Pith.

General Economics · arXiv:2608.00038 v1 · submitted 22 Jul 2026 · 58 references

Google's AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy

Zanna Iscenko, Scott Strand, Yiyuan Chen, Guillaume Aimard, Mihai Codreanu, Vivek Sampathkumar, Alex Imas, Julian Jacobs

show 10 more authors

Evalyne Muiruri, Juan Mateos-Garcia, Jia Jen Ng, Samirah Javed, Josh Martin, Omar Ajmeri, Denis Calin, Andrew Kim, Fabien Curto Millet, James Manyika

The paper says

“This paper introduces the AI & Economy ATLAS (Activity, Task, Landscape, and Adoption Study), an ongoing economic research initiative using Google AI usage data.”

more of the abstract

The first iteration of ATLAS is built on 15 million de-identified interactions across the Gemini App, Google AI Mode, and Gemini API. Using privacy-preserving algorithms as well as established and bespoke classification methods, we map AI usage to over 800 occupations, 4000 tasks, 300 household activities, 150 countries, and 140 languages. We then make a number of observations on what the data reveals about AI's diffusion, and its usage at work and in day-to-day life. In the workplace, we show that while AI adoption spans occupations covering just above 88% of US employment, penetration remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope. Outside of work, AI spans activities making up about 98% of Americans' non-sleep time, with disproportionately high use in high-friction tasks such as engaging with government and professional service providers, likely delivering economic value that standard national accounts may miss. Globally, adoption scales with national wealth and has broad linguistic distribution, with English queries representing only around a third of volume. As we build upon ATLAS and expand its scope and capabilities, we will continue to provide large-scale empirical evidence to inform the public, policy and academic questions about the ongoing AI transformation.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 4 Aug 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Big, careful descriptive snapshot of real AI usage, but the headline '<10% automation' figure rests on a cluster-level intent classifier validated only on synthetic data. This line is wrong

4 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 58 references say.

The paper says, at §3.6, Fig. 7

The review did not quote this passage.

Pith says: Objection 1 of 4 · major

The claim that end-to-end automation is less than 10% of non-routine cognitive conversations is load-bearing and rests on an intent classifier applied to cluster summaries, not individual conversations. The paper itself acknowledges that clusters 'could nest multiple user intents'; assigning one label per cluster means mixed-intent clusters will be labeled by the dominant content of an auto-generated summary, potentially undercounting automation.

more of objection 1

Validation in §2.1 and Appendix A.2.2 uses synthetic data generated by the same model family and reports no human-labeled conversation-level test for the intent construct. Please provide precision/recall against human raters on real conversations, or restrict the headline to a preliminary, classifier-relative share.

The paper says, at §3.4, Figs. 3–4

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

The occupational coverage (68%) and median task saturation (21%) are defined relative to arbitrary thresholds—50 unique users per occupation and 25 unique users per task—and no sensitivity analysis is reported. Because the 'broad but shallow' conclusion is central, these thresholds should be varied (e.g., 10/25/100 users) or replaced with a continuous concentration measure to show that the qualitative finding is not an artifact of the chosen cutoffs. Without this, the 'shallow penetration' inference is partly built into the threshold choice.

The paper says, at §2.3, §3.6

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

The sample is a two-week window (April 6–19, 2026) from consumer surfaces and free API only; paid/enterprise API is excluded. The paper correctly lists this as a limitation, but the <10% automation figure in the executive summary is not consistently qualified.

more of objection 3

Enterprise usage is likely the most automation-intensive, and a two-week window may be seasonally atypical (tax-related queries are noted in §3.3). Please restate all 'automation is rare' conclusions as applying to the included consumer/free surfaces and period, and provide robustness checks on alternative windows or partial enterprise data if available.

The paper says, at §2.1, App. A.2.2, App. B

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

The accuracy of the work/non-work classifier, the O*NET/ATUS mapping, and the bespoke intent/expertise classifiers is validated only with synthetic ground truth and internal human ratings; no public validation item set or external benchmark is provided. Because all headline percentages inherit any systematic misclassification (e.g., a bias toward 'collaborative' labels), the authors should release a de-identified validation sample with human labels, or at minimum report per-category precision/recall against a held-out human-annotated set of real conversations.

more of objection 4

Without this, the mapping accuracy cannot be independently assessed.

5 minor objections

The paper says, at Abstract & §4.3

Pith says: minor objection 1 of 5

The Executive Summary states household time savings of 'just 30 minutes per week' while the model in §4.3 uses a range of 0.5–5% (≈5.4–53 minutes/week). Clarify that 30 minutes is an illustrative scenario within the range, not the conservative lower bound.

The paper says, at Abstract & §3.4

Pith says: minor objection 2 of 5

The abstract says '68% of occupations globally' and '88% of US employment' without noting that the 68% is an unweighted occupation share while the 88% is employment-weighted; please make this distinction explicit in the abstract.

show 3 more

The paper says, at §3.4, fn. 14

Pith says: minor objection 3 of 5

The upper/lower bounds on employment coverage at different task-saturation thresholds are informative; consider presenting them in the main text rather than a footnote, since they qualify the 'shallow diffusion' claim.

The paper says, at References

Pith says: minor objection 4 of 5

Several references have formatting issues: 'BLS , 2025' contains a stray space; Comin and Mestieri is listed as 'None 2014'; some NBER working papers and arXiv items are missing complete identifiers. Please standardize.

The paper says, at General

Pith says: minor objection 5 of 5

No data or code availability statement is provided. Even if the full data cannot be released, a statement about what ancillary materials (e.g., synthetic validation prompts, human rating instruments) will be made available would aid replicability.

What would settle it

Draw a stratified random sample of ~5,000 raw work conversations from the same surfaces (or, if privacy rules prohibit, use cluster summaries together with per-cluster conversation counts), have independent human annotators label each conversation/cluster with the five intent categories defined in Section 3.6, and compare against the pipeline's cluster-level labels. Recompute the automation share for non-routine cognitive work using human labels, weighted by cluster size. If the human-labeled share is materially above 10%, or if human annotators frequently disagree with the automated label on mixed-intent clusters, the central 'mostly assistive' conclusion is not robust. Ideally also release the human-labeled validation set and classifier code for independent reproduction.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 4 Aug 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 41 s · 43,543 tokens read, 3,545 written
  3. Editor wrote the verdict above 26 s · 44,349 tokens read, 2,700 written
  4. Referee major revision: 4 major objections, 5 minor 1 min 5 s · 43,882 tokens read, 5,310 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Top of mind: this is the most detailed real-usage map of a consumer AI assistant we have, and it is a genuinely careful descriptive study. But the headline 'less than 10% automation' number is softer than the abstract implies, because it comes from a cluster-level intent classifier validated mostly on synthetic data. The paper is more hedged than its own press would suggest; still, that one number should not be treated as measured.

What is actually new and valuable: the scale (15M interactions across Gemini App, AI Mode, and API), the mapping to O*NET occupations and ATUS activities, the penetration-adjusted cross-country comparisons, and the language-switching analysis. The distinction between occupation coverage (68%) and task saturation (21% median) is a genuinely useful way to frame shallow diffusion. The wage-intensity gradient and the ATUS time-use regressions are plausible and, importantly, do not depend on the contested intent classifier. The authors are transparent about the pipeline and compare themselves to Anthropic and OpenAI rather than ignoring them.

The soft spots, in proportion: the stress-test concern about the intent classifier lands. The classifier labels cluster summaries, and clusters can nest multiple user intents; a single label per cluster will push mixed clusters toward the dominant surface intent. Validation on synthetic data generated by the same model family cannot catch model-specific bias in the 'automation' construct. So the '<10%' figure is a directional signal, not a firm estimate. The two-week window and the exclusion of paid/enterprise API usage mean work-related automation-intensive use is likely underrepresented. The 68% and 21% numbers also depend on arbitrary user thresholds (50 and 25), which the paper acknowledges but which make point estimates fragile. I would push back on the 'circularity' framing in the Pith Report: every classifier-based study defines its outcome; the real issue is validation, not self-reference. The household $100B value is properly presented as a scenario range, so it is less concerning than the automation claim.

Bottom line: this is a valuable descriptive contribution for labor economists and anyone studying AI diffusion. It deserves a serious referee, and I would bring it to a reading group. The main revision I would ask for is a human-labeled conversation-level validation of the intent classifier, or a clear downgrade of the automation claim to 'preliminary cluster-based estimate.'

The strongest objection

The most consequential claim—that end-to-end task automation represents less than 10% of AI conversations in non-routine cognitive work, implying AI is mostly assistive—depends entirely on the bespoke intent classifier in Section 3.6. The classifier does not label individual conversations; it labels cluster summaries, and the paper itself notes that clusters 'could nest multiple user intents.' With one label per cluster, any mixed-intent cluster will be assigned to whichever intent dominates the auto-generated summary. If those summaries are optimized for O*NET/ATUS taxonomy assignment rather than for intent, automation signals could be systematically underrepresented. Validation (Section 2.1, Appendix A.2.2) uses synthetic ground truth generated by the same model family used for classification, so it cannot detect model-specific biases in the 'automation' construct. Additionally, the sample is a two-week window from consumer surfaces plus free API; paid enterprise API logs are excluded (Section 2.3), so the work statistics may miss the most automation-intensive usage. The paper is carefully hedged, but the quantitative '<10%' figure is not yet established at the precision required to support the central policy claim.

The reader's weakest assumption already identifies the pipeline's classification accuracy—including the intent classifiers—as the key uncertainty. My stress-test sharpens this to the specific cluster-level intent classifier behind the <10% automation figure, which is the most policy-relevant and least externally validated component. The paper's own caveats in Section 3.6 and Section 2.3 support this concern: clusters can nest multiple intents, and paid enterprise usage is excluded. A human-labeled validation test would directly settle whether the automation share is robust or an artifact of classifier design. Since the reader already issued a CONDITIONAL verdict, my concern does not change the verdict; it reinforces the need to release validation data or an independent benchmark before the headline should be taken as established.

The referee's report

If the measurement pipeline is valid, this is one of the largest and most detailed empirical maps of real-world AI usage, directly informing debates on AI and labor, task-based models, household production, and the global digital divide. It complements existing industry studies (Anthropic, OpenAI) with a wider set of surfaces and introduces a household time-use taxonomy, a penetration-correction method for cross-country comparisons, and new evidence on language usage. The paper explicitly frames its findings as early observations and provides extensive methodological detail and privacy safeguards. However, the central quantitative claims—especially the <10% automation share and the 21% median task saturation—rest on proprietary LLM classifiers validated only on synthetic data and internal human ratings, with no public benchmark or released validation items. The economic correlations (wage elasticity, GDP elasticity) are descriptive and are not yet backed by causal identification. Thus the paper's significance is conditional on substantial external or human validation of the classification pipeline.

Pith's reading of the paper (a machine's summary, not the authors')

The paper builds a privacy-preserving census of AI use by mapping 15 million de-identified Gemini conversations to standard economic categories: occupations and tasks (O*NET/SOC), daily activities (ATUS), countries, and languages. Its central claim is that generative AI has diffused remarkably widely—touching 68% of occupations and activities covering 98% of Americans' waking hours—but has penetrated shallowly: the median occupation uses AI for only 21% of its tasks, and end-to-end automation accounts for under 10% of conversations in non-routine cognitive work. A sympathetic reader would see this as the first large-scale behavioral evidence that, at this stage, AI is a complement to human work and household life rather than a substitute. The paper also argues that home use, especially in high-friction bureaucratic tasks, generates value that GDP misses, and that global adoption tracks national wealth closely. Why it matters: it moves the debate from exposure projections to observed usage, giving policymakers an empirical foundation for questions about displacement, productivity, and the digital divide.

What carries the argument. The load-bearing machinery is the ATLAS measurement pipeline: 14.65 million de-identified interactions (Gemini App, AI Mode, and API, April 6–19, 2026) are PII-redacted, summarized, clustered by the OCTO tool, and then assigned to the O*NET/SOC work taxonomy and the ATUS time-use taxonomy, plus bespoke intent and expertise classifiers. A penetration adjustment using web-traffic referral shares is applied for cross-country comparisons. This machinery converts unstructured chat into statistically comparable task- and activity-level counts, which is what makes the headline percentages possible; every one of those percentages inherits the pipeline's classification accuracy.

Core claim. On its own terms, the paper's discovery is that actual AI usage is broad but shallow and overwhelmingly collaborative. At work, usage spans 68% of detailed occupations representing 88% of US employment, yet only 3% of occupations show AI usage for more than three-quarters of their tasks, and the median occupation shows usage for 21% of tasks. For non-routine cognitive work—the category that dominates interactions—the intent classifier attributes end-to-end automation to fewer than 10% of conversations, with the rest falling into partial drafting, review, ideation, or information retrieval. Outside work, over 86% of conversational interactions occur, and AI is over-represented in exactly the

The premise it rests on. The entire statistical picture rests on the automated classifiers—work/non-work triage, O*NET task and occupation assignment, intent, and expertise—being accurate on real conversations; they were validated only on synthetic ground truth and internal human ratings, with no public benchmark, and the sample is a two-week window from one provider's consumer surfaces.

If the paper is right (Pith's inferences, not the authors')

If correct, exposure-based projections of massive white-collar displacement are not matched by current real-world usage: the gap between 'affected' and 'automated' is large and measurable.

Household AI use in planning, research, and bureaucracy implies a flow of non-market time savings (the authors' scenario range is roughly $15–150 billion per year in the US) that standard GDP accounting omits.

The steep income and education gradient of usage implies that, absent deliberate diffusion measures, AI benefits may concentrate among higher-wage workers and wealthier countries, widening existing divides.

The 9–12% higher turn counts and 18–20% more tokens for non-primary-language interactions suggest investments in multilingual capability carry a measurable efficiency dividend for users.

The government-services over-representation (roughly 20 times its time share) and off-hours usage suggest AI is already functioning as a de facto extension of public-service access, with implications for digital-government design.

If the shallow-usage pattern reflects current capability rather than intrinsic demand, the 21% median task saturation is a lower bound; one testable extension is to track whether saturation rises as model capabilities improve in the same panel.

The paper's own limitation—no outcome data—suggests a natural extension: link usage clusters to downstream task completion or time-use diaries to estimate actual productivity gains rather than conversation counts.

The language-symmetry result (work and non-work code-switching nearly identical) could be tested against other providers' data; if it holds, it implies that perceived multilingual quality gaps are not driving task-sorting behavior.

The 'interest-adoption gap' between search interest and usage in emerging markets invites a causal study: does lowering access costs (e.g., cheaper data plans, SMS-based AI) convert interest into usage, as suggested by some cited pilots?

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

This paper introduces the AI & Economy ATLAS (Activity, Task, Landscape, and Adoption Study), an ongoing economic research initiative using Google AI usage data. The first iteration of ATLAS is built on 15 million de-identified interactions across the Gemini App, Google AI Mode, and Gemini API. Using privacy-preserving algorithms as well as established and bespoke classification methods, we map AI usage to over 800 occupations, 4000 tasks, 300 household activities, 150 countries, and 140 languages. We then make a number of observations on what the data reveals about AI's diffusion, and its usage at work and in day-to-day life. In the workplace, we show that while AI adoption spans occupations covering just above 88% of US employment, penetration remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope. Outside of work, AI spans activities making up about 98% of Americans' non-sleep time, with disproportionately high use in high-friction tasks such as engaging with government and professional service providers, likely delivering economic value that standard national accounts may miss. Globally, adoption scales with national wealth and has broad linguistic distribution, with English queries representing only around a third of volume. As we build upon ATLAS and expand its scope and capabilities, we will continue to provide large-scale empirical evidence to inform the public, policy and academic questions about the ongoing AI transformation.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-23

The authors' main claim, as the classifier picked it out 0.69

“In the workplace, we show that while AI adoption spans occupations covering just above 88% of US employment, penetration remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.15 · states limitation 0.65 · human participants 0.54 · controlled comparison 0.03 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.07 · mentions code or data availability 0.04

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
This paper introduces the AI & Economy ATLAS (Activity, Task, Landscape, and Adoption Study), an ongoing economic research initiative using Google AI usage data. contribution 0.99 0.76 0.03 asserted 0.99
The first iteration of ATLAS is built on 15 million de-identified interactions across the Gemini App, Google AI Mode, and Gemini API. method 0.92 0.88 0.03 asserted 1.00
Using privacy-preserving algorithms as well as established and bespoke classification methods, we map AI usage to over 800 occupations, 4000 tasks, 300 household activities, 150 countries, and 140 languages. method 0.93 0.94 0.04 asserted 1.00
We then make a number of observations on what the data reveals about AI's diffusion, and its usage at work and in day-to-day life. contribution 0.76 0.82 0.07 unclear 0.56
In the workplace, we show that while AI adoption spans occupations covering just above 88% of US employment, penetration remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope. result 0.99 0.96 0.23 asserted 1.00
Outside of work, AI spans activities making up about 98% of Americans' non-sleep time, with disproportionately high use in high-friction tasks such as engaging with government and professional service providers, likely delivering economic value that standard national accounts may miss. result 0.79 0.93 0.19 hedged possibility 0.73
Globally, adoption scales with national wealth and has broad linguistic distribution, with English queries representing only around a third of volume. result 1.00 0.91 0.08 asserted 1.00
As we build upon ATLAS and expand its scope and capabilities, we will continue to provide large-scale empirical evidence to inform the public, policy and academic questions about the ongoing AI transformation. future work 1.00 0.13 0.04 asserted 0.78

What kind of paper the classifier reports this is: empirical study 0.68 · evidence from analysis of existing data 0.90

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,435 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (45)
Figure 1 of the paper
Figure 1. White-Collar Professions Are Overrepresented in AI Usage Relative to US Employment Notes: This figure compares the shares of US work-related Gemini interactions within broad occupation groups to shares of civilian US employment in those groups. US employment shares at the Major Occupation Group (SOC) level sourced from Occupational Employment and Wage Statistics May 2024 data (Bureau of Labor Statistics, 2025b). Gemini shares are derived from interactions aggregated across surfaces (App, AI Mode, API) classified into O*NET SOC occupations. OEWS data does not contain information on…
Figure 2 of the paper
Figure 2. Using Gemini for Work in Arts and Design Occupations is More Common in Non-OECD Countries While Computer and Math and Business and Finance Occupations Are More Common in the OECD Notes: Gemini shares are derived from interactions aggregated across surfaces (App, AI Mode, API) based on the user’s country/region and classified into O*NET SOC occupations. This chart compares the broad occupation group shares in the 38 OECD member countries to non-OECD countries. 13
Figure 3 of the paper
Figure 3. Gemini Usage is Observed Across 68% of Detailed Occupations, Including the Vast Majority of White-Collar Professions Notes: This chart uses the 22 non-military SOC major occupation groups on the vertical axis and for each of those groups, depicts the share of component detailed occupations where we observe at least 50 Gemini users. The Gemini usage shares are derived from interactions aggregated across surfaces (App, AI Mode, API) classified into O*NET SOC occupations. O*NET SOC codes are mapped to corresponding SOC codes in the May 2024 OEWS data (Bureau of Labor Statistics, 2025b)…
Figure 4 of the paper
Figure 4. For Occupations with Observed Gemini Usage, Workers Are Generally Using AI for Less Than One Quarter of Their Tasks Notes: This chart displays the distribution of task saturation across detailed occupations with non-zero AI usage. “Task Saturation” is defined as the percentage of O*NET tasks within an occupation that see meaningful interaction volume (at least 25 unique users globally). Occupations where zero tasks meet this threshold are excluded. The vertical axis represents a count of occupations, while the horizontal axis states what percentage of tasks are saturated in those…
Figure 5 of the paper
Figure 5. Occupations Without Substantial Gemini Usage Have Greater Shares of Manual Tasks Notes: This figure uses the categories from Autor and Thompson (2025), assigning each O*NET task, covering occupations across the economy, to one of the five categories on the horizontal axis. It then compares the group of detailed O*NET occupations with zero task saturation (omitted from
Figure 6 of the paper
Figure 6. Non-Routine Cognitive Tasks Are Over-Represented in Gemini Usage Relative to the O*NET Task Universe While Manual and Interpersonal Tasks Are Under-Represented Notes: This figure uses the categories from Autor and Thompson (2025), assigning each O*NET task, covering occupations across the economy, to one of the five categories on the vertical axis. The horizontal axis represents the shares of tasks in each of those categories for the relevant distribution, starting with the O*NET task universe. It then compares the O*NET taxonomy distribution to the representation of work-related…
Figure 7 of the paper
Figure 7. Using Gemini for End-to-End Task Automation Is Rare Except for Routine Cognitive Tasks Notes: For de-identified conversation cluster summaries, Gemini is used to label the user intent to determine whether they are seeking to automate the task or complement human work occurring outside of Gemini. This chart depicts the shares of conversations assigned to the intent categories on the vertical axis for each of the task categories with different colored bars. The task categories are based on Autor and Thompson (2025). Routine and Non-routine Manual tasks have been combined into a single…
Figure 8 of the paper
Figure 8. Gemini Usage Is Concentrated in Cognitive Tasks Across the Expertise Spectrum With Usage for Low-Expertise Non-Routine Cognitive Tasks Being Particularly Overrepresented Notes: The baseline O*NET taxonomy is used to categorize tasks into expertise quartiles, where expertise scores are calculated based on the method in Section 3.1 of Autor and Thompson (2025). The task categories on the horizontal axis are also based on Autor and Thompson (2025). Routine and non-routine manual tasks have been combined into a single category. The bars in the figure represent the share of Gemini…
Figure 9 of the paper
Figure 9. Gemini Usage Is Higher in Occupations With Higher Median Earnings and Educational Attainment Notes: Gemini usage intensity is a measure of our observation of AI usage adjusted for the employment population associated with a given detailed occupation. Gemini usage intensity equals the total AI conversations for a detailed occupation in our sample divided by the total employment in that occupation in the OEWS 2024 data. That is also the source for occupation-level annual median earnings. OEWS data does not contain information on Military Specific Occupations so those are omitted. In…
Figure 10 of the paper
Figure 10. Over 86% of Conversational AI Usage Occurs Outside of Formal Work, Dominated by Leisure, Education, and Household Activities Notes: This figure shows the global share of conversational AI requests (AI Mode and Gemini App) across major life activities, classified according to the American Time Use Survey (ATUS) Tier 1 major activity categories (excluding sleeping). Work conversations shares is inclusive of the share that was considered “work” in the first step of our pipeline. Non-work interactions account for 86.5% of all AI conversations, with nearly half concentrated in two…
Figure 11 of the paper
Figure 11. AI Assistance Over-Indexes in Cognitive and Bureaucratic Tasks, and Under-Indexes in Physical and Location-Bound Activities Notes: This figure compares the share of U.S. non-work AI conversations against actual human time allocation across ATUS Tier 1 major categories using dumbbell plots. Both shares are calculated out of total non-work active time (excluding work and sleep) to ensure direct comparability. Cognitive, informational, and administrative tasks such as Education, Consumer Purchases, and Professional & Personal Care Services are heavily over-represented in AI…
Figure 12 of the paper
Figure 12. Human Time Allocation Strongly Predicts AI Conversation Volume Across Daily Activities Notes: This scatter plot illustrates the relationship between the share of human time allocation (x-axis) and the share of U.S. AI conversations (y-axis) across ATUS Tier 2 (intermediate-level) activity categories, plotted on logarithmic scales to accommodate broad variations in category size. Each point represents an individual Tier 2 task category (excluding sleeping, working, and uncoded activities). The solid 45-degree line represents exact parity, where an activity captures the same share of…
Figure 13 of the paper
Figure 13. Regardless of the Activities Granularity Level, Time Allocation Explains a High Percentage of the Variance in AI Conversations Notes: This figure displays linear regression estimates (with robust standard errors) quantifying the positive relationship between human time use shares and AI conversation shares across ATUS task categories (excluding work and sleeping). Across both broad (Tier 1/2) and granular (Tier 3) activity levels, human time allocation strongly and statistically significantly predicts where AI is utilized. At the most aggregate level, human time use explains nearly…
Figure 14 of the paper
Figure 14. Conversational AI is Used Across a Wide Set of Productive Household Tasks and Is Particularly Common in Selected Professional, Government Services and Civic Obligations Notes: This figure compares U.S. AI conversation shares with ATUS human time allocation across two sets of economic activities using dumbbell plots. All shares are calculated out of total non-work active time (excluding work and sleep). Panel A (Productive Household Tasks) displays top activities meeting the economic "third-person criterion", tasks you could theoretically pay a third party to perform, such as…
Figure 15 of the paper
Figure 15. Across Countries, Higher Income Correlates With More Productive Household AI Use, While Developing Nations Concentrate Heavily on Education Notes: This figure presents country-level scatter plots and linear macro trend lines exploring how national income (Purchasing Power Parity GDP per capita, log-transformed) relates to the distribution of conversational AI usage across four core life domains: Work & Work-Related, Valuable Household Tasks, Education, and Socializing, Relaxing, & Leisure. Each data point represents an individual country or territory. Higher national income is…
Figure 17 of the paper
Figure 17. AI Assistance Concentrates in Education, Services, and Household Management Rather Than Physical Domestic Chores Notes: This figure compares AI conversation shares against ATUS human time allocation across the 10 most over-represented and 10 most under-represented intermediate (Tier 2) activity domains (excluding work and sleep). The dumbbell plots highlight the gap between human time allocation and AI consultation. Relative to human time, AI is heavily sought out for knowledge-intensive and organizational activities such as education, financial and legal services, and practical…
Figure 20 of the paper
Figure 20. High-Usage Regions Drive Conversational AI Volume Disproportionately to their Population Notes: This figure compares the share of global population against the share of global conversational AI usage across the five per-capita usage quintiles at the country/region level. Multipliers denote the ratio of usage share to population share for each quintile (for instance, countries and regions in the “Medium” quintiles represent 1.2x their population in terms of usage). The quintiles are based on country-level conversation volume per capita, adjusted for Gemini Apps penetration. China…
Figure 21 of the paper
Figure 21. AI Conversational Usage (Gemini Apps and AI Mode) Shows a Clear, Positive Correlation With GDP Per Capita (PPP) Notes: This figure displays the relationship between penetration-adjusted conversations per capita and GDP per capita (PPP) on a log-log scale. Per-capita conversations are adjusted for Gemini’s penetration using StatCounter data and are indexed to a maximum value of 100. GDP per capita data is sourced from the World Bank (2024a) World Development Indicators (WDI) database. For countries/regions with missing data in WDI, we use values from the International Monetary Fund…
Figure 22 of the paper
Figure 22. Access to the Internet Acts as a Bottleneck on Per Capita Conversational AI Usage, but Does Not Explain the Full Disparity Notes: This figure illustrates the regional impact of adjusting adoption intensity for internet access. The baseline metric is conversations per capita normalized by Gemini App’s usage share, while the internet-access adjusted metric utilizes internet users (calculated using World Bank (2024b) internet access percentages from the World Development Indicators (WDI) database) in the denominator instead of the total population. Both metrics are scaled to a maximum…
Figure 23 of the paper
Figure 23. Global Gemini API Usage Per Capita Shows a Pronounced Concentration in Traditional, High Income Economies Notes: This figure displays the per-capita intensity of Gemini API requests by country/region quintiles. Unlike other metrics in this report, conversation counts here include enterprise/educational and paid usage. Usage locations are determined by the IP address originating the API request, meaning the map reflects the server’s location rather than the end-user’s (e.g., requests handled by a US-based server for a chatbot in Mexico will be mapped to the US). Per-capita rates are…
Figure 24 of the paper
Figure 24. maps the per-capita intensity of work-related conversations across conversational AI surfaces. The resulting geographic distribution is broadly consistent with overall usage patterns, showing very few variations. Highly connected, high-income countries continue to lead work-related AI adoption on a per-capita basis. Latin America also shows a strong usage of Conversational AI for work-related usage, with most of the region falling the upper per-capita usage tiers
Figure 25 of the paper
Figure 25. Developing Nations Lead in the Share of Work-Related AI Use Notes: This figure maps the percentage of conversations classified as work-related by country. The work percentage is calculated as the ratio of work conversations to total conversations. Classification of work status is determined by our automated work/non-work classifier applied to conversation content. The analysis excludes China and countries/regions with populations under 1,000,000. This inversion is likely driven by differing economic constraints around digital engagement. In lower-income regions, where metered…
Figure 26 of the paper
Figure 26. Middle East and East Asian Countries and Work Conversations Tend to Have Higher Turn Counts With Conversational AI Notes: This figure displays normalized average turn counts per conversation, split by income group and region, and categorized by work status. Turn counts are weighted, and averages are indexed to the global mean (Global Mean = 100). The underlying data combines Gemini Apps and AI Mode. The metric does not control for language-specific communication styles, which may influence dialogue length. While adoption volume remains concentrated in high-income nations, analyzing…
Figure 27 of the paper
Figure 27. Conversational AI Usage is Multilingual, With Non-English Languages Accounting for the Majority of Global Interactions Notes: This figure displays the global distribution of primary interaction languages across conversational AI surfaces (Gemini App and Google AI Mode), showing the top 15 individual languages by global request share with remaining languages aggregated under “Other.” Regional variations of a base language (e.g., Brazilian and European Portuguese) are consolidated into single language categories. Across 143 privacy-cleared languages, non-English interactions…
Figure 28 of the paper
Figure 28. Rates of Non-Primary Language Use are Remarkably Symmetric Across Work and Personal Domains, Disproving High-Stakes Code-Switching Notes: This log-log scatter plot compares the share (%) of AI conversations conducted in a language other than a country’s primary language across work activities (y-axis) versus non-work activities (x-axis) for countries with over 1,000 total conversations. The figure uses for defining work and work related usage conversational ATUS AI conversations combined with the work/non-work classifier bucket. Each point represents an individual country. The…
Figure 29 of the paper
Figure 29. Language-Switching Does Not Seem To Be Systematically Correlated with Economic Value Notes: This figure shows the percentage of AI conversations conducted in a non-primary country language across major American Time Use Survey (ATUS) activity domains. Activity categories are sorted by overall non-primary language share (requiring at least 100 conversations per activity-country pair). Non-primary language usage is highest in community, civic, and personal identity domains—led by Volunteer Activities (21.9%), Religious & Spiritual Activities (20.5%), and Government Services & Civic…
Figure 30 of the paper
Figure 30. Occupational Classifier Misclassifications Concentrate Most Often Within Closely Related White-Collar Domains Notes: This heatmap displays off-diagonal Excess Confusion Lift multipliers for SOC Major Groups (SOC 2-digit occupations), isolating conceptual classification overlap from overall category sample size. A multiplier of 1.0x indicates that misclassifications occur at the exact rate expected by random chance based on category size. Values above 1.0x indicate systematic conceptual overlap (e.g., 2.0x–3.0x confusion lift between Management and Business/Financial Operations, or…
Figure 31 of the paper
Figure 31. Time-Use Misclassifications Are Heavily Confined to Context-Silent Tasks Notes: This heatmap displays off-diagonal Excess Confusion Lift multipliers for ATUS Tier 1 major activity categories, measuring systematic misclassifications patterns beyond base-rate category frequency. A value of 1.0x represents base-rate chance error. Systematic confusion is tightly localized to categories where conversational summaries lack context, such as distinguishing whether a caregiving or assembly activity is performed for a household member versus a non-household member. Modest overall confusion…
Figure 32 of the paper
Figure 32. Gemini Detailed Occupation Task Saturation Shares With O*NET Importance Weighting Notes: Similar to
Figure 33 of the paper
Figure 33. Distribution of Task Expertise Scores Across the O*NET Task Universe Notes: We calculate scores measuring the expertise of tasks in the O*NET universe by replicating the methodology in Autor and Thompson (2025). Expertise for each task is defined as 100 - the average Standard Frequency Index (SFI) of the lemmatized terms in its O*NET task statement, where SFI is an index combining word frequency and word entropy provided for 140,000 English language words in Zeno et al. (1995). This figure depicts the distribution of resulting values. The dashed lines indicate the quartiles…
Figure 34 of the paper
Figure 34. Granular Task Comparison Highlights AI Over-Representation in Specialized Knowledge and Under-Representation in Routine Chores Notes: Compares U.S. AI conversation shares against ATUS human time shares across the 10 most over-represented and 10 most under-represented detailed daily tasks (minor categories), ranked by absolute percentage point difference (excluding working category and sleep). The dumbbell plots illustrate that at the most granular level, AI assistance disproportionately concentrates in education work, comparison shopping, financial planning, and government service…
Figure 35 of the paper
Figure 35. Human Time Use Positively Predicts AI Consultation Across Hundreds of Granular Household Activities Notes: Log-log scatter plot illustrating human time use shares versus AI conversation shares across all individual ATUS Tier 3 minor activity categories (excluding sleeping, working, and uncoded tasks). Each point represents a distinct granular task. The solid 45-degree line indicates parity between human time spent and AI consultation share. Despite wide variation across hundreds of niche activities, the overall positive relationship persists even at this micro-level, confirming…
Figure 36 of the paper
Figure 36. Bureaucratic Compliance, Licenses, Taxes, and Civic Obligations Dominate Government-Related AI Conversations Notes: Relative share of specific tasks within the Government Services & Civic Duties domain (and associated travel). The x-axis displays the relative share (%) of each specific task out of all AI conversations classified under government and civic interactions. Obtaining government licenses, paying taxes/fines/fees, and navigating civic obligations (such as voting requirements, local civic procedures, and immigration compliance) represent the largest share of citizen…
Figure 37 of the paper
Figure 37. Privacy-Preserving Conversation Summarizing Clusters Also Reveal High Citizen Demand for Help with Legal Requirements, Procedures, and Compliance Notes: Word cloud visualizing frequent substantive terms across AI conversation summaries in government and civic tasks (with word size scaled by relative frequency). Some prominent terminology centers are “administra￾tive”, “documentation”, “compliance”, “requirements”, “application”, “procedures”, “legal”, and to some extent “bureaucratic”, “immigration”, “taxes”, “driving”, and “visa”. 93
Figure 39 of the paper
Figure 39. Non-Physical Cognitive Tasks Show Significantly Higher AI Representation Ratios than Physical Activities Notes: Probability density functions of AI representation ratios for physical vs. cognitive ATUS Tier 3 tasks. We use Gemini to classify a set of 456 activities in the ATUS taxonomy into physical or cognitive tasks. The vertical dashed line marks exact parity between AI share and human time share. The distribution for cognitive/non￾physical tasks is shifted substantially to the right, with a mean representation ratio (defined as the AI conversation share divided by the ATUS…
Figure 40 of the paper
Figure 40. Conversational AI Usage Per Capita Shows Higher Adoption in Higher-Income Countries, Even Prior to Gemini Penetration Adjustment Notes: Global distribution of conversational AI (Gemini Apps and AI Mode) per capita usage by country/region quintiles (unadjusted for penetration). Calculations are based on weighted conversation counts across both platforms. Several countries/regions are excluded due to consumer-level access being unavailable, or queries primarily occurring through enterprise Workspace accounts. Per-capita rates are calculated using World Bank (2024c) population…
Figure 41 of the paper
Figure 41. Unadjusted Per Capita Usage Highlights High Differences in Adoption Between Countries Notes: Global population share versus unadjusted AI conversation share across five per-capita usage quintiles at the country/region level. Multipliers denote the ratio of usage share to population share for each quintile. China and regions with populations under 1,000,000 are excluded. The figure utilize the same methodology as in Section 5 (unadjusted for penetration version). Note that the analysis aggregates country-level data, which may mask within-country inequality in adoption and usage. 96
Figure 42 of the paper
Figure 42. Unadjusted Conversational AI Usage Intensity Maintains a Strong Positive Correlation With GDP Per Capita Notes: Unadjusted per-capita conversation counts versus PPP GDP per capita on a log-log scale. Per-capita conversations are indexed to a maximum value of 100. GDP per capita data is sourced from the World Bank (2024a) WDI database (or International Monetary Fund (2024) WEO where missing). Only countries/regions with more than 1,000,000 inhabitants with official availability of Gemini Apps are featured; Mainland China is excluded. The figure utilize the same methodology as in…
Figure 43 of the paper
Figure 43. Adjusting for Internet Access Elevates Relative Adoption Metrics for Connectivity-Constrained Regions Notes: Regional comparison of per-capita vs. per-internet-user adoption (unadjusted for Gemini penetration). The internet-access unadjusted metric utilizes internet users (calculated using World Bank (2024b) internet access percentages from WDI) in the denominator instead of total population. The comparison is aggregated at the regional level, which masks country-specific variation and digital divides within regions. Both metrics are scaled to a maximum regional value of 100. The…
Figure 44 of the paper
Figure 44. Established Digital Hubs Maintain Per-Capita Leadership in Work-Related AI Conversations Prior to Penetration Adjustment Like Sub-Saharan Africa Notes: This figure displays the per-capita intensity of work-related conversations, unadjusted for usage share, by country/region quintiles. The metric divides work conversations per capita by the StatCounter Gemini usage share proxy. China and countries with populations under 1,000,000 are excluded. To prevent extreme outliers from small sample sizes, countries with a Gemini usage share of 0.1% or lower are excluded. . 99
Figure 45 of the paper
Figure 45. Interacting in a Non-Primary Language Drives Longer Sessions and Higher Token Verbosity Notes: Regression point estimates (with 95% confidence intervals) evaluating how interacting in a non-primary language affects session length (turn count), verbosity (total tokens), domain technical depth score, and media usage relative to primary-language interactions. The Domain Knowledge Score maps qualitative ratings (Low, Medium, High) to a 1–3 scale. Four progressive model specifications are shown, incorporating language fixed effects and increasingly granular ATUS activity category fixed…

Showing 41 of the paper's 45 figures. Figures are the authors', reproduced from the paper.

References (58)
  1. [1] Acemoglu, Daron and David Autor, “Skills, Tasks and Technologies: Implications for Employment and Earnings,”Chapter 12, Handbook of Labor Economics, 2011, p. 1043–1171. and Pascual Restrepo, “Automation and New Tasks: How Technology Displaces and Reinstates Labor,”Journal of E... on Pith
  2. [4] Autor, David H., Frank Levy, and Richard J. Murnane, “The Skill Content of Recent Technological Change: An Empirical Exploration,”The Quarterly Journal of Economics, 11 2003,118(4), 1279–1333. Barop, Johanna and Joseph Levine, “AfriGPT: LLMs over SMS,” on Pith
  3. [6] Becker, Gary S., “A Theory of the Allocation of Time,”The Economic Journal, 9 1965,75 (299), 493–517. Bick, Alexander, Adam Blandin, and David J. Deming, “The Rapid Adoption of Generative AI,” Working Paper 32966, National Bureau of Economic Research September on Pith
  4. [7] , , , Nicola Fuchs-Schündeln, and Jonas Jessen, “Mind the Gap: AI Adoption in Europe and the US,”Brookings Papers on Economic Activity, 2026,Spring. Björkegren, Daniel, Jun Ho Choi, Divya Budihal, Dominic Sobhani, Oliver Garrod, and Paul Atherton, “Could AI Leapfrog the Web? E... on Pith
  5. [8] BLS, “2024 American Time Use Survey (ATUS) Activity Lexicon,” U.S. Bureau of Labor Statistics on Pith
  6. [9] BLS,2018 Standard Occupational Classification ManualU.S. Bureau of Labor Statistics on Pith
  7. [11] or Business and Financial Operations Occupation (SOC 13). The same issues apply to time use categories: an adult asking AI for instructions on building a nest box with a 10-year old could do so for their own child (‘Arts and crafts with household children’) or their visiting n... on Pith
  8. [12] , Avinash Collis, W. Erwin Diewert, Felix Eggers, and Kevin J. Fox, “GDP-B: Ac- counting for the Value of New and Free Goods,”American Economic Journal: Macroeconomics, October 2025,17(4), 312–44. , Bharat Chandar, and Ruyu Chen, “Canaries in the Coal Mine? Six Facts about the... on Pith
  9. [13] , Danielle Li, and Lindsey Raymond, “Generative AI at Work,”The Quarterly Journal of Economics, February 2025,140(2), 889–942. Bureau of Labor Statistics, “Educational attainment for workers 25 years and older by detailed occupation,” U.S. Department of Labor on Pith
  10. [14] , “Occupational Employment and Wage Statistics May 2024 National Data,” U.S. Department of Labor on Pith
  11. [16] Charikar, Moses S., “Similarity estimation techniques from rounding algorithms,” in “Pro- ceedings of the thiry-fourth annual ACM symposium on Theory of computing” STOC02 ACM May 2002, p. 380–388. Chatterji, Aaron, Thomas Cunningham, David J Deming, Zoe Hitzig, Christopher Ong... on Pith
  12. [17] Comin, Diego and Bart Hobijn, “An Exploration of Technology Diffusion,”American Economic Review, December 2010,100(5), 2031–59. and Marti Mestieri, “Technology Diffusion: Measurement, Causes, and Consequences,” None on Pith
  13. [19] Original version October 2024, revised May on Pith
  14. [20] Dell’Acqua, Fabrizio, Edward McFowland, Ethan Mollick, Hila Lifshitz, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R. Lakhani, “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligen... on Pith
  15. [21] Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock, “GPTs are GPTs: Labor market impact potential of LLMs,”Science, 2024,384(6702), 1306–1308. England, Paula, Michelle Budig, and Nancy Folbre, “Wages of Virtue: The Relative Pay of Care Work,”Social Problems, November... on Pith
  16. [23] Gentzkow, Matthew, Bryan Kelly, and Matt Taddy, “Text as Data,”Journal of Economic Literature, September 2019,57(3), 535–74. Gimbel, Martha, Joshua Kendall, and Ryan Nunn, “What We Do and Don’t Know About How AI is Affecting the Labor Market,” The Budget Lab May on Pith
  17. [25] , “Google Search’s I/O 2026 updates: AI agents and more,” Google Blog May on Pith
  18. [26] Gottfried, Jeffrey, William Bishop, Monica Anderson, Michelle Faverio, Eugenie Park, and Colleen McClain, “Americans and AI 2026: Chatbots, Smart Devices and Views on Impact,” Report, Pew Research Center June on Pith
  19. [28] Hertog, Ekaterina, Setsuya Fukuda, Rikiya Matsukura, Nobuko Nagase, and Vili Lehdonvirta, “The future of unpaid work: Estimating the effects of automation on time spent on housework and care work in Japan and the UK,”Technological Forecasting and Social Change, 2023,191, 12244... on Pith
  20. [29] Substack blog. International Monetary Fund, “World Economic Outlook Database, October 2024,” Wash- ington, DC: International Monetary Fund on Pith
show 38 more
  1. [30] Ipsos and Google, “Google / Ipsos Multi-Country AI Survey 2026,” Report, Ipsos January on Pith
  2. [31] Jones, Charles I. and Peter J. Klenow, “Beyond GDP? Welfare across Countries and Time,” American Economic Review, 2016,106(9), 2426–2457. Kalyani, Aakash, Nicholas Bloom, Marcela Carvalho, Tarek Hassan, Josh Lerner, and Ahmed Tahoun, “The Diffusion of New Technologies,”The Qua... on Pith
  3. [32] Korinek, Anton, “Generative AI for Economic Research: Use Cases and Implications for Economists,”Journal of Economic Literature, January 2023,61(4), 1281–1317. and Joseph E Stiglitz, “Artificial Intelligence, Globalization, and Strategies for Economic Development,” Working Pap... on Pith
  4. [34] Light, Richard J., “Measures of response agreement for qualitative data: Some generalizations and alternatives.,”Psychological Bulletin, November 1971,76(5), 365–377. Lindenlaub, Ilse, Ryungha Oh, Maria Alejandra Rodriguez, and Laura Veldkamp, “Beyond Exposure: Predicting AI A... on Pith
  5. [35] Lundin, JessicaM., AdaZhang, NihalKarim, HamzaLouzan, GuohaoWei, DavidIfe- oluwa Adelani, and Cody Carroll, “The Token Tax: Systematic Bias in Multilingual Tokenization,” in “Proceedings of the 7th Workshop on African Natural Language Processing (AfricaNLP 2026)” Association f... on Pith
  6. [36] Mathiowetz, Nancy A., “Errors in Reports of Occupation,”Public Opinion Quarterly, 01 1992,56(3), 352–355. Mellow, Wesley and Hal Sider, “Accuracy of Response in Labor Market Surveys: Evidence and Implications,”Journal of Labor Economics, 10 1983,1(4), 331–344. 68 METR, “Task-C... on Pith
  7. [37] Milkie, Melissa A., Liana C. Sayer, Kei Nomaguchi, and Hope Xu Yan, “Who’s Doing the Housework and Childcare in America Now? Differential Convergence in Twenty-First- Century Gender Gaps in Home Tasks,”Socius: Sociological Research for a Dynamic World, February 2025,11. Misra,... on Pith
  8. [38] , , , , and Juan M. Lavista Ferres, “Measuring AI Diffusion Across U.S. Geographies: County, State, and Metro Estimates,” Technical Report MSR-TR-2026-18, Microsoft Research May on Pith
  9. [39] Møgelvang, Anja, Camilla Bjelland, Simone Grassini, and Kristine Ludvigsen, “Gen- der Differences in the Use of Generative Artificial Intelligence Chatbots in Higher Education: Characteristics and Consequences,”Education Sciences, 2024,14(12). National Center for O*NET Develop... on Pith
  10. [41] Ramey, Valerie A., “Time Spent in Home Production in the Twentieth-Century United States: New Estimates from Old Data,”The Journal of Economic History, 2009,69(1), 1–47. Reid, Margaret,Economics of Household Production, London: Chapman and Hall, on Pith
  11. [42] Sajadieh, Sha, Loredana Fattorini, Raymond Perrault, Yolanda Gil, Vanessa Parli, Lapo Santarlasci, Juan Pava, Nestor Maslej, Russ Altman, Erik Brynjolfsson, Carla Brodley, Jack Clark, Virginia Dignum, Vipin Kumar, James Landay, Terah Lyons, James Manyika, Juan Carlos Niebles, ... on Pith
  12. [43] Schreyer, Paul and W. Erwin Diewert, “Household Production, Leisure, and Living Standards,” September 2014, pp. 89–114. Scur, Daniela, Raffaella Sadun, John Van Reenen, Renata Lemos, and Nicholas Bloom, “The World Management Survey at 18: lessons and the way forward,”Oxford Re... on Pith
  13. [44] Syverson, Chad, “What Determines Productivity?,”Journal of Economic Literature, June 2011,49(2), 326–65. Tamkin, Alex, Miles McCain, Kunal Handa, Esin Durmus, Liane Lovitt, Ankur Rathi, Saffron Huang, Alfred Mountfield, Jerry Hong, Stuart Ritchie, Michael Stern, Brian Clarke, ... on Pith
  14. [45] Verma, Gaurav, Siddharth Suri, and Scott Counts, “How Copilot Changed the Pace of Work in Word — And How We Measured It,” Technical Report MSR-TR-2026-24, Microsoft June on Pith
  15. [46] 70 Warrens, Matthijs J., “A Formal Proof of a Paradox Associated with Cohen’s Kappa,”Journal of Classification, October 2010,27(3), 322–332. Wilmers, Nathan, “Generative AI and the Future of Inequality,”An MIT Exploration of Generative AI, mar 27 on Pith
  16. [47] 71 Appendix A Data and Methodology This section provides further detail about the data sources for ATLAS, the methodological steps in preparing the dataset and privacy preservation measures we took at every step. In brief, the fully automated pipeline for data processing proce... on Pith
  17. [48] On each of the surfaces noted above, we sample approximately 5 million interactions—full conversations (for Gemini App and AIM) and request-response pairs (for Gemini API)—which satisfy basic data cleaning and materiality filters.65 These samples are then re-weighted by the ov... on Pith
  18. [49] and industry (Tamkin et al., 2024; Eloundou et al., 2024; Tomlinson et al., 2025; Verma et al., 2026; Chatterji et al., 2025). To support research, as well as other internal use cases, Google DeepMind has developed Observation Clustering and Taxonomy Organisation (OCTO), a too... on Pith
  19. [51] to ensure classification outputs conform to one of the provided taxa only. This approach is then applied recursively to traverse the taxonomy, assigning clusters at each level to the most suitable category within the set of descendants from the category selected at a higher le... on Pith
  20. [54] for examples of classified conversations for human raters to review to protect the privacy of our users. B.3.1 Inter-rater Agreement We start the measurement of inter-rater agreement by identifying the plurality consensus among the three human annotators, defined as the catego... on Pith
  21. [56] to omit tasks that are not important. We exclude any occupation for which fewer than 50% of tasks have available importance ratings. 89 Figure 33:Distribution of Task Expertise Scores Across the O*NET Task Universe Notes:We calculate scores measuring the expertise of tasks in ... on Pith
  22. [57] 90 D Additional Tables and Figures, Household Section Figure 34:Granular Task Comparison Highlights AI Over-Representation in Specialized Knowledge and Under-Representation in Routine Chores Notes:Compares U.S. AI conversation shares against ATUS human time shares across the 1... on Pith
  23. [100] GDP per capita data is sourced from the World Bank (2024a) WDI database (or International Monetary Fund (2024) WEO where missing). Only countries/regions with more than 1,000,000 inhabitants with official availability of Gemini Apps are featured; Mainland China is excluded. Th... on Pith
  24. [1977] kappa evaluation scale. Mean pairwise Cohen’s and Fleiss’ kappas show ‘moderate’ or ‘substantial’ agreement between the LLM classifiers and human annotators. For example, the SOC major groups classification achieved a plurality agreement kappa of 0.83, and Fleiss’ kappa of 0.6... on Pith
  25. [1983] Caselli, Francesco and Wilbur John Coleman, “Cross-Country Technology Diffusion: The Case of Computers,”American Economic Review, May 2001,91(2), 328–335. Cerutti, Eugenio, Antonio Garcia Pascual, Yosuke Kido, Longji Li, Giovanni Melina, Marina Mendes Tavares, and Philippe Win... on Pith
  26. [1988] Noy, Shakked and Whitney Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence,”Science, 2023,381(6654), 187–192. OECD,OECD Policy Insights on Well-being, Inclusion and Equal OpportunityDecember on Pith
  27. [1992] Bridgman, Benjamin, “A Disaggregated View of Household Production Trends,”AEA Papers and Proceedings, May 2023,113, 619–22. , Andrew Craig, and Danit Kanal, “Accounting for Household Production in the National Accounts: An Update 1965–2020,”Survey of Current Business, 2022,102... on Pith
  28. [2001] can best improve economic research on AI use. B.1.2 Classifier Validation Under Information Constraints Assessing classifier performance in contexts like ATLAS is inherently difficult due to lack of ground truth data. Even if unrestricted access to raw logs were possible, the ... on Pith
  29. [2002] to avoid evaluating probabilistically distant pairs. To maximize recall, this process is repeated across multiple independent projection vectors to yield an approximate nearest neighbor graph. Finally, the resulting graph acts as the input to affinity clustering, a large-scale... on Pith
  30. [2014] , “Do-it-yourself Digital: the Production Boundary, the Productivity Puzzle and Economic Welfare,”Economica, August 2018,86(344), 750–774. and Leonard Nakamura, “Time Use, Productivity, and Household-Centric Measurement of Welfare in the Digital Economy,” on Pith
  31. [2018] Bound, John, Charles Brown, and Nancy Mathiowetz, “Measurement Error in Survey Data,” 2001, p. 3705–3843. Bresnahan, Timothy F and Manuel Trajtenberg, “General Purpose Technologies, Engines of Growth?,” Working Paper 4148, National Bureau of Economic Research August on Pith
  32. [2019] , , and , “Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses,”Strategic Management Journal, May 2021,42(12), 2195–2217. Fleiss, Joseph L., “Measuring nominal scale agreement among many raters.,”Psychological Bull... on Pith
  33. [2021] Kuznets, Simon, “National Income, 1929-1932,” in “National Income, 1929-1932,” National Bureau of Economic Research, June 1934, pp. 1–12. Lancaster, Kelvin J., “A New Approach to Consumer Theory,”Journal of Political Economy, 1966,74(2). 67 Landis, J. Richard and Gary G. Koch,... on Pith
  34. [2022] Gronau, Reuben, “Leisure, Home Production, and Work–the Theory of the Allocation of Time Revisited,”Journal of Political Economy, 1977,85(6), 1099–1123. Gupta, Ashim, MaitreyMehta, ZhichaoXu, andVivekSrikumar, “FoundinTranslation: Measuring Multilingual LLM Consistency as Simp... on Pith
  35. [2023] Ahuja, Kabir, Harshita Diddee, Rishav Hada, Millicent Ochieng, Krithika Ramesh, Prachi Jain, Akshay Nambi, Tanuja Ganu, Sameer Segal, Mohamed Ahmed, Kalika Bali, and Sunayana Sitaram, “MEGA: Multilingual Evaluation of Generative AI,” in Houda Bouamor, Juan Pino, and Kalika Bal... on Pith
  36. [2024] Google, “Google I/O 2026: Sundar Pichai’s opening keynote,” Google Blog May on Pith
  37. [2025] Artificial Analysis, “Multilingual AI Model Benchmark,”https://artificialanalysis.ai/ models/multilingual2026. Audoly, Richard, Miles Guerin, and Giorgio Topa, “Do Job Postings Show Early Labor-Market Effects of AI?,”Liberty Street Economics, 2026, (May 14). Liberty Street Eco... on Pith
  38. [2026] Work in progress, Oxford Internet Institute in partnership with Africell. Bateni, MohammadHossein, Soheil Behnezhad, Mahsa Derakhshan, Mohammad- TaghiHajiaghayi, RaimondasKiveris, SilvioLattanzi, andVahabMirrokni, “Affinity Clustering: Hierarchical Clustering at Scale,” in “NI... on Pith
Cited by 4 Pith papers
How the literature uses it

What this paper does with the arXiv papers it cites

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

What it cites · 8 sentences without a confident label

Uses data or values from 2 papers
  • [ref] analyzed AI task usage, determining 57% of usage suggests augmentation while 43% of usage suggests automation of human capabilities.

    Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations · The Broad Picture · 0.86

  • [ref] use household internet browsing data from US home devices to find that household AI adoption may translate to an effective time savings of roughly 43%–64%, but these estimates may differ for non-browsing activities.

    2603.03144 · Non-Market Welfare: Valuing Household Time Savings · 0.83

Compares with 1 paper
Cites as support 1 paper
  • In this scenario, existing advantages are locked in and potentially amplified through several channels: * Task Selection: When task choice is left to the worker, lower-skilled workers may apply AI inappropriately, resulting in performance drops, whereas higher-skilled workers select tasks where the tool provides high utility [ref] . * Complementary Judgment: Effective utilization of AI in non-routine cognitive work requires high-level human capital inputs such as architectural judgment, error-detection, and code-reading ability to verify and integrate AI outputs [ref] . * Seniority-Biased Demand: If senior workers leverage AI more effectively, firms may substitute away from entry-level hiring while keeping senior employment stable, leading to a seniority-biased technological change that shifts career ladders and exacerbates wage premiums [ref] .

    How much does AI impact development speed? An enterprise-based randomized controlled trial · Discussion: The Adoption Margin and Implications for Inequality · 0.97

Cites as background 1 paper

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment