Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable. ======================================================================== # Artificial Intelligence and Labor Market Reallocation Authors: Hie Joo Ahn, Nicholas A. Carollo Discussant: Gad Levanon Video: https://www.youtube.com/watch?v=HSCV6-Vfc4U&t=13817s ## Talk (03:50:17 – 04:13:04) [03:50:17] >> Okay, thank you Lisa. So next up is Hijuan and she'll be talking about AI and labor market reallocation. [03:51:10] Thanks for including our paper in the program. Last but not least, the title is AI and labor market reallocation. [03:51:16] This is joint work with Nick Carlo um who's there to answer difficult questions and the user disclaimer applies. [03:51:40] So our focus is large language models effects on labor market reallocation and its functioning. So for that we construct unemployment rate, labor force flows, market tininess and various other labor market outcomes of workers with [03:51:54] different AI exposure and adoption and the purpose is to assess reallocation pressure you know likely arising from large language models and its effects on structural unemployment. So we essentially want to now cast large language models effects on the natural [03:52:07] rate of unemployment you know based on two approaches. One is a theoretical one that extends Jangman Robert's model of structural unemployment. The other one is statistical one that combines basian trend cycle decomposition and flows model of unemployment in order to [03:52:21] recover the natural rate by exposure AI exposure and adoption. [03:52:26] So we find the evidence for weaker hiring for the AI exposure workers and increased reallocation pressure in the aggregate labor market. Number one, unemployment rate rose higher among the workers highly exposed to AI high, you [03:52:40] know, with higher AI exposure and with a job of higher AI adoption than others because their job finding rate and job switch rate declined, you know, more and persistently than others. However, their [03:52:53] activity switch rate you know within the same job rose more suggesting increased task reorganization among them which amounts to you know increased dispersion in the individual level job finding [03:53:05] propensity and you know likely due to AI you know suggest increased reallocation pressure in the aggregate labor market. [03:53:12] So we estimate that large language models likely raise the natural rate by about a tenth or two ten percentage point even though you know these estimates are subject to you know large uncertainty. So we interpret this result as suggesting upside risk to the natural [03:53:27] rate of unemployment likely due to LLMs and the contribution of this paper is to construct labor force flows by AI exposure and adoption along with other labor market outcomes you know to assess large language models effects on the [03:53:42] aggregate labor market and its functioning. So here's the plan of the talk. I'm going to first talk about the measurement and then we are going to look at the stocks and flows together and think about the implications for the natural rate of unemployment. [03:53:55] So measurement so we are going to use the AI exposure and adoption scores in compiled based on open AI and liccast respectively and assign these scores to individuals in [03:54:08] the CPS to construct the stops and flows and also you know to job vacancy data from jolts after constructing occupation level vacancies based on lycast data. [03:54:19] Hence we can recover market tightness by AI exposure adoption. So let me fix the idea and the concept of AI exposure and adoption. [03:54:28] So AI exposure captures AI's feasibility of conducting human task. So we use Elon's measure compiled based on open AI. Of course we did the robustness checks with alternative measures and this measure vary by occupation hence to [03:54:43] reflect you know workers characteristics. So we interpret this as reflecting you know supply side effects of AI. AI adoption however captures actual labor demand you know for AI integrated roles for production. So we [03:54:58] measure AI adoption with a light cast job postings data with a share of job postings within an industry that mentions AI or machine you know learning skill tags and this measure vary by industry and over time so can be matched [03:55:12] with the CPS micro data. Of course this measure is observed with the measurement errors that we are all interested in. So we actually look at the correlation of this measure with the census bureau's BTOS you know adoption at the industry [03:55:26] level and we find out that these are you know positively and significantly correlated and also we consider varying coverage of AI and ML skills in lycast you know for robustness checks. [03:55:40] So actually a important player in this you know analysis is to how we categorize you know different exposure and adoption. Okay. So for AI exposure you know we categorize we consider four [03:55:52] groups based on the scores. So exposure itself is a fraction concept. So we consider 0 to 25% 25 to 50 50 and 75% and 75 to 100%. So note that this is not [03:56:05] the cortile concept. For example the labor for share of the top exposure group is only 4%. Okay. And for the adoption we consider 90% plus percentile [03:56:17] and 50 to 90 25 to 50 and 0 to 25 you know percentile which is cortile. So the reason why we consider this type of categorization is to capture the right tail behavior of the exposure and [03:56:30] adoption which is important. Lastly we consider the join intensity of AI exposure adoption. So we consider the cutoff of above and below mean of each measure. In this way we can consider both supply and demand side of the AI [03:56:45] intensity. So we consider four groups high exposure high adoption low exposure low adoption and high exposure low adoption low exposure high adoption. [03:56:54] Okay. So let's take a look at the occupation by AI exposure. So top exposure group u 4% of the labor force include software developers, computer programmers and accounting and auditing [03:57:08] clerks. Okay. And the mid high group includes secretaries and administrative staff. And mid low groups 40% of the labor force includes nurses, managers and teachers. And the low exposure group [03:57:22] one-third of the labor force includes janitors, laborers and nurses and cooks. [03:57:29] Let's take a look at the joint categorization. So high exposure high adoption group. One quarter of the labor force includes software developers in computer service industries because we consider occupation for exposure [03:57:43] industry for adoption. Okay. And lawyers and judges in the legal service industry. So low exposure and low adoption which is a gray box 41% of the labor force includes laborers in construction industries and waiters and [03:57:57] waitresses in restaurant industry. So let's take a look at the off diagonal elements. High exposure low adoption group one-third of the labor force includes managers in the construction industries and managers and food service [03:58:11] industries and low exposure high adoption 6% of the labor force includes assemblers in motor vehicle industries and retail salesperson in the computer stores. So you can see that the occupation itself has low exposure but [03:58:23] industries you know have been adopting um you know um AI quite a bit. Okay. So the last step of the measurement is imputing missing AI scores for CPS micro [03:58:36] data. So unemployed individuals report in industry and occupation. So we use that information to match with with the AI scores. But unemployed individuals they report previously in an industry [03:58:48] and occupation but 10% of unemployed individuals don't report previously don't have previous industry occupation or they report military occupation which [03:58:59] lack AI scores. So we follow you know the following step to impute the AI scores for these people. So number one, we first assign legged AI scores if the person was employed in the previous [03:59:13] month still missing. We use the mean of the consecutive rotation groups AI scores. [03:59:20] So we essentially want to use the you know adjacent months information to impute the missing AI scores. Still missing we use we go to step three which is we use the average across all rotation groups. So essentially the [03:59:33] intuition is that we use the second half rotation groups information to impute the first and you know the other way around as well. Okay, still missing you know so so far you know now we only have 5% you know of the [03:59:47] you know missing um individuals with missing AI scores. So as the last step still missing we assign the mean scores of the demographics and education and other you know observer characteristics of those who only work to one employment [04:00:00] spell because these workers are likely to have really weak job attachment. So in our analysis we use step three as the baseline you know excluding you know individuals with missing AI scores but we do the step we use step four as a [04:00:15] robustness checks and then we find out that the two um you know the two treatment in resulting similar um results okay let's enjoy the result [04:00:28] so you are looking at the unemployment rate by AI exposure 12 months moving average to smooth out you seasonality and the measurement errors. [04:00:38] So the top line is the bottom exposure group and you can see that strong cyclicality there and the yellow line is top exposure groups unemployment rate. [04:00:49] Okay. So the vertical line dash line is the end of the November 2022 you know when the LLMs were introduced. Okay. So we see that the top exposure group's unemployment rate rose by 1.3 percentage [04:01:02] point you know through July 2026 and 1.7 percentage point November last year. Whereas the bottom exposure groups unemployment rate rose by 410 percentage point so far this year and 7/10 [04:01:16] percentage point you know at the end of the last year and middle groups you know unemployment rate changes you know go somewhere between 7/10 and 4/10 percentage point. So considering that this top exposure group's unemployment [04:01:29] rate is least cyclically sensitive, the fact that it's unemployment rose more than the bottom group actually means that it reflects more than cyclic weakening and hence likely to reflect you know large language models effects. [04:01:43] So you see similar uh and the you know the driver of the rise in unemployment rate is actually the job finding rate. [04:01:51] So again that you see that the yellow line the top exposures group job finding rate declined persistently and it has stayed lower than other groups. Okay. If you take a look at the separation rate [04:02:04] um their job loss rate separation rate you know edged up a little but differential in the the differences across the groups is actually relatively muted. So we consider the job finding rate was the main margin of the rise in [04:02:18] unemployment rate. job to job transition rate. So yellow line is again the top exposure groups job to job transition rate. You see that it has declined more than other groups you know and then stayed lower than [04:02:30] other groups. Okay. And it's important to know that job to job transition is another you know margin of the hiring and hence reflects you know weaker demand you know for these high exposure workers. [04:02:42] However the panel shows activity switch rate. So CPS also asked whether your activity changed if you didn't change your job at the same job. You see that the yellow line top exposures group you know activity change activity you know [04:02:57] change rate you know increased more than other groups. So which actually suggest increase task reorganization you know among this top exposure group. [04:03:06] So if you look at the um um the different disagregation by adoption, you see the similar pattern again the yellow line is the top deskile um adoption groups and unemployment rate. It rose by [04:03:20] 1.2 percentage point you know whereas the bottom you know cortile adoption group you know unemployment rate rose by 410 percentage point and again the main margin of adjustment was the job finding rate. [04:03:33] So let's take a look at the joint categorization. [04:03:37] So yellow line again is the unemployment rate of the high exposure high adoption group and their unemployment rate rose by about 1 percentage point you know since the introduction of the large language model. The black dashed line is [04:03:51] it's uh the unemployment rate of the high exposure and low adoption um low adoption group. Their unemployment rate rose by 510 you know a half ten percentage point. You see that the gap actually reflects the effect of the [04:04:03] adoption. Again the main margin of the main driver was the job finding rate. [04:04:09] You see that the yellow line the job finding rate of the high exposure high adoption group declined more and then stayed at the lower level than other groups. [04:04:19] So you might wonder wait a minute you know we looked at the previous you know work also looked at you know the AI's effects on the unemployment rate but their conclusion has been the effect is small to negligible but why do we see [04:04:33] the different result because we focus on the right tail of the AI exposure adoption and their interactions you know the margin that has not been you know well understood in the literature [04:04:45] so what have we learned So we have learned that large language models likely weakened you know demand for workers exposed to new technology and raised the task reorganization among them suggesting increased disparities in [04:04:59] job finding rates and buildup of the reallocation pressure. [04:05:05] Okay. But wait a minute there are also confounding factors you know now you may have alternative explanations that can screw with all these empirical findings. [04:05:14] Let's you know go one by one. So um so there can be a lot of you know convoluting factors you know that's correlated with you know large language models you know adoption or AI adoption. [04:05:26] So statistically you know address this concern you know we considered the linear probability model of the job findings you know considering very detailed you know fixed effects. So for example we considered you know 3D industry interacted with the time fixed [04:05:39] effects and then we looked we tried to identify the effects of the interaction the effects of the AI exposure and adoption interaction and then we found out the statistic significant negative effects you know from 2023 whether the [04:05:54] effects were not statistic significant you know before that. So something was happening and it seems like an interaction has been important and we also you know considered the adop adoption potential IBS and you know alternative fixed effects that capture [04:06:09] the heterogeneous effects of the monetary policy but we see you know we still get that you know similar result and now another convoluting factors could be unwinding of the tech overhiring in 2022. Okay, so in our [04:06:23] analysis, you know, we got rid of the tax sector and then we found out essentially the same result and the third factor could be the telework. So in our analysis regression analysis, we found out that telework actually raises you [04:06:37] know job finding probability whereas AI exposure and adoption lowers that. So certainly it captures different effects and also we looked at AI's effects on hours worked and wage growth and AI [04:06:51] tends to have positive effects on hours worked and wage growth but this upside effect attenuated after 2023 which is overall consistent with the you know weakened demand story that I have been [04:07:04] we have been claiming. Okay, but the question is but is does it have a pervasive effect? So here I would like to show you the geography of AI exposure and adoption. So you're looking [04:07:16] at the metro areas with a greater AI exposure tend to be tend to also exhibit higher AI adoption. Okay. So if there's a you know negative interaction effects you know the effects can be pervasive and then the effects you know reach [04:07:31] beyond you know young entrance to the labor force. [04:07:34] So we looked at you know changes in the compositional you know uh we looked at compositional shift across this um you know exposure and adoption group among new entrance you know with prime age and permanent job losers who reach at the [04:07:49] retirement ages. And then we found out that within this you know unemployed group you know there has been increased you know um the composition shift toward the high exposure high adoption group as you can see from you know red bars you [04:08:03] know so the red bar captures the average you know from 2025 and the blue bars capture the average in 2022 before LM's introduction okay so now let's think about implications [04:08:18] for structural unemployment So here you know for policy to explore policy implication you know we want to know how this AI diffusion affected structure unemployment. So to assess you [04:08:32] know that possibility you know we rely on a theoretical model of structural unemployment. So we actually use a Jangman rapper's you know model of structural unemployment that relates structural unemployment as a function of the mismatch between job seekers and [04:08:45] vacancies. So essentially the intuition is that you know if you see the you know dispersion in the job finding propensity across individuals that actually signals that some workers are having hard time to the sectors with the favorable [04:08:58] reemployment prospect. Okay. So essentially want to estimate how much large language models you know contributed to the changes in the dispersion in the you know job finding propensity controlling for all the other [04:09:10] cyclical you know economic conditions. So here's um you know our model that we derived. So SUT is a structural unemployment and the blue term is a cross-sectional dispersion of the individual level job finding propensity. [04:09:25] Okay. And UT is the unemployment rate and FT is the job finding rate. Okay. So we recover the blue term that is the cross-sectional standard deviation of the individual job finding propensity [04:09:38] using the fitted value of the linear probability model of the job finding um UE transitions um you know so that's what we call as individual level job finding propensity [04:09:50] or higher ability okay and then we also calculate the AI per one you know to assess you know AI exposure adoptions you know effects and let's take a look at the picture. So panel A shows the [04:10:03] mean higher ability by AI exposure. So you can see that the top line is the you know job finding propensity of the the bottom exposure group and the um you know orange dash line is the you know [04:10:16] the job finding propensity of the top exposure group. You see that you know the top exposures groups you know job finding pro propensity you know declined and persistently likely due to AI and that actually results in increased [04:10:30] dispersion in the job finding propensity. So the solid line is with AI factors included and the dashed line is without AI factors. Okay. And this cross-sectional dispersion increase by [04:10:43] 1.1 percentage point. And then we actually rely on this structural model to find the implications for you know changing the structural unemployment. So assuming 4% unemployment rate and then you know 30% of the job finding rate and [04:10:57] the matching elasticity of.3. So 1.1 percentage point increase in the dispersion implies two t percentage point increase in structural unemployment and one if we assume in alternative vacancies you know which is [04:11:10] alpha.5 that means that you know one10 percentage point increase in structural unemployment. So lastly okay okay so time's up so let me just consume one minute quickly you know go over the last slide. [04:11:25] So lastly we consider um we directly estimate ustar using the flows you know measures that we constructed so far. So here the concept of the ustar is the um the unemployment rate that is composed [04:11:38] only of the trend components of the separation and job finding rate. Okay. [04:11:42] So we actually estimate this ustar for you know different groups of AI exposure and adoption and we aggregate this sub using the label for share and the merit of this analysis is that we can identify [04:11:55] durable large language models effects controlling for pre-trend LLM trends and the cycles of different groups. So we can do the counterfactual analysis. So our model says that you know so that's at the figure the black line is the [04:12:09] changes in the aggregate natural rate which you know according to a model has risen by about two ten percentage point. [04:12:16] So if we shut down you know contribution of the high exposure high adoption group the aggregate natural rate decreases by one/10en percentage point. If we shut down further of the high exposure low adoption group then the aggregate [04:12:28] natural rate is reduced by five basis point. So all told you know high exposure group likely to have contributed to rise in the to the rise in natural rate by about 15 basis point [04:12:41] but this estimates are within the 68% of the posterior intervals. So not that statistic significant hence that we interpret this result as suggesting upside risk to the natural rate of [04:12:54] unemployment. Okay. So this is it. Thank you so much. ## Discussion — Gad Levanon (04:13:04 – 04:28:34) *Shared discussion for Session 3: Job Openings, Occupations, and Technological Dynamism; the discussant covers every paper in the session.* [04:13:04] Okay, thank you Hiju. And our discussant for this session is Gad Lebanon. [04:13:51] Thank you. Hi everyone. H it's a a pleasure to be here. Um I am very happy to uh to been assigned those two papers. I'm I'm a [04:14:03] user more of than a creator of data and I work a lot with job postings and I uh also work a lot on AI and its impact. So [04:14:15] I really learned a lot from those uh two papers. Um I will briefly talk I I will um everything here is brief but I will uh [04:14:27] give some big picture um comments on the the two papers um and I will sprinkle a little uh big and bold ideas here and there uh to to [04:14:40] add to the conversation. Um just one general uh comment also listening to the discussion here uh in in recent hours. I I think [04:14:54] what we've learned or listened to today mostly is how to use alternative data to improve uh official data. But I think the the real opportunity of alternative data is actually to use the alternative [04:15:08] data itself to do analysis. And I think researchers should spend more of their time improving those alternative measures to make them uh better and more [04:15:21] useful. So I'm very happy that Lisa and her colleagues decided to do that. Um, one uh clarifying comment. Lisa kept talking about the burning glass data and [04:15:36] I'm the co-founder of the Burning Glass Institute. So, I just want to tell you that it's completely two different organizations. [04:15:44] It's not our data. The only reason they are called the same is because the CEO of Burning Glass Technologies that later became Litecast is Matt Seagelman and he's also the president of the Burning [04:15:57] Glass Institute. But we are completely different organization. It's not our data. If anything, there is a mutual feeling of hostility and animosity between the organization. [04:16:10] erh they indeed sued us to want us to change our name. So um so you know my bias is going in um so I'll start with [04:16:23] with Lisa's paper which I think is I didn't read all papers but I think it's probably the best paper we have on the biases of online uh job openings. Um [04:16:38] uh and I'm not going to go go over everything they found, but um I think you you can think of it as there is the raw postings. [04:16:51] There are adjusted postings that you can adjust yourself using publicly available data like the jolts breakdowns by industry and and region. And then there is the adjustment that you can make with [04:17:05] Lisa's method. Um and uh I think uh this is a a a better it's clearly better um than raw postings and that's what I I [04:17:19] think the the author showed the how how big the the the difference when you adjust um using their their method. I I think one thing I would recommend for the paper is [04:17:33] also to try since a lot of people are able to adjust themselves using the jolts publicly available data. How much how different is your your results your adjusted [04:17:46] results would be compared to adjusting it with publicly available data. I believe it will be better because you have information [04:17:58] on variation within industry that you can't just do with jobs. [04:18:05] Um yeah, so that that is one uh suggestion to to improve the data or or show that it is more useful. [04:18:15] Um here is what I So I'm very proud of my hair. I in my age I have a full set no hair loss [04:18:29] and very little white hair. But uh the white hair that I do have is trying to because of trying to make sense of time [04:18:40] series of of job postings. And uh I I think um most users of uh of postings are looking at changes over [04:18:54] time as opposed to just using level. So if let's say QCW in terms of quality is at like 97 and CES is at 90 [04:19:07] I think uh levels of posting are at 70 and time series of posting are like 13 14. So that's I think where a lot of the [04:19:20] work I think needs to happen. So I encourage um I encourage you to use it. [04:19:26] H I think this method can help but uh to fully fix it I think uh there's a lot of un um glamorous work in dealing with things like uh uh job posting coming and [04:19:40] going scraping from other job boards all kinds of a lot of non-economic reasons so I think calling for the community h to um [04:19:52] to to do uh that I think that's there is the great potential when it comes to posting is improving the the time series. [04:20:02] Um how can it help the rest of us? I think you are producing a set of weights uh if we want to better improve our uh postings. If you publish those weights [04:20:15] or better if BLS publishes those weights uh and takes over uh this I think it will be a good public service. In fact, I think BLS should take all have their [04:20:28] own postings data set and uh have another program. I'm sure I'm not the first one who thinks about it, but big and bold ideas that is one. Um and I think the the method is bigger [04:20:42] than this paper. I think any public any any private data that has an employer identified you can and I'm curious did like is any other data set ADP or or [04:20:55] anything are they using something like that probably not so I think there is a lot of opportunity here to check the biases of many other databases using your your method including our own burning glass [04:21:09] institute data and be happy to to try it on So that's uh the first paper. I'm already behind. Um so let me get to the other paper. Um and give you some context. [04:21:24] So first of all uh there is a big debate. Has a AI already displaced jobs? I think the answer is yes. The big question is how much? I think it's above 100,000 and below 5 [04:21:38] million. How did I get to those numbers? I I made them up but I think 95% of the people would agree that that's the number of displacement that already [04:21:49] happened but it's a big uh range and uh 100,000 is very different than 5 million. So we need to know better. Um I think no no studies found aggregate [04:22:02] effect. Um some found for young workers h in most exposed jobs. uh but uh this paper um did find aggregate effect. So this in terms of I think [04:22:16] you're holding me back upside down. I have five minutes not two. Um so so um and I think well I am in trouble now. uh so let me think what I'm going to the [04:22:30] point here that there are reasons anything that based on exposure measures I think it's uh the exposure measures that are lundu and felton they have very little to do with job [04:22:44] displacement so any paper that uses those measures is um will it will be a miracle to find a a significant impact and that's where the the paper comes in [04:22:57] I I think they it's a high uphill climb to to get significant results with using this exposure measures. But the the big idea here is that yeah, let's add to the [04:23:10] exposure measure a um an adoption measure because even if the technology there if the employers don't make an effort to adopt the technology, you won't see labor [04:23:23] market impacts. Uh so I think that that is a a good idea. Um as I mentioned before I I question the exposure measure. I've worked it to myself. I think it's [04:23:36] it's such a tiny part of what how AI is impacting um the labor market. So using it as the sole proxy is is is [04:23:48] a non-starter in my opinion. But adding to the to adoption measure I I think is in the step in the right direction. I think the adoption measure essentially as it's built now is more of [04:24:02] a function of the occupation distribution of the of each industry. So I I think you could find better adoption measures. But when I looked at the [04:24:14] results I kind of tend to believe them. erh the the ones at the top um are indeed I I'll finish in in two minutes so forgive me we don't have any other presentation after that right so we can [04:24:28] go for a long time um so uh um the ones at the top I really do believe that are in the top and the one at the bottom are are at the bottom I think for good reasons I [04:24:44] think there are ways to improve the the adoption um adoption measure and I am happy to chat with the authors afterwards on how to do that. Um what do I make of the results? So first just I [04:24:59] think the paper needs a a a clear table that shows all the specifications for all the outcome variables and the results. Otherwise it's hard to to tell um what's really happening here. But I [04:25:13] do believe that adoption helps when you interact it with the exposure measure. [04:25:20] Um I think taking it to CPS is the right move because you can then see flows and not just talks and it seems like the results are stronger for for [04:25:32] um for flows. I I would strongly recommend using this also on the ACS and the OEWS. [04:25:40] They are much larger. The OES has its issue in time series, but once there are ways to get around it and and it's extremely strong and well measured. Um, but at the end of the day, I think the [04:25:54] results are only as good as the exposure and adoption measures behind them. And I think that's not just for this paper, but for the community at large, it is you are bound to uh to fail if that's [04:26:07] like your exposure measure. That's how you measure the impact of AI. Um well, it's up to you, but I have four big and bold ideas. If you want to give me two more minutes, [04:26:21] one minute. Okay. So, um I thought like because I was asked by the the organizer to also think go beyond those two papers and think about other things. So I just [04:26:32] want to give you a quick look at um how profiles which are LinkedIn profiles and postings can do for studying AI things that you cannot do in your uh with [04:26:47] government official data. So for example use cases in postings uh there are a lot of information in job postings not just about the candidate but about what the project that they are going to work on. [04:27:02] We we created an AI use cases um from postings and career histories. Uh hundreds of thousands of uh of AI use cases like build a chatbot for customer [04:27:14] service um build an AI tutor things like that. Now you could uh once you have this use case you can connect the use case uh to the impacted occupation like [04:27:28] in the example of customer service chatbot the impacted occupation is a a customer service worker. So you can use that to create a an kind of an an a new [04:27:41] kind of adoption or deployment index that is based not on theory but on actual actions of that employers are are taking. Uh you can also uh build like if [04:27:53] you now I asked the BLS map me the all the jobs that were created because of AI. There is no way that that you can do that. But with for example postings and [04:28:06] career and LinkedIn you can do that. You can identify all AI companies. you can identify AR roles in nonAI companies and I we started working on that and I I [04:28:20] think you'll be able to have like a a number that is quite believable of the number of jobs that would not have existed if AI didn't exist. So that's another example that uh where you can ## Q&A (04:28:34 – 04:37:59) *Shared floor discussion for Session 3: Job Openings, Occupations, and Technological Dynamism.* [04:28:34] use it. Um and let me pause here. I very gen thank you for your generous time counting and uh thank you >> and thank you Gad for that nice uh [04:28:54] discussion. So we did uh and do have some time for for questions. We're running a bit behind, but uh we've had some great uh presentations and just wanted to leave it open if anybody has [04:29:07] any questions or comments. And I see there's some microphones floating. There you go. [04:29:14] >> Yeah. Uh I guess I introduced myself on recovery. Um so just following up on discussion at the end. Um the adoption measure. Um so so with the exposure [04:29:27] measure I can see clearly you know this is it's a notion of sub substitution right with the adoption measure a company posts uh you know requirements for any ice I'm not so sure this is about substitution [04:29:42] and not compliment right and and so I was even more surprised to see then that if I read this correctly we saw I guess high high was the biggest effect I could have easily seen it's high low that that [04:29:57] um this is the biggest effect. Anyway, I'm not so sure that just from the naive sort of outside view that your adoption measure is really about substitution which I suppose you wanted to uh measure [04:30:10] but anyway that's my question. Thanks Mike institute. Um so having done [04:30:32] some work in AI and looking at onet data and I know we have a job posting data. [04:30:38] Um, I think there's a limit in terms of what we can learn just from existing databases. And one of the things I would encourage us to think about is is there a need to follow up Gad's comment on use cases to do a lot more qualitative research um sort of in real time to [04:30:53] figure out what's going on in industry right now. Um, so that I think the databases themselves are slow to catch up like ONET is slow. Um and I think that uh some sort of vehicle by which we can get data perhaps through the job [04:31:06] post and talking to incumbents and occupations or industry experts in terms of what they're doing could be a very useful way of sort of approaching this. [04:31:28] Anybody else? Okay, looks like not >> Oh, sorry. [04:31:39] >> Thank you. John Paul for city. Just a question for Lisa about your nuisance parameter. You're talking in part about firms who appear in the Burning Glass, but then you don't see them actually having vacancies. So, there are firms [04:31:53] who will just go fishing in the labor market and they're always posting just to see what comes up. So I guess one way you observe that in the data is you're looking for firms who are always [04:32:05] posting and and never hiring very rarely hiring. So have you looked at that kind of pattern? Can that help you understand your nuisance parameter and maybe do something uh additionally to maybe take [04:32:18] those firms out of the the burning glass data set? [04:32:34] Yeah, thank you. I think that's a great question. Um, yes. So, these are firms who um they have postings. They didn't have hires and they also don't have unfilled vacancies according to the Joel's definition and Jolt's definition has these requirements like active [04:32:48] recruiting. Um, and so it's I don't think it's that likely that they're the the phantom vacancies, but I think your other question is about the firms that just always have a posting all the time. [04:33:00] And so that's really what's going and so those are what what the white cast and glass people do is just every now and then they call it a new posting. Um, and so so like after 60 days they the same [04:33:13] thing's been up. So like fast food they just always have these openings up, right? Um, and uh so the they recycle it. We did try our best to look at the rates of that the rates of that across industries for instance and it doesn't [04:33:28] really match the patterns. Um but I don't actually uh for instance like food and accommodation services it doesn't have a particularly high nuisance factor. Um but I haven't actually crossed that with size to see if it's small accommodation look at that. Thank [04:33:43] you. PG, did you? [04:33:57] >> Um, thank you so much for the great questions and the um support of the discussion. Um really appreciate you know all of um your comments and also questions um raised by the audience. So [04:34:11] um let me address you know a few um issues that the discussion raised. the first of all you know whether this AI exposure adoption that we currently have is the best thing certainly I think it's an open research you know area and you [04:34:24] know we are actually interested in you know exploring the novel cultures so that you know so in that sense your comment about maybe perhaps we can get comparisons about you know what what AI is really doing for the labor market outcomes based on use cases and also you [04:34:38] know as Michael said you know the qualitative research on the measuring um you know the AI is large policy effects you know based on the you know the real time data well you know there certainly you know open um you know research [04:34:52] question and you know we're very interested in you know pursuing um that route as well um but about the issue that whether this AI adoption is you know the substitute or compliment and [04:35:04] why does it really have destruction effect it's actually open question it's also an empirical question while we claim in our papers that AI adoption represents you know demand for AI you [04:35:18] know using role but we didn't you know make any assumption about whether it's a compliment or substitute you know so but empirically we found out that workers who are highly a exposed AI and adoption and also workers you know who you know [04:35:32] whose job you know has is exposed to high AI adoption their unemployment goes more and they have experienced a lower and you know job rate and also they experienced you know uh larger decline in the job buying rate. So it's an [04:35:46] empirical question what's really doing and certainly it's also a you know like research question as well and they comment about you know what AI adoption is really measuring um so it's actually a pretty good question they will get the [04:36:01] same adoption you know like orders or similar coariation or the magnitude if we aggregate the exposure measures using the occupation and employment weight we'll get the same thing or not that's [04:36:14] question and um but you know but you know in this paper we actually show that there's an interesting margin that has not been considered in the literature. [04:36:24] So for example you know there are managers you know who who are in the industry with the higher or lower adoption. So for example, there are managers in the construction industries and managers in the computer services industry and they seem to have you know [04:36:38] different types of the labor demand jobs and same for teachers but teachers are low exposure group but there are teachers who are in K12 schools and teachers in college and universities and they seem to have different labor market outcomes. So in that sense our claim [04:36:52] again is that the what is captured the interaction between exposure and adoption can be you know something important predictor of the you know labor market uh outcomes. Um so Nick do you have something [04:37:06] okay thank you so much you learned a lot. [04:37:09] >> Okay let's thank our authors uh presenters and discuss them. [04:37:17] Okay. Uh, so if you signed up for dinner, we have an announcement. So the the location is the Sapphire Room and it's on the penthouse level. So if you go out these doors and to the right, take the elevator to the penthouse level. Um, and it's at the end of the [04:37:32] hall is where you'll find the restaurant. Um, and that starts about 6:00 where the reception time. [04:37:56] 12 seconds.