Households that adopt ChatGPT reallocate saved time into online leisure rather than increasing productive browsing, implying generative AI sharply raises the efficiency of household digital tasks. Adoption has been fastest among younger and higher-income households, producing a widening 'generative AI divide' in who benefits.

Main Finding

Adoption of ChatGPT by U.S. households (measured on home devices) causally increases households’ online leisure time while leaving time spent on productive digital tasks roughly unchanged. Households predominantly use ChatGPT in the context of productive non-market activities (education, job search, informational research), implying that generative AI raises the efficiency of productive online tasks and enables a reallocation of freed-up time toward leisure. Mapping the time reallocation into productivity implies large implied efficiency gains for adopters (preferred calibration: ~76%–176%, roughly a doubling).

Key Points

• Data and scope
  • Uses Comscore panel browsing data for >200,000 U.S. household home devices, 2021–2024 (pre- and post-ChatGPT release).
  • Focuses on household-level changes in browsing behavior before/after ChatGPT (released Nov 30, 2022).
• Adoption patterns and inequality
  • Rapid private adoption but with a widening “generative AI divide”: higher and faster adoption among higher-income and younger households; little sign of convergence.
• Predicting adoption
  • Constructed an ex-ante household “exposure” measure (Jan–Dec 2021) that captures how much a household’s prior browsing overlapped with chatbot-capable website tasks. A 1 SD higher exposure predicts a 2.5 percentage-point higher probability of ChatGPT use by Dec 2024.
• Causal effects (IV long-difference estimates)
  • Instrument: pre-ChatGPT exposure (2021) → plausibly exogenous predictor of post-release adoption.
  • Main IV result: ChatGPT adoption substantially raises leisure browsing time (paper reports an increase of roughly "150 log points") and increases the leisure share of browsing duration by ~30 percentage points, while total time on productive online activities remains statistically unchanged.
  • Interpretation: ChatGPT substitutes for multi-site productive browsing (and/or shortens task durations), freeing time for leisure.
• Mechanisms / browsing context
  • High-frequency analysis (30-minute windows around ChatGPT visits) shows ChatGPT use is concentrated around productive websites (search, education, job search, informational sites) rather than leisure sites.
  • Declines in other browsing after adoption are concentrated in categories that ChatGPT can substitute for (search, news, Stack Overflow), consistent with productivity substitution.
• Productivity mapping
  • Uses a time-allocation model (adapted from Aguiar et al., 2021) and estimated Engel-curve elasticities (identified via an IV using local precipitation shocks) to convert time savings into implied productivity gains.
  • Preferred calibration implies efficiency gains in productive digital tasks of about 76%–176% for adopters (authors characterize this as approximately doubling efficiency).
• Limitations
  • Observed browsing is limited to household-owned desktop/laptop devices (no mobile device or offline activity observed).
  • No direct measure of task output or welfare beyond browsing-time reallocation; mapping to productivity depends on model assumptions and elasticity estimates.

Data & Methods

• Core data: Comscore web browsing panel capturing all qualified browsing on household-owned machines (timestamps, visit duration, URL, session id) plus demographic bins (income, age, location, household size).
• Exposure measure construction:
  • Web-scraped website content and used LLM-based classification to assess the overlap between each website’s activities and ChatGPT capabilities.
  • Aggregated site-level overlap across each household’s 2021 browsing to form an ex-ante “exposure” score (higher if a larger share of pre-ChatGPT browsing was on sites amenable to chatbot substitution).
• Identification strategy:
  • Main causal design: long-difference regression instrumenting household ChatGPT adoption with pre-ChatGPT exposure, controlling for demographic-by-region fixed effects and browsing composition covariates.
  • Rationale: pre-release exposure predicts later adoption but (plausibly) is exogenous to post-release shocks that would independently change leisure/productive browsing, conditional on controls.
• Mechanism tests:
  • LLM-based classification to label browsing sessions as “leisure” vs. “productive” (non-market) activities.
  • High-frequency windowed comparisons (30 minutes before/after ChatGPT visits) contrasting users’ context to demographically similar non-users.
  • Category-level analyses showing declines concentrated in AI-substitutable categories.
• Productivity mapping:
  • Quantitative time-allocation model translating time reallocations into implied productivity changes.
  • Required Engel-curve elasticities for leisure vs. productive digital activities estimated via an IV exploiting variation from local precipitation shocks (used to identify how leisure and browsing respond to shocks in total browsing time).

Implications for AI Economics

• Non-market productivity matters: Generative AI’s welfare and economic impact extend beyond workplaces — household-level efficiency gains and time reallocations are potentially large and should be included in aggregate assessments of GenAI’s economic value.
• Distributional concerns: The documented “generative AI divide” (by income and age) implies uneven distribution of household-level productivity and welfare gains, which could amplify broader inequality unless adoption barriers for older and lower-income households are addressed.
• Complementarity/substitution across digital platforms: Adoption reduces visits to websites and services that GenAI can substitute (search, news, Q&A sites), implying platform-level demand shifts and potential implications for online ad markets and content providers.
• Policy and intervention relevance:
  • Interventions (training, digital literacy programs, subsidies) targeted at low-adoption groups could equalize access to non-market productivity benefits.
  • Regulatory and welfare assessments of GenAI should account for private household gains (not just workplace productivity), as these may be quantitatively substantial.
• Research agenda:
  • Extend measurement to mobile devices and offline household activities to quantify total-time and cross-device reallocation.
  • Directly measure output/quality of household tasks (learning outcomes, job search success, financial decisions) to complement time-based productivity inferences.
  • Study interactions between home and workplace GenAI use (complementarities, spillovers for human capital and labor-market outcomes).
  • Investigate long-run dynamics (skill acquisition, habit formation, platform responses) and heterogeneous long-term welfare consequences.

Summary takeaway: Household adoption of ChatGPT appears to increase household non-market productivity by making productive online tasks markedly more efficient, producing substantial reallocation of freed time toward online leisure — with large implied productivity gains but rising adoption inequality that merits policy attention.