Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable. ======================================================================== # Generative AI and Pre-Hire Labor Market: Evidence from a Chinese Job Matching Platform Authors: Hanming Fang, Jiayin Hu, Jiaqi Zhao, Yu Miao Discussant: Lindsey Raymond Video: https://www.youtube.com/watch?v=8cqRUc0W7Sc&t=5025s ## Talk (01:23:45 – 01:49:23) [01:23:45] Oh, do I actually need to Oh, should I get started? Yeah. [01:24:03] >> Okay. So, thank you very much for including this uh this paper in the program. My my three co-authors actually online. They I don't think they can help me answer questions, but hopefully I do a good job that they will um approve my [01:24:16] representation of them here because this is uh we're going back to the AI's impact on the labor market in the Chinese context. Okay, we are going to use um data from a large Chinese um job ma matching platform. actually one of [01:24:31] the three that was in uh uh you sui's presentation yesterday. So there's rising concern worldwide on generative AI on labor market right and uh and if we learned anything from order uh Gordon [01:24:45] Hansen's research on China shock is that policy makers and researchers should be really aware of you know how uh the labor uh the workers are responding to the shocks right so you know the uh the [01:24:59] you know the the the the rising uh uh penetration of Chinese imports to the US manufacturing uh uh places right leads from job displacement. But there was a why initially there was too optimistic [01:25:13] uh um perception that the workers are going to adjust to these shock. They will be moving to sectors where the job opportunities are rising. They will be moving to places where more jobs are emerging. Well, it turns out not to be the case and that led to of course you [01:25:28] know consequences politically and economically. But um but I so this paper will try to ask the question in the the how do labor market respond to the AI shocks, right? AI shocks is going to be [01:25:42] probably bigger than the the the first or second China shock, right? But um and both from the labor demand side and labor supply side. Okay. So right now uh the most uh best known results about uh how um gener generative AI's labor [01:25:56] market implication probably is the the the paper that on the second line right Hoseni uh Mam and Linger uh the you know the the AI seem to be seniority bias right the senior senior workers uh they [01:26:11] are getting um you know better paid and their their job jobs are not being affected but juniors positions are disappearing in large numbers Okay. So um and the research broadly uh two two [01:26:23] two branches right one is that uh rel somewhat you know uh panel's presentation yesterday is partly on the first one the productivity impacts of AI adoption uh second part is label demand adjustment in employment hiring and job [01:26:37] postings uh the paper that I already mentioned um and also yu's paper le at all 2025 is is actually also co-authored with yu who presented yesterday less is known about how labor demand side [01:26:50] adjustments are transmitted through the prehiring matching process uh and and uh and look at also at the margins of adjustment that's not monetary right so and also um uh how do workers respond in [01:27:05] their job search behavior uh to these shocks and uh finally uh this has been mentioned a couple times China is not monolithic there's uh one advantage of China is that um there are lots of [01:27:18] region you know variations across uh Chinese provinces regarding uh labor regulations and we'll see uh that allows us the opportunity to look at the potential role of regul uh uh uh [01:27:31] regulation uh on mitigating the impact of AI. Okay. So we'll study China um and China's uh diffusion and adoption of large language models is uh very rapid. [01:27:44] Uh the uh by June 2025 the there were 515 million uh active users of uh large language models defined by people who have used uh large language models uh in [01:27:59] the last six months. Right? There's uh China also has rather weak labor protection and law enforcement uh and informality is more more severe and we'll see one of the takeaway message here is that wage is not the only [01:28:12] dimension that will you know that that uh uh that's that's relevant in China right actually working hours is another dimension right and we will see that AI uh adoption um uh AI uh uh um diffusion [01:28:26] might um impact for example senior workers not so much wages but in longer working hours. Okay. And then we uh the other thing is that we bring unique data uh that has both the job postings and [01:28:39] applications. So it's it's uh uh we have the information about uh the job applicants. What kind of job do they apply and so on. Okay. [01:28:49] So research question uh we will first look at the the kind of look at the the um the the the the impact of large language model shock on the labor demand. uh similar to what have been studied in the US context uh but [01:29:03] focusing both not only on monetary margins but also on non-monetary margins especially the so-called time amenities do you get would you be required to work long hours after you know uh uh um uh uh [01:29:17] after uh the um the regular eight hour workday right so China's actually average hours of work per week is 48 hours right um and then the second thing is how are these shocks transmitted to the applicant side. How do they adjust? [01:29:32] Are they really um you know um able to um be adjust their application behavior in response to these shocks? Um and then finally, how can the regional regulatory environment mitigate these shocks and [01:29:45] the uh empirical method is quite simple. We'll be looking at occupation level exposure uh interacting with the the the event data of CHA GPT release. So you I will explain why chai diffic is relevant for China in a second. Okay. And [01:29:59] occupation level expo exposure is based on elong at all's measure. Uh so think about each job um as considering of a bunch of tasks. All right. So elong at all they will look at uh you know for each task uh in a particular job how [01:30:14] likely the large language model will uh will will uh reduce the demand uh uh of human hours um to do the same quality of work. And the only difference is that for the you know job tax are bounding to [01:30:27] jobs differently between the US and China. So the occupation level exposure we will using it will be the elong at all measure adapted to the Chinese job characteristics. Okay. Okay. So um I'm [01:30:40] going to skip the the the the um uh preview of the results. Um literature a lot of people here contributed to it. I you know uh I only list a couple but the key thing I think uh um our contribution [01:30:54] is uh to bring in the prehiring stage um into the analysis allowing us to study the question of how the labor demand changes in responses to the large language mod shock will be transmitted to the applicants behavior. Second [01:31:09] contribution is we bring in the time related amenity um into the research uh into the measion dimension we we study and we found that it's a very elastic margin in developing countries. So what we found in China may be not so relevant [01:31:24] for the on the time uh amenity amenity measure may not be so relevant for the US but for developing countries where uh uh often workers are asked to work longer than their traditional workday uh this may be an important margin uh where [01:31:37] large model is going to have impact not only on less experienced workers but also more experienced workers and then we also find that uh a regional labor protection uh does have some minor mitigating effects. on large language [01:31:52] model sharks. Okay, so let's see this. Um, okay. So, the data set is coming from one of the actually I I think uh uh one of the three one of the largest uh [01:32:04] uh online job platforms uh one of our co-auth the co-author teams um they have collaborative a agreement with with this platform right um and the sampling [01:32:17] includes the data from 2021 to 2024. We don't cover the deepseek moment. We could in the uh in in in extension of the study. We could uh because the data is uh being uh continuously uh [01:32:29] augmented. Um and uh and the uh so we uh uh have the sample of total of 777,000 valid job postings. This is job posting data from the employer side. Okay. [01:32:42] covering uh uh four years and uh and uh two 200,000 random postings per year and uh eliminating some uh postings with duplication or missing variables, missing information, we are left with [01:32:55] 770,000, right? And it covers 183 industries and uh 93 second level occupation. So um um we'll be doing our analysis at the occupation level. Okay. [01:33:09] And it the job posting has information about job title, city where the job will be located, occupation, wage, education and experience requirements, benefits, right? Whether the standard wing the the the social insurance uh benefits, [01:33:23] medical insurance benefits, things like that will be paid or not, whether how many not vacancies uh are there for that job posting as well as uh job description tax. What kind of skills are required? I think Yu's presentation yesterday, you know, focus on the job [01:33:38] description. Okay. And uh um okay. Um the second uh data set is, you know, a a random sample of um from the same platform of applicants, right? So we uh [01:33:52] draw about um uh the applicants. We randomly sample about two 20,000 applicants each year and uh then we get information of all the job positions that these the randomly sampled [01:34:06] applicant uh uh apply for. Okay. And then we also know what happened to these applications. Did the employer give a so-called uh call back? uh call back basically is any response that um the [01:34:20] employ the um the um employer gave within 24 hours to the job application whether it's in exchange of contact information whether it was an email whether it was a you know wechathat uh [01:34:34] uh indication of whether some wehat communication took place um that's the very first step of know toward a formal hiring but we the the platform unfortunately does not keep any records [01:34:47] of whether a worker and and a vacancy actually matched. Okay, so with that the the call back rate is the best we can do. Okay, so this is a employee response give a positive reply within 24 hours. [01:34:59] Um and uh so the sample uh for the for the applicants we we have the applicant side sample as well employer employer side sample. Okay. Uh then we have the large lang large large language model [01:35:12] exposure index by occupation. This I already mentioned is uh following Elod at all 2025 by GBT's GBT uh uh paper uh methodology but adapted to China context [01:35:24] by uh John Dandan and his her co-authors in 2025. basically using the only difference is that the weights used in aggregating the LM exposure to the occupation level is using the tasks uh [01:35:38] the the fraction of uh of the importance of each task in the occupation in China context right so this is built upon a larger sample to 1.25 25 million job postings from the same platform from the [01:35:53] platform where we have our data. Okay. And we'll use the baseline measure uh uh pre-shock average exposure to large language model um measured uh in the baseline is 2021. [01:36:12] One idea of the large language model expo exposure in China uh at occupation level the mean exposure is like 0.59 okay and the one stand deviation.17 okay.17 you can think of that as the [01:36:27] difference in exposure between two occupations here hotel services versus lawyer hotel services is less exposed than lawyer okay so lawyers lawyer as a profession is one standard deviation [01:36:40] more exposed to large language model uh than hotel services and the the the the uh uh among all the occupations 93 occupations the the occupation that are at the mean exposure are operations [01:36:53] management and K to2 tutotoring just give you some concrete ideas of uh how exposures work right these uh and then we have the um um a job posting and vacancy share will be guaranteed or not. Believe it or [01:37:14] not actually uh the in China in Chinese context if a company can guarantee weekends off they will make that as a very desirable feature in their job posting. If that's not mentioned then [01:37:28] you can expect that you may be asked to work overtime or on weekends. Okay. So job amenity time amenity is the one thing that that we'll be focusing on. As you can see whether their time amenity is mentioned or not in in China uh only [01:37:43] 19% of job postings actually mentioned that weekends will not we'll let you have the two uh two days off right um okay um so uh in terms of the job [01:37:57] applicant side I will just briefly mention that um the applicants on any uh in our data set on um on average any A positive call back uh is average is [01:38:15] about seven uh about three days. All right. Uh uh days to so the the callback is defined as you know receiving a positive reply uh within 24 hours. So [01:38:28] days to receive first uh positive feedback since first application means that it will basically there on average it will be the third it will take three applications for them to listen to to [01:38:40] hear from the employer. Okay. Okay. So uh the first set of results is on demand side looking at the job posting characteristics. Uh the the empirical specification is quite uh straightforward. YT will be at the J is [01:38:55] a posting. YT is the posting level outcomes including wage, whether there is time, amenity and uh uh and uh uh educational experience requirements and occup large language model exposure GJ [01:39:08] is the the the job J's occupation. Okay. And post T is equal to one if it it's after December of 2022 the month after uh CHP is uh is released. Uh and then we [01:39:23] control for time fixed effects and the city desk job city fixed effect occupation fixed effect. Okay. And we'll be focusing on this beta coefficient and standard errors are clustered at [01:39:34] occupation type level. Okay. So um so the first one uh let's actually look at the panel B first looking at the individual job c job uh posting level outcomes we see that you [01:39:48] know the uh occupations with larger large language model exposure uh they will start to require more uh education the education requirement um you know bigger than three is actually um uh [01:40:02] higher than uh uh master degree right increased education requirement bigger than bachelor uh increased a log wage decre you know the wage decreased uh and uh time a time [01:40:13] time related amenity uh become uh uh uh lower right so in terms of uh quantity you can see that quantitatively it means that one one standard deviation increase in large language model exposure which [01:40:26] is 1.17 right I mentioned earlier will lead to a wage decrease about 2% and advertised time related amenity decrease by 1.7 percentage points um on the baseline of 19%. So it's about [01:40:40] 10% uh 10% uh uh reduction in time amenity and uh education requirements and exp uh and and also uh uh uh uh uh [01:40:53] become higher. Okay. And in terms of the quantity actually we look at the occupation level we aggregate occupation level um share of job postings in a given occupation right we see that [01:41:07] occupations where large language model exposure is higher after the uh the large language charg uh release uh this share uh decrease by by uh about uh 0.5 one percentage points one standard uh [01:41:22] standard uh deviation increase in large model exposure led to 0.51 percentage points in that occupation's uh job posting share and the vacancy share dropped by similar similar percentage right so there this is consistent with [01:41:36] uh with uh uh the findings uh uh in the literature and if in terms of dynamics um uh this is the CHP release date November of 2022 um and you then we have a couple of vertical lines indicating [01:41:50] China's large model uh release is the by BU is the first one to appear just a couple months after Chai GBT and then you have um Chen right by Alibaba and do [01:42:04] by dance right do now by now by far has uh by now has the largest market share so you may be thinking no well at the right away there's actually no effect but it's the moment Chinese uh large language models uh appear they the the [01:42:19] effects right away show up right um and it's it's important to remember that uh Xiinping actually started talking about AI in 2014 and the 2017 was when you [01:42:31] know uh he made um uh AI um key kind of uh strategy to boost Chinese productivity right but for a while nothing happened right so really no one it was a you know government objective [01:42:45] but there's not a lot of concrete development yet but then when chb got released I think there was I guess there's a jolt in the Chinese AI uh circle that something can really there could be very close to bring to the [01:43:00] market and I think the labor market immediately responded right um stronger um right we do see stronger adjustment after domestic AI tools uh became available okay and the hogenous effect is this is [01:43:13] where uh we do I think contribute to the first uh type of literature that we're related to right so we this is the dividing the in uh the applicants in terms of having low experience, experience less than three years or high [01:43:27] experience, experience uh more than three years. And you can see that in terms of wages, right? So um the the the high experienced individ uh uh applicants experience a slightly lower [01:43:40] reduction in a job that require high experience. um uh uh uh uh showed a lower response in uh in reduction wages but interestingly jobs [01:43:54] that required higher experience they show a much bigger decline in the time amenities time related amenities. So even the more experienced workers even though the wages are not they are not [01:44:07] not as expected but less than than the um uh less experienced workers but they suffer also on the non-monetary margin. [01:44:16] I think that's one thing that I think is uh uh it it's just uh uh um maybe specific to China being a developing country with a relatively weak enforcement in labor laws. Okay. Uh we [01:44:30] did a bunch of robustness checks. Uh I'm going to skip uh uh these discussions for the interest of time. Uh there's a uh then we look at uh the callback rates right still at the um uh job application [01:44:45] level. Um how likely will the uh the the application submitted to a um uh you know by applicant I submitted to job J located industry GJ, right? Uh [01:44:58] occupation GJ. how likely uh will the uh uh callback rate be affected uh depending on the large language model exposure. Okay, so this is one thing we found that overall callback rate does go [01:45:11] down and a higher human capital right if you have higher education or more experience right uh the uh the the does protect you in the reduction of callback rates. So if you have more experience and more education, the callback rates [01:45:26] reduction is is smaller than than the overall callback rate reduction. Okay, quick summary. Um in terms of the large language shock on the lab uh on labor demand we found that occupational [01:45:39] composition uh changes right that there's a contraction in the job postings job postings uh among more large language model exposure exposed occupations and there's structural shift to high human capital jobs and that is [01:45:53] consistent with what we found uh you know the in the US labor market right and uh for the China context the time related amenities seem to reduce more for high experienced jobs. Right? So the callback rates also decline in general [01:46:08] but less for high human capital jobs. So next question we want to ask is give only one minute. So uh uh what do what do the applicants uh search behavior? [01:46:17] How do the applicant search behavior change? Right? So we'll be looking at uh whether do they are they able to adjust to these shocks right this big shock that is um um uh hitting the Chinese labor market. So here the only thing we [01:46:31] I would like to explain that the larger engine model exposure at the at the uh applicant level is the um the the larger motor model exposure of the applicants pre-arch occupation. So if I if I was in [01:46:45] in occupation X right what is the large model exposure of occupation X and then I will be looking at outcome for example we'll be looking at whether I switch to apply for jobs in occupations that are [01:46:57] less exposed large model right so so uh I guess I'm really running out of time let me just uh uh say that summarize the results right so we see some um mitigation effort by the job applicants [01:47:12] but not enough to um overcome the demand the the changes in the in the demand. [01:47:18] Right? So uh demand side wage declines does not transmit to applicant wages but because if you the demand side wage declines and accurate applicants adjust their expected wage in the job they apply accordingly. They maybe they can [01:47:32] still find a job right but they are not adjusting uh in is sufficiently. So let me just mention uh last thing is that you know they do they really switch to large model uh uh um the less exposed [01:47:46] jobs uh there's some effect right there's some effect right the uh uh and they do leave um um uh job that are more exposed right um and and so on okay so [01:47:59] uh regional uh um regulation we say that we found that regional regulation does have some effect on uh mitigating um the wage negative wage effect uh by by the large model but not so much on the time [01:48:14] amenities and which is consistent with Chinese labor law mostly protect uh workers uh dispute workers in their disputes with their employers on terms in terms of wages but not so much on [01:48:26] non-wage benefits. So um anyway so this is uh what I have to say I'm sorry for running over time. Thank you. [01:48:44] Oh post is similar to the job posting vac similar 54% reduction. [01:49:14] Uh well, first off, thank you all for having me. Very excited to be here. ## Discussion (01:49:23 – 02:01:45) [01:49:23] >> Okay. Um very excited to be here. This is a great topic for a conference. So I'll be discussing the paper. I'm Lindsay Raymond from MIT and I kind of want to start with just a [01:49:36] road map. So I don't think I need to kind of summarize the paper. Um and my comments will primarily focus on three things. The identification and timing um some questions on measurement and then some changing composition in the [01:49:51] sampling that kind of make uh and how that affects how we interpret the results. [01:49:57] Um, and before getting into it, I think I want to say that this is a really important paper on a topic where there's a lot of questions and kind of little good evidence. So, I was very excited to see the paper. Um, and really think that [01:50:12] is particularly important to ask how these these these effects might vary across countries. So, very nice to read, not a US- ccentric paper. Um, and all of my comments here about, you know, baking the paper better. I think there's a very [01:50:26] incredible data set you guys have. Um and so kind of I'll I'll also make a couple of suggestions which of course you're free to ignore. [01:50:34] Um so starting with the identification strategy which is something that has been used elsewhere. Uh taking this idea of a post you know chat GBT shock interacted with some measure of LLM exposure. [01:50:48] Um, and first I think that chat GBT, which was released at the end of November in 2022, was never available in China. And so if I were going to think about how I would make this, you know, [01:51:02] this identification strategy specific to China, you say the earliest LM was a BYU test [01:51:16] launch which happened in March 2023. I would kind of use the model whatever whatever you think based on your knowledge of what the relevant shock would be. I think that's what you want to use to define the post indicator. [01:51:29] Um I also think that this type of identification strategy any sort of timevarying uh shock that's correlated with LLM exposure is going to be correlated with [01:51:43] this indicator and we are going to interpret as the effects of AI and this is a point other people who have used this identification strategy have you know tried to address um what was happening in China during [01:51:56] this time. So in December of 2022, China lifted internal mobility controls um that were due to COVID. So basically I I have like all of the in-person [01:52:14] activity that was shut down during COVID is now happening. Um in January of 2023, China opens its borders. You kind of have like international travel for the first time. Um and so that you know is a [01:52:28] very particular example of like this time varying effect that is probably negatively correlated or you know correlated with some way with LLM exposure that you would have very much in the back of the mind as we're thinking about this. [01:52:42] Um, and this kind of brings me >> So, I think again, um, I didn't I don't have in the back of my mind exactly the [01:52:56] LLM exposures for the China jobs. My I would expect that, you know, if it was similar to what was happening in the US, it's kind of the in-person less structure that is you know not based in [01:53:15] fact. It could go the could go the other way as well. Um and kind of when you look at many of the event studies this one is log wage log posted wage in the postings you kind of see effects start [01:53:27] to happen very early and to take us back to where we were in 2023. [01:53:34] So we were still kind of confused about what the ta technical capabilities of these models were. So um like for instance a paper kind of well-known paper called sparks of AGI came out uh [01:53:49] in October of 2023 that was massively cited and it basically said here are all the cool things that you can do with LLMs. And I think that is just just important to remember. it is people have [01:54:02] would have to be like very smart to be able to anticipatory adjust or adjust their production in real time in 2023 to the effects of LLMs. Um and so kind of when I think about like these graphs, I [01:54:16] think this um COVID shock is probably something you should try to keep in mind. Um and then also just you know again maybe redefining the postshock to be more specific to the China LLMs. [01:54:31] Um and kind of this also brings me to this this um other really important question they talk about which is time related amenities. And I think this is a great thing to focus on in the paper basically [01:54:46] how other characteristics of jobs change beyond just posted wage. And so that is a great thing that you guys should kind of really narrow and um focus more on. [01:54:55] Um the question I have here is in the posting this is kind of defined as you may have weekends off or something to the effect which is not a non-wage amenity. If the job posting says you're [01:55:09] going to be working or you will, you know, not be working Saturdays, that means your like effective hourly wage is going to go down um relative to another job where you only work 5 days a week. [01:55:21] And so this seems to be like very much affecting the kind of effective posted wage rather than being some sort of non-wage amenity. And so I think kind of first more verification on like what [01:55:36] exactly do these uh do these um posting words mean in terms of actual um do they really mean this job will give you a weekend off or not? And then also maybe [01:55:49] this is a this offers a way to kind of redefine the posted wage in terms of an effective hourly wage using kind of the imputed actual numbers of hours a week you work per work per week and try to [01:56:02] understand using that as an outcome what the LLM exposure measure does to that um kind of and then the kind of third graph which I liked which uh you didn't [01:56:16] really have time to talk is looking at on the platform call back rates which are basically some interaction maybe an interview maybe a phone call um maybe some sort of like positive interaction between the [01:56:30] applicant and the posting and what I would encourage you to do is kind of redefine callback rates using just interview call backs like that seems to be the first order important thing that we care about is interviews um because [01:56:44] it's sort of hard to like because This is an index of a bunch of different things. It's a little hard to interpret. [01:56:50] And then secondly, this kind of relates to some of the measurement questions. [01:56:57] The data set is constructed by sampling every year 200,000 job postings. And so their quantity measures, they can't measure quantity of job postings. They measure relative shares of an occupation [01:57:10] in the data. um and see so for call back rates these call back rates are of the composition of sample that they they select every year what is the share of callbacks now suppose there's just an increase in [01:57:24] the number of applications that happen on online platforms over time but a company has the same number of job slots then just mechanically the kind of probability of getting a call back is [01:57:37] going to go down over time and that's just because there's like a higher volume of applications that are happening on the platform. And this is just an example of the way that like the comp the changing composition of the data set make it a little bit interp [01:57:52] hard to interpret kind of what these effects mean. kind of similarly when we think about you know the labor demand measures suppose we take as a given that the introduction of LLMs leads to a decrease [01:58:07] in the quantity of postings for you know software engineers it's not clear that we would measure that as a change in the share of postings that are software engineering jobs on the platform because kind of the composition of jobs is [01:58:21] changing so if you know there's more uh low blue collar service jobs that are being posted on the P platform. [01:58:31] There's more other types of white collar work that are be being posted on the platform. The shares might even stay the same or go up. And so I would say that to kind of better interpret the outcomes here. I think we really need to [01:58:44] understand what is the composition of the data better and how it changes over time. [01:58:51] um I kind of already talked about this and here I think uh the authors say there's you know two mechanisms through how they they are thinking through these effects um [01:59:04] and in the appendix when they add industry by time w uh co controls kind of the wage post the posted wage coefficient does attenuate quite a bit which sort of raises questions about you know how to interpret these effects [01:59:19] Again, I might make that your main specification. Um, particularly because the composition of postings according to what industry they are is probably changing as well. And so kind of controlling for that particularly in the [01:59:33] data you have seems quite important. Um, I already talked about this. Um, yes. And so one thing that I would kind [01:59:45] of think about emphasizing is even though there's potentially this effect of, you know, how LLM's change labor demand, an even more direct thing is how [01:59:58] they change the hiring and screening behavior in labor markets. And so one thing that you could think about doing with this data is you know before firms are restructuring production. We know at [02:00:11] least in the US people started using LLM to write cover letters. There was tons increases in in the number of applications firms gets kind of the signaling behavior the signals firms relied on started to break down. And [02:00:25] that would be a really interesting question. You might be able to study with this data that you uh would not necessarily kind of need which you would not be able to study elsewhere. It might even happen before some of these second [02:00:38] order effects. Um so with that, thank you so much again this great paper. Very happy to be here and I will end. [02:00:51] >> Okay. Uh Hannah >> um the one question I have is um have [02:01:05] you thought about automation that perhaps you could control for the uh occupation's exposure to automation? The reason I ask is um that might uh you know China has made a lot of [02:01:19] progress and there's that might could be some correlation in the occupation exposure to both LM and automation. So you know controlling for that might help. Um I've also always wondered about [02:01:31] regional differences. uh how different would the same occupation be across different you know regions whether that I know you exploit the regional labor protection but even the job definition ## Q&A (02:01:45 – 02:08:44) [02:01:45] could be different and um >> okay um so thanks for thanks for great presentation great discussion um something that struck me here is I don't [02:02:00] think this is what we've seen in the US data at all okay so My understanding of Brett Chander and others in his work has been for the most part employment has gone up. Um it there's been a junior [02:02:15] people between ages 22 and 25 haven't been doing well if they're in AI exposed occupations. [02:02:21] But I I think in the US data and whether it's because of AI or not is sort of an open question, but in the US data everybody else has been doing great. And so I think this is a real contrast to [02:02:32] that. Um and um it might be because of AI, it might be because of other factors, but I think it's important to dig into that contrast like why is is the descriptive stats that you have from China, why are they so different from [02:02:46] what we've seen in the US? >> Yeah, I was going to ask the same question. I thought like I thought that's where we going to start the paper. you would start with look like in Sig Miller ital in every single paper that we looked at like things look like [02:03:00] you know AI does okay for the labor market like not too bad it doesn't look like past automation waves and I agree with I mean as far as I know I mean I thought Lindsey also was going to comment on that this is the first paper where I see that not being the case and I'm wondering whether is it the first [02:03:15] paper that looks at the sort of posting margin and the flows there and not at the expost outcomes at the firm level or just looking at sort of like know CPS, occupational level, employment and wages [02:03:29] like do you guys know like this is a is this the first paper that just looks at sort of like what plat what is happening in platforms? [02:03:38] >> Yeah. So the thing that always strikes me when we're thinking about these exposure measures that they happen at these occupation levels and that jobs are not occupations. They're they they are constructed around subsets of the [02:03:51] tasks that go on the occupations. And so, you know, Frey Osborne had this very scary, you know, wage or job loss figure that, you know, fell to like 9% when we took within occupation variation in tasks. So, [02:04:06] Arnadal is less well known, but once we looked at the fact that jobs across firms have different bundles of tasks, we see way different effects. And I'm I what I see is very promising here is that we're not looking just at [02:04:19] occupations and you can go into the text and say how broad is this job? What kinds of bundles of tasks? And so instead of just looking at exposure and some very broad occupation level, we could actually look for this job. What [02:04:34] is the exposure? I think that would be a massive shift in perspective and could help explain contradictory effects. Um, I also think that the age could be the age effects that we're seeing in the US could be confounded with the breadth of [02:04:48] the job. And so maybe it's a breadth of job issue, not so much an age of the person. [02:04:55] >> Yeah. I I wonder also given the differences that other people have pointed out with uh outside China um whether it theoretically it might just be driven by like business cycle effects. So like Chinese firms are more liquidity constrained. we have uh you [02:05:10] know essentially an economic downturn during the period in question and so I'm just going to say in in if I need to invest in AI related capital I just don't have the money to to bring in as many many workers and and per like that [02:05:23] seems theoretically consistent with with what with basically all the cuts of the data um I have no idea how you try to separate that from like general AI but I'm curious if you have ideas on that um So quick comment and like [02:05:39] >> yeah so I want to follow up on that can you think of placebo you know there might be some calcations where this shouldn't make effect you know I can think of maybe construction workers or and can you do that and show the effect doesn't exist [02:05:53] >> so uh thank you so much Lindsay and everyone for great comments I uh a couple things I should have mentioned that this website is mo you know one of the largest uh plat manufacturing jobs. I think some of them [02:06:14] are there but uh more of these you know backend kind of uh jobs that uh that uh will be uh happening in the manufacturing uh industries um construction jobs I don't think they this is the most uh prominent website [02:06:29] but that's just a background. Uh second thing uh the big difference in terms of the aggregate numbers uh in job numbers uh I think Kevin is right that China is in a pretty big uh macroeconomic [02:06:43] downturn whether the government probably says that or not but you know uh at least we know that three there were actually there are by in June of 2026 there are 320 million workers uh in gig [02:06:57] economy right so 300 million the magnitude is it's almost the entire US population 44% of the Chinese uh for the government. Now whether [02:07:14] that's a result of the downturn of the economy or whether it's a partly attributable to China's very rapid adoptation uh adoption of of AI in workplace uh that's [02:07:28] you know I mean we we try to control for the time fix effect but again there could be um occupation interacting with time macro trends that is not being captured by the time fixes effects alone so we are not giving a very crystal [02:07:42] clear answer what is causing this right but right now we are comparing occupations with more or less LM exposure uh to get some effects of the large LM exposure on the labor demand and and the worker behavior Christine's [02:07:57] point I think the next paper will be about how job descriptions are adapting to large language model the tasks included in the a new a new bundled uh set of tasks and [02:08:17] uh and I think uh you know we I I'm sure you have done some research on that already but we would like to see whether that's happening in the Chinese data and how whether that what happening in China in terms of redesign of jobs differs [02:08:29] from the US or other other economies but there are a lot of open questions for me to and my co to work on thank you so much for the comments >> okay thank uh 15 minute break and we'll come back