Transcriptly

AI Is Killing the Career Ladder. A Stanford Economist Explains What Comes Next | Bharat Chandar

EO15:41Added Aug 31, 2026

Bharat Chandar, postdoctoral researcher at Stanford's Digital Economy Lab, breaks down why young workers in AI-exposed jobs are seeing 16% slower employment ...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00We're seeing that the jobs that are more
  3. 00:01exposed to AI, the young workers in
  4. 00:03those jobs are seeing 16% slower
  5. 00:05employment growth. So that's pretty
  6. 00:07large. The structural change in AI
  7. 00:10capabilities that are impacting the
  8. 00:12labor market, that's not going to be a
  9. 00:13temporary change. If we really unlock
  10. 00:15AI's capabilities for helping people
  11. 00:17learn, it could be much easier to switch
  12. 00:19between different professions. So I'm
  13. 00:21really hopeful that we end up somewhere
  14. 00:23closer to a career lattice that works
  15. 00:25for workers as opposed to a career
  16. 00:27ladder where there's much more risk
  17. 00:28about this technological change. I'm
  18. 00:30Barat Chandra. I'm an economist at the
  19. 00:32Stanford Digital Economy Lab and I study
  20. 00:35how AI is impacting work. I would say
  21. 00:37over the past year and a half or so, I
  22. 00:39do feel like one of the most important
  23. 00:41questions in labor economics today is
  24. 00:43about AI's impact on the labor market.
  25. 00:45Once I really started using the tools
  26. 00:47more and understood their capabilities,
  27. 00:49that became the focus of my research
  28. 00:51agenda because it felt like one of the
  29. 00:52most important questions impacting
  30. 00:54society potentially in the future.
  31. 01:04I released a study with my collaborators

  32. Canaries in the Coal Mine - What is happening to entry-level workers?

  33. 01:06Eric Bolson and Ryu Chen. We studied how
  34. 01:10jobs were changing in jobs that were
  35. 01:12more exposed to AI versus less exposed
  36. 01:14to AI. and we were tracking millions of
  37. 01:15workers across the United States using
  38. 01:17data from a payroll company called ADP.
  39. 01:19One of the key findings there is that
  40. 01:21overall we were not seeing major
  41. 01:23differences in employment changes for
  42. 01:25jobs that were more and less exposed to
  43. 01:27AI. However, when we focus on young
  44. 01:29workers, we do see more of a divergence
  45. 01:31there where the jobs that are more
  46. 01:33exposed to AI, such as software
  47. 01:34development, customer service, more
  48. 01:36administrative roles, we were seeing
  49. 01:38employment declines and jobs that were
  50. 01:40less exposed to AI, we were still seeing
  51. 01:42some continued growth and employment.
  52. 01:44And for more experienced workers as
  53. 01:46well, we were still seeing employment
  54. 01:47growth that was pretty much on trend.
  55. 01:50We're seeing that the jobs that are more
  56. 01:51exposed to AI, the young workers in
  57. 01:53those jobs are seeing 16% slower
  58. 01:55employment growth. a lot of people just
  59. 01:57starting off in their careers and
  60. 01:58they're finding it a hard time in doing
  61. 02:01that. And the reason we chose the
  62. 02:03canaries and the coal mine title is I
  63. 02:04think consistent with that view. We want
  64. 02:06to be tracking these outcomes because we
  65. 02:08think they could be indicative of
  66. 02:10potentially future transformative
  67. 02:12impacts of AI. How much of this is being
  68. 02:14driven by AI and how is that going to
  69. 02:16change going forward? We can't be sure
  70. 02:18whether this is just temporary change in
  71. 02:21the economy or if it's a structural
  72. 02:22change being driven by AI. Now, we did
  73. 02:25test some of the most plausible
  74. 02:26alternatives that we could think of. So,
  75. 02:28that includes interest rate changes.
  76. 02:30Jobs that are more exposed to interest
  77. 02:32rate changes are actually less exposed
  78. 02:34to AI. One way to think about that is
  79. 02:36things like transportation and
  80. 02:38construction are very exposed to
  81. 02:39interest rate changes, but they're
  82. 02:40really not very exposed to AI. So,
  83. 02:42that's one key thing and that makes me
  84. 02:44think that it's probably not interest
  85. 02:45rate changes that are driving our
  86. 02:46results. Other things that we tested
  87. 02:48include tech over hiring. So, we can
  88. 02:50take out the tech sector, we get similar
  89. 02:51results. Take out computer jobs. We
  90. 02:53tested some of these different
  91. 02:54alternatives and we were still getting
  92. 02:55this very similar results. So, like
  93. 02:57you're saying, if it's a structural
  94. 02:58change in AI capabilities that are
  95. 03:01impacting the labor market, that's not
  96. 03:02going to be a temporary change. That's
  97. 03:04potentially going to be a long run
  98. 03:05change. And the longer that we can track
  99. 03:07this and if those trends still seem to
  100. 03:08hold up, that would be indicative of AI
  101. 03:11potentially impacting work. Now, that
  102. 03:13said, we don't have an experiment where
  103. 03:14we can compare a world with AI to one
  104. 03:16without AI and do a clean comparison.
  105. 03:19There's definitely a lot more work to be
  106. 03:21done here to really tease apart the
  107. 03:23impact of AI.

  108. Young Workers Lost Their Edge. Where’s the new one?

  109. 03:26If you think about what young workers
  110. 03:28are doing when they're entering the
  111. 03:30workforce, a lot of it is
  112. 03:31implementation, doing things that rely
  113. 03:34on the book knowledge that they learned
  114. 03:35while they were at school. Whereas the
  115. 03:37things that they don't have as much
  116. 03:38experience with an ability to to do is
  117. 03:41relying on the tacet knowledge or the
  118. 03:44sort of experience that you can only get
  119. 03:45by doing things on the job. also more
  120. 03:47social interaction and more strategic
  121. 03:49thinking. Tacid knowledge I think of as
  122. 03:52things that rely on a lot of hyper local
  123. 03:55context or strategic thinking or social
  124. 03:58interaction or things that you only
  125. 04:00build via experience on the job. So
  126. 04:02those are the types of things that are
  127. 04:04maybe not written down as much in a
  128. 04:06book. For young workers, it's more
  129. 04:07directly overlapping with the AI
  130. 04:09capabilities. And those could be the
  131. 04:11sort of situations where more
  132. 04:12experienced workers might have a
  133. 04:14relative advantage compared to AI and
  134. 04:16also compared to young workers. When it
  135. 04:18comes to training young workers, it's
  136. 04:20totally right that firms will want to
  137. 04:22hire young people if they want to have a
  138. 04:25middle management or more experienced
  139. 04:26staff going forward. Now the issue here
  140. 04:28is even though that they have some
  141. 04:30incentive to do that so that they have
  142. 04:32workers in the future, they might not
  143. 04:34have enough incentive to do that. So
  144. 04:36they might not hire as much young people
  145. 04:38as they should from a social perspective
  146. 04:40and they might not train them as much as
  147. 04:42they should. And the reason that's the
  148. 04:43case is because those young people don't
  149. 04:45have to stay at the company forever.
  150. 04:47They can just go leave to another
  151. 04:48company. So it's true that they will
  152. 04:50still want to hire some of them, but
  153. 04:52they might not want to hire as many as
  154. 04:54would be beneficial to society. And it's
  155. 04:56just kind of this mismatch between what
  156. 04:58is the incentive of the individual
  157. 05:00private company versus what is the
  158. 05:02incentive of society as a whole. Now the
  159. 05:04more optimistic take that I could give
  160. 05:06here is that if AI really is as capable
  161. 05:09of helping people learn uh and as a tool
  162. 05:11for education maybe could speed up the
  163. 05:13process at which that happens that could
  164. 05:15also require a lot of changes in the way
  165. 05:16that we organize our education system
  166. 05:19potentially universities or even at a
  167. 05:20lower level than that to help people
  168. 05:22learn faster and better. There are three
  169. 05:24things that I think AI is going to be
  170. 05:26much less capable of doing certainly in
  171. 05:28the short to medium term. One physical
  172. 05:30tasks unless we see a big advance in
  173. 05:32robotics. Number two is strategic
  174. 05:35thinking and guiding what needs to be
  175. 05:37done. And number three is social
  176. 05:39interaction. I think the strategic
  177. 05:40thinking is increasingly important and
  178. 05:43it's going to be even more important
  179. 05:44going forward potentially because it
  180. 05:46does seem like in the future a lot of
  181. 05:48work might look like guiding AI agents
  182. 05:51to do implementation while you're
  183. 05:52telling them and guiding them on what
  184. 05:54needs to be done. And so that sort of
  185. 05:56strategic thinking, expressing what it
  186. 05:58is that needs to be done or what I want
  187. 06:00to be produced, I think that's going to
  188. 06:02be a pretty key skill and that's kind of
  189. 06:04the role of what a manager does within a
  190. 06:06company. So that sort of managerial work
  191. 06:08and strategic guidance could potentially
  192. 06:10be a quite important skill going
  193. 06:12forward. When I think about young
  194. 06:14workers, how can they develop those
  195. 06:16sorts of skills? building and using the
  196. 06:17tools as much as possible and getting
  197. 06:20used to to that sort of mode of work.
  198. 06:22The faster that that can happen, the
  199. 06:24better that they might be uh in terms of
  200. 06:26adjusting to labor market disruptions or
  201. 06:29these technological changes.
  202. 06:33I do think it's very helpful to compare

  203. From Career Ladder to Career Lattice

  204. 06:35AI to some of these historical changes.
  205. 06:37So for example, the industrial
  206. 06:39revolution. I think one comparison
  207. 06:41between AI and that period that was a
  208. 06:43case where it was actually the most
  209. 06:45skilled workers who faced more risk from
  210. 06:47the industrial revolution. So one case
  211. 06:49that comes to mind is the lites who were
  212. 06:51these kind of skilled textile workers
  213. 06:54and the new inventions that came about
  214. 06:56during the industrial revolution
  215. 06:57actually led a lot of them to lose their
  216. 06:59work and those were kind of the more
  217. 07:00skilled workers in society. Something
  218. 07:02that you might be seeing that's kind of
  219. 07:03similar here is that it's more of the
  220. 07:05knowledge workers in more educated roles
  221. 07:07that might be facing greater AI
  222. 07:09exposure. So I think that's an
  223. 07:10interesting comparison. If we think
  224. 07:12about things like electricity or the IT
  225. 07:15revolution, so basically over the course
  226. 07:17of the 20th century, a lot of those were
  227. 07:19actually kind of the opposite where it
  228. 07:21was kind of this middle skill or
  229. 07:23low-skilled work that tended to be more
  230. 07:25exposed to that technology. Whereas the
  231. 07:28most skilled, the highest educated
  232. 07:30people benefited a lot more from the
  233. 07:32development of this new technology. So
  234. 07:33we still have to see going forward, is
  235. 07:35AI going to look more like the first
  236. 07:37case or the second case? I do think
  237. 07:39there's something worth bearing in mind
  238. 07:40here. So one way that AI might be
  239. 07:42different than past historical episodes
  240. 07:44is just the rate of capabilities
  241. 07:46improvement. Even today it's much more
  242. 07:48capable of doing different tasks than it
  243. 07:49was 3 years ago and I do think there's
  244. 07:51this question about as new work gets
  245. 07:53created there's new demand for existing
  246. 07:56work etc. Are those going to be done by
  247. 07:59humans or are the AI capabilities going
  248. 08:01to advance fast enough that AI is also
  249. 08:03going to be doing that kind of work? And
  250. 08:04I think that's the one area where we
  251. 08:06could think that potentially AI could be
  252. 08:07different than prior technologies.
  253. 08:09There's been a lot of discussion about
  254. 08:11how we can use AI to augment workers and
  255. 08:14make them better off as opposed to
  256. 08:15potentially just substituting them from
  257. 08:17the workforce and automating all
  258. 08:19everything that they're doing. Where I
  259. 08:21was going with this essay is just trying
  260. 08:22to suggest one concrete solution that I
  261. 08:26think could potentially augment workers
  262. 08:28quite a bit. It using AI as a tool for
  263. 08:31helping people learn. I think an example
  264. 08:33of a person who's augmenting themselves
  265. 08:35with AI right now. An example of that
  266. 08:38would be a startup founder with a really
  267. 08:40lean team that's able to do a lot more
  268. 08:42tasks because they have access to the
  269. 08:44AI. All of the different functions that
  270. 08:46previously they wouldn't have had any
  271. 08:47idea how to do. Now they can do it
  272. 08:50themselves because they have access to
  273. 08:51these AI tools. I I think that's a very
  274. 08:53good example in fact of augmentation.
  275. 08:55Whether you're more automated or
  276. 08:56augmented really depends on what are the
  277. 08:58tasks that you're focusing on. Are you
  278. 09:00increasing the scope of tasks that you
  279. 09:02can do or are your tasks getting shrunk
  280. 09:04by the introduction of this technology?
  281. 09:06The reason that I think that this could
  282. 09:08be wonderful in terms of augmentation is
  283. 09:10that when we think about technology that
  284. 09:12benefits workers, often it is increasing
  285. 09:15the set of tasks that they're able to
  286. 09:17do. In contrast, things that automate
  287. 09:19work that substitute for workers, those
  288. 09:21are things that take away some of the
  289. 09:23tasks that workers have to do and now
  290. 09:25they have to do fewer things. know I
  291. 09:26think the goal is to try to find ways to
  292. 09:28augment workers to make them more
  293. 09:31capable of doing things. And one of the
  294. 09:33best ways that we know historically for
  295. 09:35doing that is by educating them. With
  296. 09:37education, workers are able to do a lot
  297. 09:39more than they could do before. I do
  298. 09:41think there's something worth bearing in
  299. 09:43mind here. So one way that AI might be
  300. 09:45different than past historical episodes
  301. 09:47is just the rate of capabilities
  302. 09:49improvement. And I think we have an
  303. 09:51opportunity right now for one of the
  304. 09:54biggest changes in learning capabilities
  305. 09:57that we've had in 100 years if not
  306. 09:59longer. And that's in using the AI tools
  307. 10:01for personalized learning. For me, using
  308. 10:04AI for augmentation, there's a couple
  309. 10:07branches to that. And something that
  310. 10:08increasingly I'm using it for is
  311. 10:09actually for math. There are areas where
  312. 10:12I might need to write down a model or
  313. 10:14prove something. And it's really, really
  314. 10:16good at that. The way that that's
  315. 10:18augmenting is that it's easier to check
  316. 10:19if something is correct than it is to
  317. 10:22necessarily write it from scratch. And
  318. 10:24so I also potentially view that as a as
  319. 10:27a significant way of augmenting my work.
  320. 10:29Now, on the other hand, things that I
  321. 10:30don't do with AI, I personally don't
  322. 10:33really use it for writing. And the
  323. 10:34reason I don't use it for writing is
  324. 10:35that writing helps me think and it helps
  325. 10:38me understand a problem really well when
  326. 10:40I do it myself. It's not that I don't
  327. 10:42trust the AI tools to do the writing.
  328. 10:45It's more that I would get way less
  329. 10:47value out of the writing if I didn't do
  330. 10:48it myself and understood what it was
  331. 10:50that I was talking about. I think in
  332. 10:51deciding what we want to delegate and
  333. 10:53what we want to preserve as human, I
  334. 10:56think a lot of that depends on what it
  335. 10:58is that humans want and some of that is
  336. 11:00about values like what is right, what is
  337. 11:02wrong. Some of that is also just
  338. 11:04expressing our preferences. So what do
  339. 11:07we want to build? What would make us
  340. 11:08better off? What would make us happier?
  341. 11:10What are the things that we would
  342. 11:11actually want to use AI for for
  343. 11:13implementation? That's something that we
  344. 11:15have to express to the AI. I think those
  345. 11:17sorts of tasks, it's not obvious to me
  346. 11:19how that's going to be automated, you
  347. 11:21know, in the short to medium term at
  348. 11:22least because some of that both it
  349. 11:25depends on also our reflection. We need
  350. 11:27to think through what it is that we
  351. 11:29want. Sometimes we learn about what it
  352. 11:30is that we want as we reflect on it and
  353. 11:33as we think more about it. And so that
  354. 11:35guidance about what it is that we should
  355. 11:37build, what it is that we should
  356. 11:38implement, that I view as at least in
  357. 11:40the short to medium term being more
  358. 11:42characteristically human than AI. And I
  359. 11:44view the AI is more on the
  360. 11:45implementation side.
  361. 11:48Imagine that AI really reduces the
  362. 11:51benefits of learning something new. An
  363. 11:53interesting thing about this is that
  364. 11:54that's a world where potentially
  365. 11:55inequality is much lower in the labor
  366. 11:57market. If the barriers to getting at
  367. 11:59the top of a field or getting the
  368. 12:02highest quality output in a given
  369. 12:04occupation or something, if that barrier
  370. 12:07becomes much lower because AI can do a
  371. 12:09lot of the hardest tasks, then that's
  372. 12:11actually a world where potentially
  373. 12:12inequality is lower because the
  374. 12:14difference between people who know a lot
  375. 12:16in school and are very capable could be
  376. 12:19not that different from people who don't
  377. 12:21try that hard in school. It's kind of
  378. 12:22interesting because there's this
  379. 12:24potentially trade-off between inequality
  380. 12:26and investments in learning. On the
  381. 12:28other hand, if AI really increases the
  382. 12:31benefits of this sort of strategic
  383. 12:32thinking, even these social skills,
  384. 12:34etc., that could actually increase the
  385. 12:36benefits of trying really hard in school
  386. 12:38because if I can develop the strategic
  387. 12:40thinking skills, then I'll be really
  388. 12:42valuable in the labor market.
  389. 12:44I would encourage young people, students
  390. 12:46to use the AI tools as much as they can
  391. 12:50uh build with them and really focus on
  392. 12:52developing that kind of strategic
  393. 12:53thinking. How do you best make use of
  394. 12:55these tools? Where are the areas where
  395. 12:57they're not as good and what are areas
  396. 12:59in which you as a human can add a lot of
  397. 13:01value? I think there are some very
  398. 13:03complicated ways of thinking about how
  399. 13:05AI might affect thinking going forward.
  400. 13:07There are some newer interventions that
  401. 13:10are happening here uh education style
  402. 13:12modes with AI usage to make people focus
  403. 13:16on the critical thinking skills as
  404. 13:18opposed to just offloading the task. So
  405. 13:20there are different sort of platforms. I
  406. 13:22know uh Khan Academy for example has one
  407. 13:24where you can use the AI tool but it's
  408. 13:26not going to give you the answer. It's
  409. 13:28going to help you think through how to
  410. 13:31get to the answer as opposed to just
  411. 13:32giving it to you off the bat. I do think
  412. 13:34that we could imagine a world in the
  413. 13:36future where if we really unlock AI's
  414. 13:39capabilities for helping people learn,
  415. 13:41it could be much easier to switch
  416. 13:42between different professions based on
  417. 13:45how demand for those jobs is evolving
  418. 13:47over time. And if some uh job becomes
  419. 13:50much more important in the economy, if
  420. 13:52we can find a way to help people
  421. 13:54transition faster, that could really
  422. 13:56unlock a lot of potential. So I'm really
  423. 13:58hopeful that we end up somewhere closer
  424. 14:00to a career lattice that works for
  425. 14:02workers as opposed to a career ladder
  426. 14:04where there's much more risk about this
  427. 14:06technological change.
  428. 14:08For the first time I struggled to

  429. Next Episode

  430. 14:11assemble questions that chat GPT would
  431. 14:13get wrong. I was devastated was thinking
  432. 14:16how am I going to stay ahead of AI? I
  433. 14:20actually think that's the wrong
  434. 14:21question. My name is Ken Ono. I'm a
  435. 14:23mathematician and I also work in the
  436. 14:25space called AI for math. I'm a
  437. 14:27professor at the University of Virginia
  438. 14:28on leave and I'm the founding
  439. 14:30mathematician at Axiom Math. My view on
  440. 14:32intelligence now has changed quite a
  441. 14:35bit. We in this world aren't doing the
  442. 14:39best we can at educating our children.
  443. 14:43And I don't say that to be critical of
  444. 14:45educators. I am an educator. It's always
  445. 14:47a treat to visit a kindergarten class, a
  446. 14:51first grade class when it's bring your
  447. 14:53parent to school day so they can talk
  448. 14:55about what they do. Oh, I know all the
  449. 14:57prime numbers or I'm really good at
  450. 14:59adding that wonder and I want to just
  451. 15:01bottle up this energy because if we
  452. 15:04could maintain that wonder in the world
  453. 15:07and the energy that children have when
  454. 15:11everything around them is new. Think
  455. 15:13about where we would be today. The
  456. 15:15ability and the potential to be someone
  457. 15:17like Romanagen or at least creative in a
  458. 15:20productive way. I think it resides in us
  459. 15:23all. Who owns your identity? You do.