh,
1:22:51 like scholars from
1:22:52 a lot of
1:22:54 A variety of disciplines,
1:22:55 as they pointed out to the,
1:22:57 uh,
1:22:57 fact that diffusion of tech innovations is not uh like a natural evolution.
1:23:01 Rather,
1:23:03 what will be automated and how is uh like political decisions
1:23:07 and that the institutions are central to that decision making.
1:23:11 So my question is,
1:23:11 what do you think about the role of
1:23:13 development organizations in giving uh like policy direction and
1:23:17 AI policymaking?
1:23:19 Thank you.
1:23:22 Thank you very much for your articles.
1:23:25 My name is Chagoba.
1:23:25 I'm a founder of an AI startup.
1:23:28 My question for the three presenters,
1:23:30 um,
1:23:30 I,
1:23:31 um,
1:23:31 assume that the work that you have done
1:23:34 was based on the generative AI,
1:23:35 um,
1:23:36 that exists today.
1:23:37 Um,
1:23:38 there's a lot of indication that Agentic AI would be actually the
1:23:41 most important when it comes to the impact on economies and,
1:23:44 uh.
1:23:44 Society,
1:23:45 so do you think that
1:23:46 if you look at your work again
1:23:49 with the prospective agent AI,
1:23:51 do you predict the same results or not?
1:23:54 Thanks.
1:23:56 Thank you.
1:23:57 Um,
1:23:58 this is not really a fair question,
1:24:00 but I still want to hear your insights.
1:24:03 Uh,
1:24:04 Daniel,
1:24:05 uh,
1:24:05 Mexico has a huge education system crisis right now.
1:24:09 How do you see the use of AI
1:24:12 for education development in under the serving areas of the developing world,
1:24:17 especially when infrastructure challenges are acute.
1:24:21 And government has stepped back from education as a policy priority.
1:24:26 Thank you.
1:24:28 Hello,
1:24:29 I'm Abdullah.
1:24:30 Um,
1:24:30 my question is kind of broad,
1:24:32 but also specifically for Catherine as well.
1:24:34 So our role as economists is kind of to
1:24:37 learn from the mistakes of the past that we've made
1:24:39 in terms of global economy and like not make them in the future.
1:24:42 Um,
1:24:43 and Catherine mentioned that economics doesn't change,
1:24:45 but
1:24:46 If all of this predictive inference is wrong,
1:24:48 and we've seen that they've been proven wrong,
1:24:50 but radiology and all of those examples,
1:24:52 perhaps then economic economics needs to change in
1:24:55 order for our predictions to be right,
1:24:56 because the consequences
1:24:58 of being wrong are
1:25:00 that we might just end up in a bubble,
1:25:01 like the dot com bubble.
1:25:03 And
1:25:03 the consequences are multi-tiered in that case.
1:25:05 So how do you guys think
1:25:07 in your knowledgeable opinions that we need to
1:25:10 change the respective inference to be more right,
1:25:12 and make sure that we don't end up in another
1:25:14 global crisis where people are impacted,
1:25:16 technology is impacted,
1:25:18 and this just ends up being a repetition of the same bubble again.
1:25:21 Thank you.
1:25:21 Thank you.
1:25:22 Last question,
1:25:22 and then we'll get responses.
1:25:24 Um,
1:25:25 first of all,
1:25:25 thank you very much for the insightful discussion and the presentation.
1:25:28 I really appreciated it.
1:25:29 Um,
1:25:30 going right to the question.
1:25:31 Um,
1:25:32 so it's a question regarding the use of AI in the development field as one of,
1:25:35 if not the most versatile and volatile element of the digital public goods.
1:25:40 So for digital ODA,
1:25:42 um,
1:25:42 in the terms of the soft infrastructure,
1:25:44 what will be the main incentives for donor
1:25:46 countries to actually provide aid to the recipients?
1:25:49 And on that note,
1:25:50 what kind of the policy implications can we infer from that?
1:25:53 Thank you.
1:25:55 OK,
1:25:55 there's one more,
1:25:56 please go ahead,
1:25:56 but we'll have to cut it there because
1:25:59 Thank you so much.
1:26:00 My name is Maya.
1:26:01 Um,
1:26:01 I really appreciated the fact that you were challenging this view
1:26:04 that AI will replace jobs and instead will change them.
1:26:08 So my question is,
1:26:09 if you envision this idea that maybe AI will create
1:26:13 new industries on jobs that will support um developing countries,
1:26:17 and if so,
1:26:18 what would you picture that those new jobs or industries to be like?
1:26:23 So maybe you could take a minute each to
1:26:26 pick and choose whichever question you want to respond to.
1:26:28 Thank you.
1:26:32 Uh,
1:26:33 great,
1:26:33 yeah,
1:26:33 lots of great questions.
1:26:34 So,
1:26:35 um,
1:26:35 I'd,
1:26:36 I'd say in terms of there are a few questions that we're thinking about
1:26:39 the future,
1:26:39 new jobs,
1:26:40 etc.
1:26:41 um.
1:26:42 I think
1:26:43 in in the near term there's
1:26:45 there's some
1:26:47 wins
1:26:47 where AI seems quite amenable to improving education
1:26:51 and health service delivery in in many countries
1:26:54 that can't afford
1:26:55 great human providers for all all people
1:26:58 um and so
1:26:59 I think there is,
1:27:00 there is a lot of opportunity it's,
1:27:02 it's difficult to to know exactly how to integrate
1:27:06 uh AI assistance into classrooms and I think that's a challenging problem.
1:27:10 I don't think there's any.
1:27:11 Um,
1:27:12 we don't quite know what the optimal structure
1:27:14 of teaching is that integrate this,
1:27:16 um,
1:27:17 so there's a lot of questions there.
1:27:19 And then
1:27:20 there there also is a lot of uncertainty about the long run trajectory about
1:27:24 what capabilities future AI will have and
1:27:26 also as it's increasingly integrated in the economy
1:27:30 if these kind of
1:27:31 bold predictions about
1:27:33 economic effects will will pan out and I think
1:27:35 there's some chance that it that it will.
1:27:37 I think
1:27:38 the job implications for low income countries,
1:27:40 I just have a lot of uncertainty about what that could look like.
1:27:43 I,
1:27:43 I do think we will be surprised by by the new jobs that emerge though.
1:27:48 Catherine.
1:27:49 Um,
1:27:51 I just wanted to highlight two things I
1:27:53 took from the questions about what we're probably
1:27:55 getting wrong right now in the economics of
1:27:57 AI with our focus on high-income countries.
1:28:02 I think the first thing is
1:28:04 that as you can hear,
1:28:06 we basically
1:28:07 hate that,
1:28:07 you know,
1:28:08 sort of going off the task-based model saying
1:28:10 thinking about AI as task replacement is,
1:28:13 you know,
1:28:13 not helpful.
1:28:14 But I thought what was useful is just a reminder
1:28:17 that the West or developing country or,
1:28:19 or high income countries have outsourced
1:28:21 and a lot of tasks
1:28:24 to low-income countries.
1:28:26 And therefore,
1:28:26 we should think a little bit about that before we just say,
1:28:29 you know,
1:28:29 tasks will change and shift and so on.
1:28:32 Um,
1:28:32 the other thing I want to say is,
1:28:33 you know,
1:28:34 again,
1:28:34 the economics of technology tends to draw a lot of reassurance from the fact that,
1:28:39 you know,
1:28:39 when we study what happened to telephone operators with automatic switchboards,
1:28:43 when we look at what happened to accountants,
1:28:45 when the calculator was invented,
1:28:47 it looks really good,
1:28:49 uh,
1:28:49 in 10 years.
1:28:51 But we tend to forget about the pain,
1:28:53 the upheaval,
1:28:54 the horribleness of the 10 years in the intervening period,
1:28:58 and I think that's again something that we don't do enough of.
1:29:03 Thank you.
1:29:04 Um,
1:29:04 I'm gonna address briefly two of the questions.
1:29:06 One was about generative versus agentic AI,
1:29:09 and,
1:29:09 uh,
1:29:09 our framework is not specific to generative AI.
1:29:12 In fact,
1:29:12 I lead a team of technical researchers.
1:29:14 The main area we've been working on over the last couple of years is agentic AI.
1:29:18 So that's been deeply on our mind when we wrote.
1:29:20 AI as normal technology
1:29:21 and it is agnostic to what form of AI will be at the frontier in 10 years,
1:29:25 20 years,
1:29:27 and I want to give one note of caution around the hype around generative
1:29:30 AI and I say this as someone you know who finds it very useful.
1:29:33 I spend several hours a day using coding agents and they're fantastic.
1:29:37 But when you look at
1:29:39 the kind of agentic AI that you would need to have transformative rapid
1:29:43 transformation of the economy,
1:29:44 I think you need three things.
1:29:45 One,
1:29:46 you need them to operate with little human intervention relatively autonomously.
1:29:50 If you're,
1:29:51 if you have a human in the loop,
1:29:52 you're not really achieving that much in terms of productivity gain.
1:29:55 You need them to operate
1:29:58 in tasks which have high costs of errors.
1:30:02 So
1:30:03 if you,
1:30:04 you know,
1:30:04 if you're just getting responses from a chatbot,
1:30:06 that's very different from doing medical diagnoses or something like that.
1:30:11 So we identified these
1:30:13 three aspects of autonomy and agency and generality.
1:30:17 It has to be one general purpose system as opposed to something that
1:30:20 you had to spend 10 years developing for a particular kind of worker.
1:30:23 When you look at these three factors,
1:30:25 there are 0 deployed systems right now
1:30:28 as far as I could tell,
1:30:29 and we predict that it's going to be 0 for a few years to come.
1:30:33 So unless that changes,
1:30:34 while agentic AI is very exciting from a technological perspective,
1:30:37 I would not count on it being a counterexample of the kinds of things I talked about.
1:30:42 OK,
1:30:42 so very quickly on new jobs,
1:30:44 um,
1:30:45 we're big on the potential for AI to create lots of new jobs.
1:30:49 Obviously can't predict with any certainty,
1:30:51 but I think.
1:30:51 The needs to be very attuned to this possibility.
1:30:55 I really like David Autor's paper called Rebuilding
1:30:58 Middle Class Jobs Using AI.
1:31:00 So he talks about historical transformations that went from
1:31:03 the artisanal expertise of,
1:31:04 so let's say,
1:31:05 potters to the mass expertise of people who could use machines
1:31:09 to perform those in a much more efficient manner.
1:31:11 That was a lot more people
1:31:13 and depending on demand effects,
1:31:14 you might actually have an increase in labor if the demand for,
1:31:17 you know,
1:31:17 pottery goes way up because it's cheaper,
1:31:19 right?
1:31:20 So a similar thing might happen with AI.
1:31:22 You might have a result of pervasive cognitive automation.
1:31:26 Could be that you have mass expertise of things that
1:31:29 today require 10 years of training like law and medicine.
1:31:32 A lot of the aspects of those professions could be performed
1:31:36 by people.
1:31:36 With let's say 6 months to a year of training
1:31:39 who are using it in a way it's it's a human AI team
1:31:43 and to take advantage of that,
1:31:45 what we will need
1:31:46 is new rapid credentialing so that you can have some licensing of these people,
1:31:49 but it doesn't involve the same 10 year pipelines that we have today.
1:31:53 Thanks.
1:31:53 So we're passed our time.
1:31:54 I will not try to summarize this session,
1:31:56 but the three big takeaways for me were one,
1:31:59 and I made a note of these
1:32:01 one,
1:32:01 be humble and acknowledge the uncertainty.
1:32:04 2,
1:32:05 the more things change,
1:32:06 the more they stay the same.
1:32:07 And 3,
1:32:08 be serious about your research,
1:32:10 collect the evidence to help guide the decision-making,
1:32:13 which,
1:32:13 as it turns out,
1:32:14 I guess are just life lessons.
1:32:15 So,
1:32:16 uh,
1:32:16 so,
1:32:17 I would like to thank the panelists uh for their wisdom,
1:32:19 and I thank you all for joining.
- add-style
- lp-body-content
A video recording of the Session 2 on Day 1 of The Annual Bank Conference on Development Economics 2025 "Development in the Age of Populism." This session discusses "Artificial Intelligence".
CHAIR: Gaurav Nayyar, Director of the World Development Report 2026, World Bank
PANELISTS:
Paper 1: AI as Normal Technology (Arvind Narayanan, Princeton University)
➜ Presentation | ➜ Paper
Paper 2: How does Competition Policy Need to Change in a World of Artificial Intelligence? (Catherine Tucker, Massachusetts Institute of Technology)
➜ Presentation | ➜ Paper
Paper 3: Could AI Leapfrog the Web? Evidence from Teachers in Sierra Leone (Daniel Björkegren, Columbia University)
➜ Presentation | ➜ Paper
DISCUSSANT: Han Sheng Chia, Policy Fellow, Center for Global Development