Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=

# The Household Impact of Generative AI: Evidence from Internet Browsing Behavior

Authors: Michael Blank, Gregor Schubert, Miao Ben Zhang  
Discussant: Martin Beraja  
Video: https://www.youtube.com/watch?v=VdvT0JzwMHU&t=23038s


## Talk (06:23:58 – 06:47:54)

[06:24:05] browsing behavior. So, please. >> Right. Uh, awesome. Thank you very much for putting us on the program. It's a pleasure to be here. Um, I am Michael Blank for those of you I haven't met.  
[06:24:17] Uh, I'm an AP in finance here at Stanford. And this is joint work with Gregor Schubert who is at UCLA and Ben Zang who's at USC and also in the audience.  
[06:24:28] So uh there's been you know a lot of great work and we're still kind of learning a lot about kind of the adoption of generative AI tools in the workplace and by firms. Uh a question that we think though is kind of equally interesting and that we think is  
[06:24:41] relatively underappreciated so far is to what extent how uh households have been adopting generative AI for their own personal and household uh use. Uh why do we think this predominantly two reasons.  
[06:24:54] The first is that we think a fair read of the evidence suggests that if anything, generative AI so far has been predominantly adopted or a little bit more adopted by uh households for use outside of the workplace. So this is  
[06:25:07] suggested uh through you know a few different sources. Uh for one thing uh the paper by OpenAI that looked at chat logs and kind of classified them according to what uh the purpose of a given chat was suggests that you know a  
[06:25:21] majority of chats that is if anything been growing over time are based off of uh things in uh you know personal uh tasks uh rather than uh things for work.  
[06:25:33] And then there are also uh surveys that uh suggest that again people have predominantly been coming into contact and starting to use generative AI tools uh for use outside of the workplace rather than in their jobs. Uh and then in addition we think that you know with  
[06:25:47] households there aren't really the organizational and coordination frictions to adopting generative AI and beginning to kind of uh you know transform one's life using it or potentially transform one's life using it. And so we think that it's  
[06:26:01] potentially kind of maybe an early indicator or uh you know uh impact of generative AI uh that kind of is more profound than just kind of uh adoption rates. It can kind of change how people are allocating their times in potentially very important ways that  
[06:26:15] could be kind of missing from the GDP statistics. All right. So the main question that we're going to try to answer in this paper is how does generative AI affect households at home and in doing so create value?  
[06:26:27] And to unpack what I mean by this, this is not going to just be a matter of looking at what we observe households doing when they're using generative AI at home. Rather, when I say how does generative AI adoption affect households, I'm going to be trying to go  
[06:26:41] after a more causal treatment effect. Meaning that I want to be able to measure the change in a household standard of living and time allocations after adopting generative AI relative to a counterfactual world in which they had not adopted it.  
[06:26:55] And then when I talk about how does generative AI create value, what I really am trying to get after is what types of economic activities does the adoption of generative AI crowd in. And there are a few kind of different possibilities that of what activities  
[06:27:10] could be crowded in just to go through a couple of extreme examples. One possibility is that even if generative AI kind of helps people do productive tasks more efficiently, they might not want to spend a ton more time doing those productive tasks. If it's  
[06:27:24] something like completing their tax returns, for example, they might just want to get it done as quickly uh as they can and then use the excess time savings that generative AI gives them to engage more in the leisure activities that actually give them kind of  
[06:27:37] intrinsic uh welfare value. But it's also the case that, you know, generative AI might give some people perhaps with less formal education uh and technical backgrounds more opportunities to invest in their human  
[06:27:50] capital if generative AI tools help democratize learning new skills uh finding new jobs or opportunities in the labor market uh and so on. And so these are, you know, examples of very different types of activities that  
[06:28:02] generative AI uh once adopted at home could be crowding in. And we want to try to uh kind of understand if uh you know what exactly has being uh crowded in and what implications it could have for welfare. All right. And kind of  
[06:28:16] something that's going to be in the background of these questions is the distributional implications depending on whether generative AI adoption at home uh exhibits kind of a gradient with respect to income or age or education levels. And so to the extent that's  
[06:28:30] going to be possible in our data, we're going to make sure to keep track of exactly which types of households are adopting generative AI, which is going to kind of mediate how we then are going to think about these treatment effects.  
[06:28:40] Okay. So what do we do in this paper? So the kind of core of our paper is going to be a novel source of household time allocation data. And in particular, we're going to use data from comscore, which is a digital marketing company  
[06:28:53] that uh gives us a panel tracking 200,000 households who we can see the second by- second timestamped internet browsing behavior of. So, as I'll explain in a bit more detail uh in a couple of slides, com score incentivizes  
[06:29:06] people to put a tracker on their home computer and their cell phone and we can then see all of the online uh browsing activity that that household engages in.  
[06:29:16] Uh so with this data, we're going to first off measure who exactly in our panel has apparently adopted chat GPT as well as who based off of their pre-release of chat GPT browsing um  
[06:29:28] allocations seems to be most xanti exposed or find it most xanti useful to adopt chat tpt once it actually comes out. And then we're also going to uh look at exactly how people allocate their browsing time between different  
[06:29:42] types of activities. And in particular, we're going to focus on breaking down website uh visits into uh activities that we could classify as productive versus classifying as leisure in a household production sense.  
[06:29:55] Okay. So, using this data, we're going to have five main findings. The first is that there's been faster and more persistent adoption after the release of chat GPT of generative AI tools by younger and higher income households, which is consistent with survey uh  
[06:30:09] evidence. And more generally, there's been more adoption among the households who we can say have an X anti higher usefulness of using chat GPT. So the people for whom we're doing activities before it's released that would uh have  
[06:30:23] a a greater overlap with chat GPT are more likely to begin using these tools once they're made available. So this second finding is going to be helpful in that it's going to give us a source of variation in adoption that's going to be useful for trying to get at a a  
[06:30:37] plausibly causal treatment effect of adopting chat GPT. That's what these next two findings relate to. Uh in particular the third finding is that according to our baseline estimates adopting chat GPT leads households to spend significantly more time on leisure  
[06:30:52] related browsing activities and correspondingly a lower share of time on productive browsing activities.  
[06:30:58] Uh but this is going to have to be interpreted in line with our fourth finding which is that when we look at kind of highfrequency browsing windows to try to isolate what exactly people are using chat GPT in their homes to do we find that they are predominantly  
[06:31:12] utilizing chat GPT to perform what we classify as productive digital activities. So this is a bit kind of counterintuitive or hard to square. What do we make of the fact that people are doing more uh time on leisure activities overall even though the direct effect of  
[06:31:26] adopting chat GPT is seemingly uh related to productive activities. This is where our fifth and final part is going to come in. What we do in this part is we estimate a structural uh model of time allocation which when combined with estimates of the  
[06:31:40] elasticity of the angle curve between leisure and productive uh browsing activities, we're going to be able to infer that uh chat GBT adoption has led to a scaled efficiency gain of 75 to  
[06:31:53] 175% of productive uh or for productive activities rather. Why though does leisure time allocation go up when you adopt chat GPT? Well, according to an estimate of the angle curve, uh leisure  
[06:32:06] goods are uh time uh luxury goods, meaning that when somebody has more time overall that's freed up by having more productive uh uses uh more productive browsing activity, they're going to want to do more of the things that they  
[06:32:20] actually get intrinsic enjoyment out of and do less of the time necessity goods which are going to be productive goods according to our estimates.  
[06:32:27] Okay, so in the interest of time, why don't we go right into what we do starting with the data. So, as I said, our main data source is going to be the comscore desktop panel that we have access to right now from 2021 to 2024.  
[06:32:39] This is going to be a panel of around 200,000 uh machines, which I'll refer to interchangeably as households, which as I said are going to give us the exact timestamp and URLs of all browsing activities that people engage in on the machine that they have these trackers  
[06:32:53] installed on. Right? So, this is a ton of data. So we have to kind of aggregate it in a way that makes it kind of usable and kind of lends itself to our analyses. So the way in which we're going to do this is by aggregating the browsing information to the machine by  
[06:33:07] quarter by website type level where we're going to consider two uh particular classifications of sites. The first is going to be uh whether or not a site seems to be associated with completing a productive or leisure household task. And then the second  
[06:33:21] classification is whether the site seems to have overlap with chat GPT meaning that its functions can seemingly be completed uh with uh chat GBT rather than uh going to the website itself.  
[06:33:33] Okay. And then we are then going to also uh through these aggregated measures uh estimate the extensive margin of time allocation to a particular website category which is just going to be whether or not a household visits the site type in a given quarter as well as  
[06:33:48] intensive margin uh based measures in which we're going to be able to say for a given machine in a given quarter how much of their time when they're browsing is allocated to that particular uh type of website.  
[06:34:02] Right? And then very importantly for our purposes, uh, comscore also collects self-reported household characteristics for the people in the panel, which are going to include the households, uh, income, age, and MSA, as well as whether  
[06:34:15] or not the machine is predominantly used for work or home use. And we're only going to keep for the purposes of our analysis machines that are reported as being for home use.  
[06:34:26] Okay. And then there's also a lot more kind of data that comes with our sample that we haven't yet made much if any use of like smartphone based browsing, the text of people's uh search queries into Google uh online shopping outlays. A lot  
[06:34:40] of these things uh we are very eager to hear if people have uh suggestions for how we can begin incorporating some of this richer uh uh digital footprint type data into our analysis because we suspect that there's uh potentially a lot more that we can do to kind of uh  
[06:34:54] augment what we've done so far. Okay. So, uh, first off, I just want to kind of because this is a relatively new data source, uh, go over what the distribution of households in our panel is between age and income buckets. So,  
[06:35:07] not surprisingly, our panel is not exactly representative of the US population. I don't know about people in this audience, I would not want a tracker to be put on my computer. I don't quite know who the types of people who who would are, but um you know it  
[06:35:21] tends to be the case that our sample over represents uh households at the lower and the higher ends of the income spectrum. We think that the lower ends might be kind of college students. The upper ends might be people who are kind of, you know, relatively um welloff in terms of wealth and maybe spend more  
[06:35:35] time at at home. Uh and then we also underrepresent older households and tend to over represent younger households. So throughout we're going to always rewe our sample so that it replicates the ACS sample weights. But of course it's a caveat that that's only going to be um  
[06:35:50] you know okay to the extent that kind of in a given bucket the households that we observe are reasonably representative of the population.  
[06:35:58] Okay. So uh let's now get into exactly how we construct our measures. So first off to measure generative AI adoption we're going to do something very uh parsimmonious and simple. We're just going to focus on the extensive margin  
[06:36:11] adoption of chat GPT over time as a broader proxy for adopting generative AI. And so in particular, we're going to say that you are a chat GPT adopter if we ever can observe you visiting either  
[06:36:22] chat GPT or openai.com at some point in the past. Uh so to be clear, adoption turns on. It becomes a one when we first see a household visiting one of of these sites and it stays on for the duration  
[06:36:35] of the panel. So this is obviously kind of a very stark way to measure um or crude way to measure uh chat GPT adoption. The reason that we do it is that we think that chat GPT was kind of the first kind of um you know uh in that  
[06:36:50] people had to starting to use generative AI tools even if they start using other tools like claude or Gemini more intensively afterward is still going to be picked up uh by our measure. Uh but to be clear also we've run robustness checks to different variants of this  
[06:37:04] including things like looking at more stringent extensive margin based measures uh utilizing intensive margin based measures or using a broader kind of set of generative AI based tools and we kind of have found uh similar results regardless of of what we've done. Okay.  
[06:37:18] Now let's go into how we classify uh sites into whether or not they are predominantly productive or leisure uh uh in nature. So the way in which we do this is for each website in our sample, we scrape domains to obtain the HTML  
[06:37:32] meta meta data [clears throat] which includes things like self- descriptions and keywords that websites maintain to do things like search engine optimization.  
[06:37:40] And then what we do is we feed this metadata into GPT4.1 mini and ask it to generate the five cleaned activities that the website appears to be associated with.  
[06:37:53] Okay. And then with these five activities, we're again going to feed them into an LOM 4.1 mini and ask the LOM to classify whether or not each activity seems to be associated with a productive or a leisure task where by productive or leisure task we're going  
[06:38:08] to utilize the definition um built by agricunis that was kind of based off the American time use survey. Uh productive activities are going to be things relating to market work, education, child care, shopping and so on. Uh to be  
[06:38:22] clear, even if something like shopping is ultimately based uh on the purpose of engaging in leisure, like you're buying a book to read in the future just for fun, the act of shopping itself is not going to have an intrinsic kind of pleasure uh in it or well for some  
[06:38:35] people at least. Um and so it's going to be classified as a productive home task.  
[06:38:40] Leisure activities are going to be things with intrinsic enjoyment like gaming, social activities, uh streaming and so on.  
[06:38:47] Okay. And then as I said, the LOM is going to slap a label by their productive or leisure depending on the initial classification of activities.  
[06:38:55] We're also going to allow the LOM to spit out a mixed category label because of the fact that some sites like search engines can't really be uh given a classification into either of the two categories. Their diversity of use is too great and so we're going to separate  
[06:39:08] them in our analysis. Okay. And then for validation, we went through and did a manual uh labeling of the top domains in our sample and uh validated that it matches up reasonably  
[06:39:20] well with the LOM based labels. Okay, in the interest of time, I'm going to go straight into how we then uh measure household generative AI exposure to chat GPT. So, we're going to again start with the five activities that we uh had the  
[06:39:34] LOM label for each website. For each of those activities, we're then going to ask chat GPT after feeding in a rubric of Chat GPT capabilities to say whether or not that activity has overlap with  
[06:39:47] chat GPT or not. So whether or not chat GPT use can reasonably substitute for what activity was previously done on the website.  
[06:39:55] We're then going to aggregate these activity level exposure measures up to the website level by just summing whether uh by summing the zeros or the ones. So each website is going to have a number from 0 to five. Five corresponding to five out of five of the  
[06:40:10] websit's activities overlapping with chat GPT.  
[06:40:13] Okay. And then we are going to for our analysis build a household level exposure measure in which we are going to simply say what is your share of browsing before the release of chatgpt.  
[06:40:23] So in 2021 of websites that are classified as being exposed to the release of chatgpt. So, just to to um say that uh more carefully, it's going to be a household level weighted average where the weights are going to be the  
[06:40:38] 2021 browsing shares of all the different sites. And then the thing that's being averaged is the 01 indicator for whether or not the website overlaps with chat GBT in the sense that it has either four or five out of five of its activities overlapping with chat  
[06:40:52] GBT. Okay. And then let's get straight into what we do with these measures. So, first off, just as a sanity check, our measure of household uh chat GPT adoption uh kind of looks to exhibit time trends similar to those that have  
[06:41:06] found in surveys and that also kind of correspond to what we kind of a priority would expect in the sense that we kind of have, you know, pretty pronounced accelerations of adoption behavior around kind of model release states that generated a lot of buzz like the release  
[06:41:19] of multimodal capabilities or the more recent release of a genetic capabilities.  
[06:41:25] And then we also find and this is our first finding that there's more rapid adoption by highincome households and by young uh households. It's not quite monotonic but the patterns uh look uh kind of uh to to exhibit the income  
[06:41:39] gradient that's positive and the age gradient that's negative. And then our second finding is that once we look at X anti-exposure, so no longer demographics, but the exposure measure based off of what households uh activities were in 2021, we find that  
[06:41:53] Xantiexposure is very strongly predictive of whether or not a household ends up actually adopting Chat GPT.  
[06:42:00] This continues to be the case even when we look within our income by age bins.  
[06:42:04] So what I'm plotting here are the coefficients from a dynam a saturated dynamic regression where the left hand side variable is the indicator for chat GPT adoption as of a given quarter t and the right hand side is just a saturated interaction of uh time dummies with the  
[06:42:19] household level exposure measure uh again based off the 2021 activity shares and what we find is that um even within income by age cells households that have more xanti uh usefulness of chatbt according to our measure are  
[06:42:33] substantially more likely to adopt it. And so because of the fact that we're able to look within household types, this is going to kind of form the backbone of our approach to trying to estimate the treatment effects of having adopted chatpt.  
[06:42:44] Uh so just to kind of uh fix ideas in terms of what exactly we're going to be uh doing uh for these uh treatment estimates, we're going to be uh making this exposure measure on the basis of the 2021 website level shares. We're  
[06:42:58] then going to have kind of a burnout period that we aren't going to incorporate into that measure. And we're then going to look at what the difference is in a household's browsing allocations between 2022 uh so January to November before the release of chat  
[06:43:12] GPT relative to 2024 after uh it's been released and household adoption is kind of diffused throughout the population.  
[06:43:20] So in particular the specification is going to be on the left hand side just regressing the change in browsing outcomes between these two pre and post periods and then the right hand side is going to be um in our indogenous specification just a dummy for whether or not the household adopted chatbt  
[06:43:34] which we're going to instrument in a two-stage lease square specification with our measure of household uh exposure which we're going to interpret econometrically as just a shift share instrument where the shares are going to be the 2022 weights and then the shifter is going to be whether or not the  
[06:43:49] website is uh overlapping with chat GPT or not. Okay. And in all of our regressions, we're going to be looking within again our age by income cells. So we're always going to be including age by income fixed effects. So that we're using the within demographic band  
[06:44:02] variation in exposure. Okay. So what do we find first? Just in the reduced form. So this is just a regression of the browsing outcome on the left hand side on the saturated set of time dummies interacted with our measure of exposure on the right hand  
[06:44:15] side. Well, what what we find on uh the lefth hand panel is that households that are more exposed to chatbt experience substantially higher browsing duration after chatbt has been released. What the right hand side graph shows given that the units look the same that the time  
[06:44:30] path of the coefficients is very similar is that this expanded browsing activity that households that are more exposed to chatb engage in is almost entirely in the leisure category. All right. And this is just showing the full set of IV uh estimates where the first stage is  
[06:44:44] very strong and again just it's saying in a poolled sense that the adoption of chatbt has predominantly led to a greater allocation of browsing time into leisure. Okay, I'm going to skip over this section entirely and just describe exactly what the structural estimation  
[06:44:59] is doing in the interest of time. So, uh, as I previewed in the introduction, what we find in in the kind of highfrequency window part of our paper is that households are predominantly engaging in productive activities when they're actually using chat GPT, even  
[06:45:12] though on an overall time allocation basis, we estimate that they're pivoting towards leisure tasks when they actually adopt chat GPT. So, what do we make of this? Well, we're going to adopt the time use framework of agard 2021 in which households are going to get  
[06:45:26] utility of spending time on either productive or leisure digital activities and the intuition behind our structural estimation can be entirely uh explained using this kind of econ 101 uh type  
[06:45:38] graph. So what we uh actually estimate are the points A and B. Point A you can think of as in the space of on the x- axis the amount of productive browsing people do and y-axis leisure browsing that people do before the release of chat GPT we estimate how the average  
[06:45:53] household moves from uh point A to point B after adopting chat GPT so pivoting towards more leisure browsing doing only marginally more productive browsing what we want to infer is how the time budget line has kind of pivoted around the  
[06:46:06] x-axis which is uh a way of saying how much more efficiently browsing activity is after adopting chat GPT. But to do that we need to know uh one particular thing which is the curve uh the uh the  
[06:46:20] slope of the ankle curve because without doing that we can't distinguish between uh the angle curve being a particular slope and the the budget line pivoting out more or less and so what we do in the paper is we estimate the angle curves as I said we estimate that  
[06:46:33] leisure activities have an angle curve elasticity substantially greater than one meaning that uh they are leisure luxuries productive activities have an elasticity substantially below one and so then through the lens of this frame framework. We can back out that  
[06:46:46] productive activities uh have become more efficient after adopting chat GBT which because uh leisures are uh time luxuries has led households to pivoted towards allocating much more of their time into leisure and uh doing the  
[06:47:01] productive tasks uh more quickly but not then allocating more overall uh time to okay um I am already over time so why don't I just wrap up by just uh quickly going over the main findings so using  
[06:47:14] our large panel of detailed housing uh household internet browsing data. We construct new measures of household uh productive or leisure browsing activities and then estimate our long difference IV specification that suggests generative AI use at home has facilitated more efficient completion of  
[06:47:29] productive tasks, but that this has led households to utilize their expanded time budget by engaging in substantially more leisure activities. The kind of next step is to then translate this into welfare effects and kind of uh speak to how much of the missing productivity  
[06:47:43] effect of GDP of chatbt use at home is missing from the GDP statistics. Thank you.  
[06:47:51] [applause] >> Our discussant is going to be Martin Baraja.  

## Discussant remarks (06:47:54 – 07:02:47)

*Shared across the three papers in this session; Martin Beraja discussed all three.*

[06:48:03] >> Hi everyone. Yes. Uh thanks Eric and Karen for inviting me to discuss these uh three papers on AI productivity and economic value. And uh so know discussing three papers is kind of hard in only 15  
[06:48:18] minutes. I'm not going to go deep. I saw my role as a discussant as know framing a bit what these how these papers fit together and uh and in fact and also sort of giving you a uh know hitting the call of Matt Jackson's this this morning  
[06:48:32] on you know there's one shallow view of this p of this session and also I think it applies to other sessions which is oh these are just three papers that measure AI use and value uh at different layers national accounts firms and households  
[06:48:47] and in a bunch of other sessions we saw the same idea A that no people are just trying to paint pictures of what is going on with AI use adoption uh value and I think the call of MAT which I wholeart wholeheartedly agree on is that  
[06:49:01] we need a bit more theory to kind of see what like what are we learning from all these different pieces of evidence and the theory that I'm going to put on the table right now is that I think these papers are pointing to something deeper.  
[06:49:14] It's not just different layers. They're pointing uh to thinking about AI not just as an as a technology that uh will automate production tasks like a fancy robot but more so that is a technology  
[06:49:28] that can accelerate learning across different agents. So firms, workers and households when they learn and codify and act on on information. Okay. And and so the the the question the organizing question for the discussion is not going  
[06:49:42] to be is AI being adopted which can be you know read as a bit shallow but does AI change how agents learn and use uh information and uh and I would say the the different papers measure different  
[06:49:55] parts of this learning process as it happens through the economy starting with households searching uh ideas and jobs and learning skills and then turning into firms and workers who then qualify some sort of knowledge into organizational capital that then  
[06:50:09] aggregates into something that we can measure in the in the in the national accounts and so know starting with the last paper that here comes first is the household's paper uh where there you know their language is that no judgment is mostly used for productive online  
[06:50:23] tasks and and I like to translate that into being about pre-market entry learning in the in the following way. So one which they talk about quite a bit in the paper and another one that I which they talk more about. I think you have  
[06:50:37] the the data there to to do so. So the the first part is is is learning uh by workers or potential workers both through education and job search different informationational activities  
[06:50:49] that end up bu being built into what we will call just human capital matches across firms and workers and and know in the in the paper they they go the they they go a long way uh trying to to look at those type of activities. There's a  
[06:51:02] separate activity that also happens at home which is someone just ideating a new project, a new firm, a startup and uh and in that process no they're doing a lot of informationational research and discarding ideas that may be good  
[06:51:17] discarding bad ideas and and focusing on on on better ideas and that all fits into entrepreneurial pre-entry signals that eventually become firms. So I would really like I I'll show you later. I think this second part is actually a big  
[06:51:31] deal. it could be a big deal quantitatively in in how it fits into aggregate GDP. Uh the paper then that goes and now know we're moving within firms and uh and here I really liked in  
[06:51:45] this paper how they separated the the functions and not just saying oh these are the tasks that are performed within the firm versus not based on own it but functions that really uh align in my mind better with what is organizational  
[06:51:58] learning. So the bottom up channel that that that was discussed this idea that there's workers that are experimenting with both the tool and the task and are learning in the process of performing those tasks how to uh performing better  
[06:52:10] and which eventually leads to some codification coming from the top of those uh activities as being firm policies or broader functional uh deployment. So  
[06:52:23] again this this this this chain from individual learning to organizational capital is kind of what is what is being uh shown in in in the in the in the paper on firms and you know an initial  
[06:52:36] reaction in that paper is well if you see some of the numbers actually they look small in terms of adoption uh so far so narrow use but I'm I think that the way I read this is just early evidence for organizational learning the firms are still figuring it out exactly  
[06:52:51] how is it that we're going to uh this uh these new AI tools and uh and again I would commend the authors to really like push this idea of how is it that we move from one layer to the to to the other layer. I thought that was the most interesting part of of the paper,  
[06:53:06] not just oh here's another survey where we're measuring AI use within firms. Um lastly the the paper on on intangibles or AI organizational capital in the argate. I love this phrase. I didn't  
[06:53:20] know it from Mor it was from Mgrim and Roberts that that one way to think of organizational capital is as the accumulated stock of solutions to the meta task coordination problem. I think that's a fantastic phrase. It's a fantastic way of thinking about this and  
[06:53:33] that what AI can do here is really reduce the unique costs of coordinating activities uh within the company that then leads to organizational capital and things that we measure eventually as higher uh productivity. And so this is  
[06:53:47] really the bridge from task exposure measures to organizational learning. How how we how we gonna uh coordinate those uh those tasks. Now for those of you who have never thought about intangibles but  
[06:54:00] know a bit about physical capital and the perpetual inventory method, let me spend like two sec two minutes on on on why is it what Carl was doing hard and and and why I give her a lot of credit for trying to do this. So if you think  
[06:54:13] about physical capital, how is it that we measure it? don't go out and count machines. That's not what we do. What we do is we count the flows. So we look at say machine purchases the flows we some sort of deflator for for those prices  
[06:54:28] and then we come up with a measure of the quantity of machines that were uh added to the stock and then we accumulate those and hope that over time know the old machines depreciated enough that they don't matter anymore and we get a measure of the stock. That is the  
[06:54:41] perpetual inventory method. This works well because what you're measuring on the investment side machines is exactly that what goes into the stock. Okay. The problem with organizational capital  
[06:54:55] intangibles now the new version being AI based organizational capital is that what you're seeing say use of AI or expenditures in AI in cloud compute whatever it is like give me your your  
[06:55:08] preferred measure is not really what goes into the stock. It's an input into what goes into stock. So you're not measuring the the investment flow. Well, in particular, you can think for example that the way to build the stock of organizational capital is not just with  
[06:55:23] compute or tokens or AI use expenditures. You need to combine that somehow with say managerial and worker time. Okay? And that by combining those two things they figured out how to say produce better, coordinate those meta  
[06:55:36] tasks. That's the knowledge that is being produced. And what you would like to measure is this knowledge flow not the use of AI and that is hard because we don't observe that other part we don't observe the managerial time we don't observe the the worker time or any  
[06:55:50] other resource that goes into the the investment of uh uh that leads to organizational capital accumulation and and so what are the risks uh or when is it that this works well so suppose that you know the imagine that the investment technology is just CRS. So here you can  
[06:56:04] see very clearly that that if you have and no they did a great job at trying to measure the the decline in in the price of of of uh of AI. If you have AI that is getting cheaper then what's going to  
[06:56:18] happen within the companies that no we're going to use a bit more AI and we're going to substitute away from other things that we were using before to build organizational capital. So if we just count and accumulate all the expenditures of AI, we're gonna be  
[06:56:33] overstating the actual increase in investment because we're not seeing all the substitution that is happening say with manager time, worker time, h whatever it is. Uh so no, this only works it works perfectly when the two are perfect compliments. So if the  
[06:56:47] managerial time and AI are perfect complements, let's say to build organizational capital, then you're getting the right measure. But otherwise, you're you're overstating it.  
[06:56:55] uh so one idea here is to use you know how do you fix this so you could use substitution elasticities and I don't know if you try this for other form other forms of IT related organizational investment that we have seen in the past  
[06:57:09] and no the extrapolation would be well maybe the institution are similar between computers and managial time than they were than they are now for AI and and managial time and then we then we move we move forward uh now the second  
[06:57:23] one I this one really I don't know what to do. I think it's a it's a it's not just a problem here. It's just a more general problem that when you're measuring AI use or when you're measuring AI investments uh and  
[06:57:36] expenditures, what you're seeing is not really the part that goes into that is really investment. You're seeing both the production and the investment side.  
[06:57:47] So imagine that you know you're spending X number of tokens and that's what we see. That's what you see for each firm.  
[06:57:53] We don't know where those tokens were used for writing emails which would be like a production task or they were used for you know ideating new products or codifying information anything that looks more like a like an investment. So  
[06:58:06] again, I think this is a call for uh perfect this is a call for you know kind of trying to separate better what is investment what what part of what we're seeing as being AI use and AI expenditure is really an investment  
[06:58:20] versus something that goes into uh production and and it's maybe automating or augmenting uh those tasks. So here I think the two papers the firm and the intangibles paper can kind of like talk to each other a bit and see whether from  
[06:58:34] the surveys we can get some idea of how much of the use is really something that looks more like investment versus something that looks more like production and with that we can do maybe some some adjustments. Um uh  
[06:58:49] good so let me just close you know again going back to to theory uh a bit on what are the no I talked about learning but what are really the two mechanism of learning that can be important here. So one that the firm paper and the national  
[06:59:02] accounts paper point to is really uh learning through firm organizational capital accumulation and uh and and that may that that that may an important part of the story might only be half the story. A more important part or as  
[06:59:17] important part that I think you know receives much less attention is whether AI can help you know potential entrepreneurs or even companies select better which projects are more likely to  
[06:59:30] be successful and households the household paper is going some way into trying to measure that. I would push you to do way more on this because I think there's we we need much more work on figuring out uh this part so that AI may not only matter by speeding up our  
[06:59:45] capital accumulation but also by improving uh selection of projects in the in the economy and and so let me just you know give you you might say well why do we care about all these things is this a big deal or not and another way of asking that question is  
[06:59:59] the question that we asked in this paper with with Edward Talamas on if we had a technology that could accelerate learning like AI high then what would the aggregate uh GDP gains uh be are those potentially large or not if because if they're not large then we  
[07:00:13] could just know close shop and just think about AI as a robot h but if they are large then I think it requires much more careful thinking and measurement about uh those two those two channels so in this paper we know we came up with with this metric vault which is the value of organizational learning  
[07:00:27] technologies that and we show that in a very large class of firm dynamics models where there is learning happening you can come up with uh the gains just by looking at two sufficient statistics that I try to capture how much learning  
[07:00:41] is going on in the economy. So one is if you see an economy where mature firms are much larger than than than young firms that is an economy where organizational capital accumulation is really slow. It takes time and and it's  
[07:00:53] important and and and another the other stat statistic is if you're seeing an economy where young firms exit much more frequently than all than older firms that is an economy where selection at entry before entry is not very good. So,  
[07:01:08] we're not doing a pretty a very good job at the economy is not doing a very good job at selecting uh uh the the right projects uh pre-entry and and when you compute those statistics in in the in the in census data, you come up with numbers that are you know older firms  
[07:01:23] are three times as large as younger firms and they die uh no seven times less than than uh than uh than younger firms. And so that gives you a a gain in terms of aggregate GD potential gain of  
[07:01:36] accelerating learning if you go to the very top firms. So this accelerating a lot 40 years. So we think about this as as an upper bound of two meaning that no accelerated learning could double GDP in the US. And um now what we don't know in  
[07:01:50] this paper with the dart is well how much does AI accelerate learning? We can look at different scenarios. Imagine you can accelerate learning by 40 years, 30, 10, two and then the gains are very different whether you have you know  
[07:02:04] whether a accelerates learning by many many years or not and also depending on the mechanism that you have in mind whether it works through mainly through organizational capital or through the better selection and and prediction of projects. So then the call for all these  
[07:02:18] you know the the papers that are working on measurement is try to frame the discussion more in terms of these mechanisms and and and and the efficiency gains in each of these mechanism in terms of for examples years accelerated uh of learning because I  
[07:02:32] think then that's going to create much more value for the rest of us that are trying to quantify uh sag gains or or macro implications of this evidence to try to map them better to to to the models there. Okay. Thank you very much.  

## General Q&A (07:02:47 – 07:10:06)

*Shared across the three papers in this session.*

[07:02:47] >> [applause] >> We've got our microphone up there if people want to line up if you have questions or comments.  
[07:03:00] >> All right. Yeah, please go ahead. >> Yeah, thank you. Um, so thank you um all for coming and presenting uh to us. I think these were amazing papers. Um, I guess I have a question maybe for the first two especially about AI diffusion.  
[07:03:15] Um so there was a lot of focus on the AI diffusion maybe within workers workers adopting AI and within organizations where um either from the top down or bottom up it's used more but if we think  
[07:03:28] about AI diffusion across like an entire industry or across the economy there's also maybe similar to the idea of a selection effect uh mentioned by the discussant um you could have new firms who pop up who are early adopters of AI  
[07:03:42] technology. I mean, I'm I'm just struck by we're um in Silicon Valley right now, the home of startups, and there are so many uh AI native firms hoping to disrupt, you know, existing incumbents and existing uh industries um with uh  
[07:03:55] some sort of solution that's fully AI uh based. Um so I was wondering um if there are any thoughts on like metrics uh that could be used to track this to see if it's important or not. And I guess if  
[07:04:07] you have any personal opinion about um this uh avenue being important for AI diffusion within the economy. Thank you.  
[07:04:18] >> Do any of the Yeah. Uh could you get a micro There's a microphone right there.  
[07:04:27] >> Is it like this maybe? Uh okay, it works. So yeah, that's a great question.  
[07:04:33] One thing we haven't so far accomplished in this paper is that we didn't match the BTOS data to LBD which is the ne next task and it's almost done and then we're going to be able to bring in the  
[07:04:45] age variable for firms right so then we can really look at the diffusion patterns for young firms versus old firms and you know basically uh you know chart out the discrepancies there and  
[07:04:58] see if there's a um extra propensity for younger firms to come in with already AI technologies for example and I do have a related paper that looks at um business formation micro data from the business  
[07:05:11] formation statistics of census bureau where uh using text analysis I'm able to sort of like fuzzily identify uh which firms are AI related uh which uh business applications are AI related and  
[07:05:26] aiming to start AI related businesses. So we have seen an acceleration uh in AI rel related business applications and formations in the last uh eight years since the you know deep learning revolution trying to commercialize these  
[07:05:41] technologies. So definitely there is some evidence of that you know the diffusion through entry and new firm formation is definitely there and maybe we could do more to quantify that. I  
[07:05:53] agree. Yeah. >> Carol, do you have anything to No, I don't have anything to add. I I to that I think uh the the linking to the age  
[07:06:06] variable uh is you know answers the question the gentle the the gentleman had. Um I I think that one does have to bear in mind if somebody is using AI  
[07:06:17] technology to disrupt a traditional business as opposed to an AI firm that's operating in that AI stack. I mean they're just serving very different roles in the economy. Both both are  
[07:06:32] good. Um but um it's the diffusion is the the first example that I used.  
[07:06:41] >> Terrific. Eric, are you go? >> Yeah, it's a question about the com score and maybe I missed it, but the adoption was, you know, if they'd ever they visited chat the website and then there were adopters there ever after,  
[07:06:55] but it seems like the the power of the comore data if you have this second by second. I think most people use or a lot of people anyways use chatpt through the website and wouldn't that give you a very detailed information about  
[07:07:09] extensive use and people who use it for a while and then stop use it. I've seen some of the adoption surveys see falling usage by by some people or or just level of intensity and I don't know if you could do more with that or if I missed a  
[07:07:22] reason why you weren't >> Yeah. No, I mean I think it's a very welltaken point. Um we have the intensive margin based measure of just you know in a given quarter how many seconds or what share of your browsing  
[07:07:35] seconds did you spend on chatbt or openai.com?  
[07:07:40] uh but it's not something that we use to kind of uh extend any of the the insights of of our analyses. It was more just kind of as a you know robustness check for the indogenous variable. But I agree with your point.  
[07:07:52] >> You can say I mean honestly that would have been the first thing I would have looked at. But but what what what is is there anything you can say about what what that showed? I mean >> yeah I mean we find similar treatment effects essentially when we scale it by uh that alternative endogenous variable.  
[07:08:04] But I think kind of to your you know more to what you're asking I think that you know there's a lot more that we can push on in terms of things like is there a generative AI divide when it comes to if you start using it do you keep using it like is the divide we see the  
[07:08:19] >> John Hartley found some evidence in their their survey that that usage was going down I don't know whether it was like a blip or noise or whether it was real did you see anything like that >> um I honestly do not think that we have  
[07:08:33] No. So we see that the intensive margin is >> usually end before when when does your data end? 2020.  
[07:08:39] >> So it's about to be extended up until the end of 2025, but currently it ends in December 24.  
[07:08:46] >> Well, no, we haven't seen any evidence of a dip so far, but I think within person that could, you know, it could be masking within person heterogeneity. So yeah, >> Eric, that's great. Just to add to that um so we are also exploiting the  
[07:09:00] substitution between computer versus mobile because uh when we talk to the open AI I think you know there's an important fact folks are here um so uh there is a kind of a massive rise of  
[07:09:13] mobile use at the end of our sample I think it's a June of 2025 I believe so that's that's another thing I think you're mentioning we're we're really excited to look into. Yeah.  
[07:09:27] >> Okay. Since we have about a minute left, I just want to ask if any of the paper authors have anything to say in reaction to Martin's excellent discussion.  
[07:09:45] [laughter] >> All right.  
[07:09:49] >> I just want to thank you. It was really nice and you gave us some new ideas to exploit entrepreneurship. Um, yeah, that's cool.  
[07:09:57] >> Okay, we have a 15minute break, a little more networking time, and then uh please be back here promptly at uh 4 o'clock and we have an amazing panel. See you then.