Many surveys show that people use ChatGPT more outside of work than for work, yet nearly all of the policy debate about AI focuses on labor markets and firms. For instance, the data from OpenAI in the graph below shows that most ChatGPT messages are sent for non-work purposes. Debates about the impact of AI on the economy and on society—and related policy proposals and economic predictions—thus tend to miss one of the most important aspects of generative AI: it is also fundamentally changing how we live our lives outside of work. But there has been little evidence so far on how this changes households’ private lives.
We address the lack of evidence in this area in a new research paper on “The Household Impact of Generative AI: Evidence from Internet Browsing Behavior.” My co-authors Michael Blank (Stanford), Miao Ben Zhang (USC) and I find that generative AI use leads to large changes in how households spend their time on the internet: ChatGPT adoption causes users to spend more time online on what one might call “leisure” browsing activities—such as interacting on social networks or video streaming—than on “productive” home browsing activities like informational research, job search, travel booking etc. We also find that ChatGPT use leads to large efficiency improvements in households’ online activity, with our estimates implying 76-176% increases in digital productive task efficiency.
Household ChatGPT adoption and the “GenAI digital divide”
The dataset we use to study household adoption of ChatGPT is a panel from ComScore that tracks all browsing activity on their home machines for a panel of thousands of households. In this data, we can determine when a household uses the ‘ChatGPT’ site for the first time, which, for simplicity, we denote as the date when this household first adopts generative AI.
The rates at which households adopt GenAI in this data broadly track the time patterns found in survey data - but one advantage is that we can track ChatGPT adoption rates continuously over time by households with different characteristics.
This data reveals a clear “GenAI digital divide”: as shown in the graphs from our paper below, both younger and higher-income households have adopted GenAI faster than older and lower-income households - and these gaps in adoption are, if anything, getting larger over time: there is no sign of convergence or “catching up” here even as GenAI has become hard to miss due to its omnipresence in the news. This means that simply knowing about the fact that GenAI is out there is likely not what’s keeping those who have never tried it from doing so - and might instead point to more fundamental barriers to adoption.
ChatGPT causes households to increase leisure browsing
Why does this adoption gap matter? Because if GenAI meaningfully changes how households spend their time—and our evidence suggests it does—then unequal adoption means unequal benefits.
To understand if GenAI-using households are browsing online differently in a way that matters, we first need a way to define meaningful categories of what the purpose of browsing is. We differentiate two different types of digital tasks—borrowing this distinction from the existing research on household time use:
“Productive” digital tasks: education, childcare, non-market work, civic activities, shopping and personal health care.
“Leisure” digital tasks: gaming, social activities / social media, TV, movies, and reading for personal interest.
We use website meta-data and an LLM classification pipeline to label about 160K online domains as belonging more to one of these categories than the other.
We then consider what happens to these activities when households start using GenAI. However, we want to ensure that we are not estimating changes in browsing behavior driven by recent changes in households’ personal lives that might lead them to both shift their digital activity, and also to use ChatGPT (e.g. losing one’s job might lead to using ChatGPT for cover letter editing, and also to more ‘job search’ online). To isolate the causal effect, we exploit the fact that some households—based on their pre-ChatGPT browsing in 2021—were more exposed to websites that ChatGPT could substitute. This “household exposure” predicts ChatGPT adoption by 2024, even among households with otherwise similar demographics, and is unlikely to correlate with other life changes.
What do we find? Households that use ChatGPT spend more of their online time on leisure, and less on productive tasks. The graph below shows our key result: leisure shares increase by 31 pp, productive shares decrease by 21 pp. There is also a category of “mixed” sites that might be used for either leisure or productive purposes, which sees (reassuringly) mixed effects. So, ChatGPT use changes how households spend their time on the internet!
We also find that overall browsing time increases for ChatGPT users - which will matter for our productivity estimates below.
In what browsing context do households use ChatGPT?
Why do households spend less time on productive tasks and more on leisure tasks if they use ChatGPT? There are two possible explanations: it could be that ChatGPT makes leisure browsing a lot more enjoyable, and so households do more of that. Borrowing terminology from the time use literature (Aguiar et al., 2021), this means that leisure browsing is a “leisure luxury good”: when a technology effectively expands your free time, you consume more it. Conversely, these findings could also be explained if productive tasks are “leisure necessities”, which get done faster with ChatGPT use, but households devote more of the freed-up time to other pursuits, such as online leisure browsing. Fortunately, we can use our detailed browsing data to figure out exactly what’s going on. In fact, one of the key advantages of our data is that we can see the context of chatbot usage, which studies from OpenAI (Chatterji et al., 2025) or Anthropic (Appel et al., 2026) on tasks done inside chatbots lack.
We use the high-frequency information on which websites households browse and consider what happens right before and after ChatGPT uses. We look at the 30-minute browsing windows around visits to the ChatGPT site and compare them to the browsing patterns of an average household with the same demographics that is not using ChatGPT, during the same time window. It turns out, ChatGPT is much more likely to be used in the context of productive online tasks - having about a 25 pp higher share of browsing time. This suggests that ChatGPT use is more likely to impact the “efficiency” of productive online tasks than leisure digital tasks.
Another illustration of this pattern is the set of specific websites that tend to be used right before and after ChatGPT: the table below shows the websites with the highest positive difference in browsing share around ChatGPT uses. As the table shows, most of the websites that people use more around ChatGPT are associated with productive browsing: Google search, LinkedIn, Grammarly, GitHub, education sites, etc.
We can also map websites to content categories that are pre-defined in the Comscore data and find a similar pattern (see graph below): ChatGPT is used with sites focused on education, web search, job search, and business topics.
What is the household productivity effect of ChatGPT?
What does all of this mean for the productivity effect of ChatGPT? Assume that ChatGPT makes some browsing activity more efficient (e.g. by replacing hours of WebMD search with a single prompt and tailored response). We show in our study that this effective increase in a household’s time budget would normally increase online leisure time more than online productive task time—households rather watch more Netflix than do more “chores” online when they get some extra time—but both should go up to some degree with more time spent online. The fact that we actually observe no increase in time spent on productive browsing tasks (and a fall in the relative share)—even though total browsing increases—suggests that households must have become more efficient at getting their productive tasks done in less time.
How much more efficient? We use a standard time allocation model based on a famous paper by Aguiar, Bils, Charles & Hurst (2021), together with assumptions about the relative size of efficiency gains in leisure tasks vs. productive tasks that are based on our measure of the share of activities on the related websites that ChatGPT can substitute. While the right assumptions here can be debated, they almost always imply very large efficiency gains in productive online tasks for private households—our preferred range implies 76-176% efficiency increases—and also some efficiency gains (albeit smaller) for online leisure tasks.
Implications for thinking about AI and economic policy
Why does this matter? Most of the policy debates around AI do not mention that there might be massive benefits to private households. While other researchers may come up with different estimates, if they are even in the same ballpark as what we find here, it means that private household benefits might be equally (or more!) important than labor market consequences of AI in terms of what matters: living lives where we can get the things done that we have to do, so we can spend more time on the things we want to do. And this should matter for regulation and ensuring society realizes the benefits from AI. For example, the recently proposed NY State Senate Bill S7263 bars AI chatbots from providing some of the information that licensed professionals provide, such as medical advice or legal advice. This bill tries to directly protect workers whose jobs may be displaced, but at the cost of reducing the benefits of AI to consumers. If the household benefits of AI are large, policy should take this into account! In any case, a lot more research should be done on the household impacts of AI, in order to expand on, and validate, our findings, and ensure that policymakers consider the large share of the productivity impacts that are taking place outside of labor markets.
Our study also has implications for maximizing the benefits from AI: as a society, we should invest in helping those households currently lagging in their private use—older and poorer households, in particular—to catch up in their adoption, e.g. by providing free training or licenses for state-of-the-art chatbots, to ensure that the benefits can reach everyone.
References
Aguiar, M., Bils, M., Charles, K. K., & Hurst, E. (2021). Leisure luxuries and the labor supply of young men. Journal of Political Economy, 129(2), 337-382.
Appel, R., Massenkoff, M., McCrory, P., McCain, M., Heller, R., Neylon, T., & Tamkin, A. (2026, January 15). Anthropic Economic Index report: Economic primitives. Anthropic. https://www.anthropic.com/research/anthropic-economic-index-january-2026-report
Blank, M., Schubert, G., & Zhang, M. B. (2026). The household impact of generative AI: Evidence from internet browsing behavior. arXiv. https://arxiv.org/abs/2603.03144
Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., & Wadman, K. (2025). How people use chatgpt (No. w34255). National Bureau of Economic Research.
Your finding that ChatGPT adoption causes a 31 percentage point increase in leisure browsing share while productive browsing falls is one of the most counterintuitive data points I have encountered on this topic. It suggests AI's efficiency gains may be showing up as more enjoyable downtime rather than more output at work, which could help explain why macro productivity statistics remain so stubborn despite high individual task efficiency numbers. The distinction between task-level efficiency and firm-level output seems to be the key missing variable in most of these debates.